Mon, 26 Sep 2011 10:24:05 -0700
7081933: Use zeroing elimination optimization for large array
Summary: Don't zero new typeArray during runtime call if the allocation is followed by arraycopy into it.
Reviewed-by: twisti
1 /*
2 * Copyright (c) 2005, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "libadt/vectset.hpp"
28 #include "memory/allocation.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/compile.hpp"
33 #include "opto/escape.hpp"
34 #include "opto/phaseX.hpp"
35 #include "opto/rootnode.hpp"
37 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) {
38 uint v = (targIdx << EdgeShift) + ((uint) et);
39 if (_edges == NULL) {
40 Arena *a = Compile::current()->comp_arena();
41 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0);
42 }
43 _edges->append_if_missing(v);
44 }
46 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) {
47 uint v = (targIdx << EdgeShift) + ((uint) et);
49 _edges->remove(v);
50 }
52 #ifndef PRODUCT
53 static const char *node_type_names[] = {
54 "UnknownType",
55 "JavaObject",
56 "LocalVar",
57 "Field"
58 };
60 static const char *esc_names[] = {
61 "UnknownEscape",
62 "NoEscape",
63 "ArgEscape",
64 "GlobalEscape"
65 };
67 static const char *edge_type_suffix[] = {
68 "?", // UnknownEdge
69 "P", // PointsToEdge
70 "D", // DeferredEdge
71 "F" // FieldEdge
72 };
74 void PointsToNode::dump(bool print_state) const {
75 NodeType nt = node_type();
76 tty->print("%s ", node_type_names[(int) nt]);
77 if (print_state) {
78 EscapeState es = escape_state();
79 tty->print("%s %s ", esc_names[(int) es], _scalar_replaceable ? "":"NSR");
80 }
81 tty->print("[[");
82 for (uint i = 0; i < edge_count(); i++) {
83 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]);
84 }
85 tty->print("]] ");
86 if (_node == NULL)
87 tty->print_cr("<null>");
88 else
89 _node->dump();
90 }
91 #endif
93 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
94 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()),
95 _processed(C->comp_arena()),
96 pt_ptset(C->comp_arena()),
97 pt_visited(C->comp_arena()),
98 pt_worklist(C->comp_arena(), 4, 0, 0),
99 _collecting(true),
100 _progress(false),
101 _compile(C),
102 _igvn(igvn),
103 _node_map(C->comp_arena()) {
105 _phantom_object = C->top()->_idx,
106 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true);
108 // Add ConP(#NULL) and ConN(#NULL) nodes.
109 Node* oop_null = igvn->zerocon(T_OBJECT);
110 _oop_null = oop_null->_idx;
111 assert(_oop_null < C->unique(), "should be created already");
112 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
114 if (UseCompressedOops) {
115 Node* noop_null = igvn->zerocon(T_NARROWOOP);
116 _noop_null = noop_null->_idx;
117 assert(_noop_null < C->unique(), "should be created already");
118 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
119 }
120 }
122 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) {
123 PointsToNode *f = ptnode_adr(from_i);
124 PointsToNode *t = ptnode_adr(to_i);
126 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
127 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge");
128 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge");
129 add_edge(f, to_i, PointsToNode::PointsToEdge);
130 }
132 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) {
133 PointsToNode *f = ptnode_adr(from_i);
134 PointsToNode *t = ptnode_adr(to_i);
136 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
137 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
138 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
139 // don't add a self-referential edge, this can occur during removal of
140 // deferred edges
141 if (from_i != to_i)
142 add_edge(f, to_i, PointsToNode::DeferredEdge);
143 }
145 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
146 const Type *adr_type = phase->type(adr);
147 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
148 adr->in(AddPNode::Address)->is_Proj() &&
149 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
150 // We are computing a raw address for a store captured by an Initialize
151 // compute an appropriate address type. AddP cases #3 and #5 (see below).
152 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
153 assert(offs != Type::OffsetBot ||
154 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
155 "offset must be a constant or it is initialization of array");
156 return offs;
157 }
158 const TypePtr *t_ptr = adr_type->isa_ptr();
159 assert(t_ptr != NULL, "must be a pointer type");
160 return t_ptr->offset();
161 }
163 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) {
164 PointsToNode *f = ptnode_adr(from_i);
165 PointsToNode *t = ptnode_adr(to_i);
167 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
168 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge");
169 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge");
170 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets");
171 t->set_offset(offset);
173 add_edge(f, to_i, PointsToNode::FieldEdge);
174 }
176 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) {
177 PointsToNode *npt = ptnode_adr(ni);
178 PointsToNode::EscapeState old_es = npt->escape_state();
179 if (es > old_es)
180 npt->set_escape_state(es);
181 }
183 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt,
184 PointsToNode::EscapeState es, bool done) {
185 PointsToNode* ptadr = ptnode_adr(n->_idx);
186 ptadr->_node = n;
187 ptadr->set_node_type(nt);
189 // inline set_escape_state(idx, es);
190 PointsToNode::EscapeState old_es = ptadr->escape_state();
191 if (es > old_es)
192 ptadr->set_escape_state(es);
194 if (done)
195 _processed.set(n->_idx);
196 }
198 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) {
199 uint idx = n->_idx;
200 PointsToNode::EscapeState es;
202 // If we are still collecting or there were no non-escaping allocations
203 // we don't know the answer yet
204 if (_collecting)
205 return PointsToNode::UnknownEscape;
207 // if the node was created after the escape computation, return
208 // UnknownEscape
209 if (idx >= nodes_size())
210 return PointsToNode::UnknownEscape;
212 es = ptnode_adr(idx)->escape_state();
214 // if we have already computed a value, return it
215 if (es != PointsToNode::UnknownEscape &&
216 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject)
217 return es;
219 // PointsTo() calls n->uncast() which can return a new ideal node.
220 if (n->uncast()->_idx >= nodes_size())
221 return PointsToNode::UnknownEscape;
223 PointsToNode::EscapeState orig_es = es;
225 // compute max escape state of anything this node could point to
226 for(VectorSetI i(PointsTo(n)); i.test() && es != PointsToNode::GlobalEscape; ++i) {
227 uint pt = i.elem;
228 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
229 if (pes > es)
230 es = pes;
231 }
232 if (orig_es != es) {
233 // cache the computed escape state
234 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state");
235 ptnode_adr(idx)->set_escape_state(es);
236 } // orig_es could be PointsToNode::UnknownEscape
237 return es;
238 }
240 VectorSet* ConnectionGraph::PointsTo(Node * n) {
241 pt_ptset.Reset();
242 pt_visited.Reset();
243 pt_worklist.clear();
245 #ifdef ASSERT
246 Node *orig_n = n;
247 #endif
249 n = n->uncast();
250 PointsToNode* npt = ptnode_adr(n->_idx);
252 // If we have a JavaObject, return just that object
253 if (npt->node_type() == PointsToNode::JavaObject) {
254 pt_ptset.set(n->_idx);
255 return &pt_ptset;
256 }
257 #ifdef ASSERT
258 if (npt->_node == NULL) {
259 if (orig_n != n)
260 orig_n->dump();
261 n->dump();
262 assert(npt->_node != NULL, "unregistered node");
263 }
264 #endif
265 pt_worklist.push(n->_idx);
266 while(pt_worklist.length() > 0) {
267 int ni = pt_worklist.pop();
268 if (pt_visited.test_set(ni))
269 continue;
271 PointsToNode* pn = ptnode_adr(ni);
272 // ensure that all inputs of a Phi have been processed
273 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
275 int edges_processed = 0;
276 uint e_cnt = pn->edge_count();
277 for (uint e = 0; e < e_cnt; e++) {
278 uint etgt = pn->edge_target(e);
279 PointsToNode::EdgeType et = pn->edge_type(e);
280 if (et == PointsToNode::PointsToEdge) {
281 pt_ptset.set(etgt);
282 edges_processed++;
283 } else if (et == PointsToNode::DeferredEdge) {
284 pt_worklist.push(etgt);
285 edges_processed++;
286 } else {
287 assert(false,"neither PointsToEdge or DeferredEdge");
288 }
289 }
290 if (edges_processed == 0) {
291 // no deferred or pointsto edges found. Assume the value was set
292 // outside this method. Add the phantom object to the pointsto set.
293 pt_ptset.set(_phantom_object);
294 }
295 }
296 return &pt_ptset;
297 }
299 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
300 // This method is most expensive during ConnectionGraph construction.
301 // Reuse vectorSet and an additional growable array for deferred edges.
302 deferred_edges->clear();
303 visited->Reset();
305 visited->set(ni);
306 PointsToNode *ptn = ptnode_adr(ni);
308 // Mark current edges as visited and move deferred edges to separate array.
309 for (uint i = 0; i < ptn->edge_count(); ) {
310 uint t = ptn->edge_target(i);
311 #ifdef ASSERT
312 assert(!visited->test_set(t), "expecting no duplications");
313 #else
314 visited->set(t);
315 #endif
316 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
317 ptn->remove_edge(t, PointsToNode::DeferredEdge);
318 deferred_edges->append(t);
319 } else {
320 i++;
321 }
322 }
323 for (int next = 0; next < deferred_edges->length(); ++next) {
324 uint t = deferred_edges->at(next);
325 PointsToNode *ptt = ptnode_adr(t);
326 uint e_cnt = ptt->edge_count();
327 for (uint e = 0; e < e_cnt; e++) {
328 uint etgt = ptt->edge_target(e);
329 if (visited->test_set(etgt))
330 continue;
332 PointsToNode::EdgeType et = ptt->edge_type(e);
333 if (et == PointsToNode::PointsToEdge) {
334 add_pointsto_edge(ni, etgt);
335 if(etgt == _phantom_object) {
336 // Special case - field set outside (globally escaping).
337 ptn->set_escape_state(PointsToNode::GlobalEscape);
338 }
339 } else if (et == PointsToNode::DeferredEdge) {
340 deferred_edges->append(etgt);
341 } else {
342 assert(false,"invalid connection graph");
343 }
344 }
345 }
346 }
349 // Add an edge to node given by "to_i" from any field of adr_i whose offset
350 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
351 // a pointsto edge is added if it is a JavaObject
353 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
354 PointsToNode* an = ptnode_adr(adr_i);
355 PointsToNode* to = ptnode_adr(to_i);
356 bool deferred = (to->node_type() == PointsToNode::LocalVar);
358 for (uint fe = 0; fe < an->edge_count(); fe++) {
359 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
360 int fi = an->edge_target(fe);
361 PointsToNode* pf = ptnode_adr(fi);
362 int po = pf->offset();
363 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
364 if (deferred)
365 add_deferred_edge(fi, to_i);
366 else
367 add_pointsto_edge(fi, to_i);
368 }
369 }
370 }
372 // Add a deferred edge from node given by "from_i" to any field of adr_i
373 // whose offset matches "offset".
374 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
375 PointsToNode* an = ptnode_adr(adr_i);
376 for (uint fe = 0; fe < an->edge_count(); fe++) {
377 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
378 int fi = an->edge_target(fe);
379 PointsToNode* pf = ptnode_adr(fi);
380 int po = pf->offset();
381 if (pf->edge_count() == 0) {
382 // we have not seen any stores to this field, assume it was set outside this method
383 add_pointsto_edge(fi, _phantom_object);
384 }
385 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
386 add_deferred_edge(from_i, fi);
387 }
388 }
389 }
391 // Helper functions
393 static Node* get_addp_base(Node *addp) {
394 assert(addp->is_AddP(), "must be AddP");
395 //
396 // AddP cases for Base and Address inputs:
397 // case #1. Direct object's field reference:
398 // Allocate
399 // |
400 // Proj #5 ( oop result )
401 // |
402 // CheckCastPP (cast to instance type)
403 // | |
404 // AddP ( base == address )
405 //
406 // case #2. Indirect object's field reference:
407 // Phi
408 // |
409 // CastPP (cast to instance type)
410 // | |
411 // AddP ( base == address )
412 //
413 // case #3. Raw object's field reference for Initialize node:
414 // Allocate
415 // |
416 // Proj #5 ( oop result )
417 // top |
418 // \ |
419 // AddP ( base == top )
420 //
421 // case #4. Array's element reference:
422 // {CheckCastPP | CastPP}
423 // | | |
424 // | AddP ( array's element offset )
425 // | |
426 // AddP ( array's offset )
427 //
428 // case #5. Raw object's field reference for arraycopy stub call:
429 // The inline_native_clone() case when the arraycopy stub is called
430 // after the allocation before Initialize and CheckCastPP nodes.
431 // Allocate
432 // |
433 // Proj #5 ( oop result )
434 // | |
435 // AddP ( base == address )
436 //
437 // case #6. Constant Pool, ThreadLocal, CastX2P or
438 // Raw object's field reference:
439 // {ConP, ThreadLocal, CastX2P, raw Load}
440 // top |
441 // \ |
442 // AddP ( base == top )
443 //
444 // case #7. Klass's field reference.
445 // LoadKlass
446 // | |
447 // AddP ( base == address )
448 //
449 // case #8. narrow Klass's field reference.
450 // LoadNKlass
451 // |
452 // DecodeN
453 // | |
454 // AddP ( base == address )
455 //
456 Node *base = addp->in(AddPNode::Base)->uncast();
457 if (base->is_top()) { // The AddP case #3 and #6.
458 base = addp->in(AddPNode::Address)->uncast();
459 while (base->is_AddP()) {
460 // Case #6 (unsafe access) may have several chained AddP nodes.
461 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only");
462 base = base->in(AddPNode::Address)->uncast();
463 }
464 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
465 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
466 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
467 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
468 }
469 return base;
470 }
472 static Node* find_second_addp(Node* addp, Node* n) {
473 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
475 Node* addp2 = addp->raw_out(0);
476 if (addp->outcnt() == 1 && addp2->is_AddP() &&
477 addp2->in(AddPNode::Base) == n &&
478 addp2->in(AddPNode::Address) == addp) {
480 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
481 //
482 // Find array's offset to push it on worklist first and
483 // as result process an array's element offset first (pushed second)
484 // to avoid CastPP for the array's offset.
485 // Otherwise the inserted CastPP (LocalVar) will point to what
486 // the AddP (Field) points to. Which would be wrong since
487 // the algorithm expects the CastPP has the same point as
488 // as AddP's base CheckCastPP (LocalVar).
489 //
490 // ArrayAllocation
491 // |
492 // CheckCastPP
493 // |
494 // memProj (from ArrayAllocation CheckCastPP)
495 // | ||
496 // | || Int (element index)
497 // | || | ConI (log(element size))
498 // | || | /
499 // | || LShift
500 // | || /
501 // | AddP (array's element offset)
502 // | |
503 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
504 // | / /
505 // AddP (array's offset)
506 // |
507 // Load/Store (memory operation on array's element)
508 //
509 return addp2;
510 }
511 return NULL;
512 }
514 //
515 // Adjust the type and inputs of an AddP which computes the
516 // address of a field of an instance
517 //
518 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
519 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
520 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
521 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
522 if (t == NULL) {
523 // We are computing a raw address for a store captured by an Initialize
524 // compute an appropriate address type (cases #3 and #5).
525 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
526 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
527 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
528 assert(offs != Type::OffsetBot, "offset must be a constant");
529 t = base_t->add_offset(offs)->is_oopptr();
530 }
531 int inst_id = base_t->instance_id();
532 assert(!t->is_known_instance() || t->instance_id() == inst_id,
533 "old type must be non-instance or match new type");
535 // The type 't' could be subclass of 'base_t'.
536 // As result t->offset() could be large then base_t's size and it will
537 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
538 // constructor verifies correctness of the offset.
539 //
540 // It could happened on subclass's branch (from the type profiling
541 // inlining) which was not eliminated during parsing since the exactness
542 // of the allocation type was not propagated to the subclass type check.
543 //
544 // Or the type 't' could be not related to 'base_t' at all.
545 // It could happened when CHA type is different from MDO type on a dead path
546 // (for example, from instanceof check) which is not collapsed during parsing.
547 //
548 // Do nothing for such AddP node and don't process its users since
549 // this code branch will go away.
550 //
551 if (!t->is_known_instance() &&
552 !base_t->klass()->is_subtype_of(t->klass())) {
553 return false; // bail out
554 }
556 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
557 // Do NOT remove the next line: ensure a new alias index is allocated
558 // for the instance type. Note: C++ will not remove it since the call
559 // has side effect.
560 int alias_idx = _compile->get_alias_index(tinst);
561 igvn->set_type(addp, tinst);
562 // record the allocation in the node map
563 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered");
564 set_map(addp->_idx, get_map(base->_idx));
566 // Set addp's Base and Address to 'base'.
567 Node *abase = addp->in(AddPNode::Base);
568 Node *adr = addp->in(AddPNode::Address);
569 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
570 adr->in(0)->_idx == (uint)inst_id) {
571 // Skip AddP cases #3 and #5.
572 } else {
573 assert(!abase->is_top(), "sanity"); // AddP case #3
574 if (abase != base) {
575 igvn->hash_delete(addp);
576 addp->set_req(AddPNode::Base, base);
577 if (abase == adr) {
578 addp->set_req(AddPNode::Address, base);
579 } else {
580 // AddP case #4 (adr is array's element offset AddP node)
581 #ifdef ASSERT
582 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
583 assert(adr->is_AddP() && atype != NULL &&
584 atype->instance_id() == inst_id, "array's element offset should be processed first");
585 #endif
586 }
587 igvn->hash_insert(addp);
588 }
589 }
590 // Put on IGVN worklist since at least addp's type was changed above.
591 record_for_optimizer(addp);
592 return true;
593 }
595 //
596 // Create a new version of orig_phi if necessary. Returns either the newly
597 // created phi or an existing phi. Sets create_new to indicate whether a new
598 // phi was created. Cache the last newly created phi in the node map.
599 //
600 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
601 Compile *C = _compile;
602 new_created = false;
603 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
604 // nothing to do if orig_phi is bottom memory or matches alias_idx
605 if (phi_alias_idx == alias_idx) {
606 return orig_phi;
607 }
608 // Have we recently created a Phi for this alias index?
609 PhiNode *result = get_map_phi(orig_phi->_idx);
610 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
611 return result;
612 }
613 // Previous check may fail when the same wide memory Phi was split into Phis
614 // for different memory slices. Search all Phis for this region.
615 if (result != NULL) {
616 Node* region = orig_phi->in(0);
617 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
618 Node* phi = region->fast_out(i);
619 if (phi->is_Phi() &&
620 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
621 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
622 return phi->as_Phi();
623 }
624 }
625 }
626 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
627 if (C->do_escape_analysis() == true && !C->failing()) {
628 // Retry compilation without escape analysis.
629 // If this is the first failure, the sentinel string will "stick"
630 // to the Compile object, and the C2Compiler will see it and retry.
631 C->record_failure(C2Compiler::retry_no_escape_analysis());
632 }
633 return NULL;
634 }
635 orig_phi_worklist.append_if_missing(orig_phi);
636 const TypePtr *atype = C->get_adr_type(alias_idx);
637 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
638 C->copy_node_notes_to(result, orig_phi);
639 igvn->set_type(result, result->bottom_type());
640 record_for_optimizer(result);
642 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;)
643 assert(pn == NULL || pn == orig_phi, "wrong node");
644 set_map(orig_phi->_idx, result);
645 ptnode_adr(orig_phi->_idx)->_node = orig_phi;
647 new_created = true;
648 return result;
649 }
651 //
652 // Return a new version of Memory Phi "orig_phi" with the inputs having the
653 // specified alias index.
654 //
655 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
657 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
658 Compile *C = _compile;
659 bool new_phi_created;
660 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
661 if (!new_phi_created) {
662 return result;
663 }
665 GrowableArray<PhiNode *> phi_list;
666 GrowableArray<uint> cur_input;
668 PhiNode *phi = orig_phi;
669 uint idx = 1;
670 bool finished = false;
671 while(!finished) {
672 while (idx < phi->req()) {
673 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
674 if (mem != NULL && mem->is_Phi()) {
675 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
676 if (new_phi_created) {
677 // found an phi for which we created a new split, push current one on worklist and begin
678 // processing new one
679 phi_list.push(phi);
680 cur_input.push(idx);
681 phi = mem->as_Phi();
682 result = newphi;
683 idx = 1;
684 continue;
685 } else {
686 mem = newphi;
687 }
688 }
689 if (C->failing()) {
690 return NULL;
691 }
692 result->set_req(idx++, mem);
693 }
694 #ifdef ASSERT
695 // verify that the new Phi has an input for each input of the original
696 assert( phi->req() == result->req(), "must have same number of inputs.");
697 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
698 #endif
699 // Check if all new phi's inputs have specified alias index.
700 // Otherwise use old phi.
701 for (uint i = 1; i < phi->req(); i++) {
702 Node* in = result->in(i);
703 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
704 }
705 // we have finished processing a Phi, see if there are any more to do
706 finished = (phi_list.length() == 0 );
707 if (!finished) {
708 phi = phi_list.pop();
709 idx = cur_input.pop();
710 PhiNode *prev_result = get_map_phi(phi->_idx);
711 prev_result->set_req(idx++, result);
712 result = prev_result;
713 }
714 }
715 return result;
716 }
719 //
720 // The next methods are derived from methods in MemNode.
721 //
722 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
723 Node *mem = mmem;
724 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
725 // means an array I have not precisely typed yet. Do not do any
726 // alias stuff with it any time soon.
727 if( toop->base() != Type::AnyPtr &&
728 !(toop->klass() != NULL &&
729 toop->klass()->is_java_lang_Object() &&
730 toop->offset() == Type::OffsetBot) ) {
731 mem = mmem->memory_at(alias_idx);
732 // Update input if it is progress over what we have now
733 }
734 return mem;
735 }
737 //
738 // Move memory users to their memory slices.
739 //
740 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) {
741 Compile* C = _compile;
743 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
744 assert(tp != NULL, "ptr type");
745 int alias_idx = C->get_alias_index(tp);
746 int general_idx = C->get_general_index(alias_idx);
748 // Move users first
749 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
750 Node* use = n->fast_out(i);
751 if (use->is_MergeMem()) {
752 MergeMemNode* mmem = use->as_MergeMem();
753 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
754 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
755 continue; // Nothing to do
756 }
757 // Replace previous general reference to mem node.
758 uint orig_uniq = C->unique();
759 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
760 assert(orig_uniq == C->unique(), "no new nodes");
761 mmem->set_memory_at(general_idx, m);
762 --imax;
763 --i;
764 } else if (use->is_MemBar()) {
765 assert(!use->is_Initialize(), "initializing stores should not be moved");
766 if (use->req() > MemBarNode::Precedent &&
767 use->in(MemBarNode::Precedent) == n) {
768 // Don't move related membars.
769 record_for_optimizer(use);
770 continue;
771 }
772 tp = use->as_MemBar()->adr_type()->isa_ptr();
773 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
774 alias_idx == general_idx) {
775 continue; // Nothing to do
776 }
777 // Move to general memory slice.
778 uint orig_uniq = C->unique();
779 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
780 assert(orig_uniq == C->unique(), "no new nodes");
781 igvn->hash_delete(use);
782 imax -= use->replace_edge(n, m);
783 igvn->hash_insert(use);
784 record_for_optimizer(use);
785 --i;
786 #ifdef ASSERT
787 } else if (use->is_Mem()) {
788 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
789 // Don't move related cardmark.
790 continue;
791 }
792 // Memory nodes should have new memory input.
793 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
794 assert(tp != NULL, "ptr type");
795 int idx = C->get_alias_index(tp);
796 assert(get_map(use->_idx) != NULL || idx == alias_idx,
797 "Following memory nodes should have new memory input or be on the same memory slice");
798 } else if (use->is_Phi()) {
799 // Phi nodes should be split and moved already.
800 tp = use->as_Phi()->adr_type()->isa_ptr();
801 assert(tp != NULL, "ptr type");
802 int idx = C->get_alias_index(tp);
803 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
804 } else {
805 use->dump();
806 assert(false, "should not be here");
807 #endif
808 }
809 }
810 }
812 //
813 // Search memory chain of "mem" to find a MemNode whose address
814 // is the specified alias index.
815 //
816 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
817 if (orig_mem == NULL)
818 return orig_mem;
819 Compile* C = phase->C;
820 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
821 bool is_instance = (toop != NULL) && toop->is_known_instance();
822 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
823 Node *prev = NULL;
824 Node *result = orig_mem;
825 while (prev != result) {
826 prev = result;
827 if (result == start_mem)
828 break; // hit one of our sentinels
829 if (result->is_Mem()) {
830 const Type *at = phase->type(result->in(MemNode::Address));
831 if (at == Type::TOP)
832 break; // Dead
833 assert (at->isa_ptr() != NULL, "pointer type required.");
834 int idx = C->get_alias_index(at->is_ptr());
835 if (idx == alias_idx)
836 break; // Found
837 if (!is_instance && (at->isa_oopptr() == NULL ||
838 !at->is_oopptr()->is_known_instance())) {
839 break; // Do not skip store to general memory slice.
840 }
841 result = result->in(MemNode::Memory);
842 }
843 if (!is_instance)
844 continue; // don't search further for non-instance types
845 // skip over a call which does not affect this memory slice
846 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
847 Node *proj_in = result->in(0);
848 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
849 break; // hit one of our sentinels
850 } else if (proj_in->is_Call()) {
851 CallNode *call = proj_in->as_Call();
852 if (!call->may_modify(toop, phase)) {
853 result = call->in(TypeFunc::Memory);
854 }
855 } else if (proj_in->is_Initialize()) {
856 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
857 // Stop if this is the initialization for the object instance which
858 // which contains this memory slice, otherwise skip over it.
859 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
860 result = proj_in->in(TypeFunc::Memory);
861 }
862 } else if (proj_in->is_MemBar()) {
863 result = proj_in->in(TypeFunc::Memory);
864 }
865 } else if (result->is_MergeMem()) {
866 MergeMemNode *mmem = result->as_MergeMem();
867 result = step_through_mergemem(mmem, alias_idx, toop);
868 if (result == mmem->base_memory()) {
869 // Didn't find instance memory, search through general slice recursively.
870 result = mmem->memory_at(C->get_general_index(alias_idx));
871 result = find_inst_mem(result, alias_idx, orig_phis, phase);
872 if (C->failing()) {
873 return NULL;
874 }
875 mmem->set_memory_at(alias_idx, result);
876 }
877 } else if (result->is_Phi() &&
878 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
879 Node *un = result->as_Phi()->unique_input(phase);
880 if (un != NULL) {
881 orig_phis.append_if_missing(result->as_Phi());
882 result = un;
883 } else {
884 break;
885 }
886 } else if (result->is_ClearArray()) {
887 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) {
888 // Can not bypass initialization of the instance
889 // we are looking for.
890 break;
891 }
892 // Otherwise skip it (the call updated 'result' value).
893 } else if (result->Opcode() == Op_SCMemProj) {
894 assert(result->in(0)->is_LoadStore(), "sanity");
895 const Type *at = phase->type(result->in(0)->in(MemNode::Address));
896 if (at != Type::TOP) {
897 assert (at->isa_ptr() != NULL, "pointer type required.");
898 int idx = C->get_alias_index(at->is_ptr());
899 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
900 break;
901 }
902 result = result->in(0)->in(MemNode::Memory);
903 }
904 }
905 if (result->is_Phi()) {
906 PhiNode *mphi = result->as_Phi();
907 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
908 const TypePtr *t = mphi->adr_type();
909 if (!is_instance) {
910 // Push all non-instance Phis on the orig_phis worklist to update inputs
911 // during Phase 4 if needed.
912 orig_phis.append_if_missing(mphi);
913 } else if (C->get_alias_index(t) != alias_idx) {
914 // Create a new Phi with the specified alias index type.
915 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
916 }
917 }
918 // the result is either MemNode, PhiNode, InitializeNode.
919 return result;
920 }
922 //
923 // Convert the types of unescaped object to instance types where possible,
924 // propagate the new type information through the graph, and update memory
925 // edges and MergeMem inputs to reflect the new type.
926 //
927 // We start with allocations (and calls which may be allocations) on alloc_worklist.
928 // The processing is done in 4 phases:
929 //
930 // Phase 1: Process possible allocations from alloc_worklist. Create instance
931 // types for the CheckCastPP for allocations where possible.
932 // Propagate the the new types through users as follows:
933 // casts and Phi: push users on alloc_worklist
934 // AddP: cast Base and Address inputs to the instance type
935 // push any AddP users on alloc_worklist and push any memnode
936 // users onto memnode_worklist.
937 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
938 // search the Memory chain for a store with the appropriate type
939 // address type. If a Phi is found, create a new version with
940 // the appropriate memory slices from each of the Phi inputs.
941 // For stores, process the users as follows:
942 // MemNode: push on memnode_worklist
943 // MergeMem: push on mergemem_worklist
944 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
945 // moving the first node encountered of each instance type to the
946 // the input corresponding to its alias index.
947 // appropriate memory slice.
948 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
949 //
950 // In the following example, the CheckCastPP nodes are the cast of allocation
951 // results and the allocation of node 29 is unescaped and eligible to be an
952 // instance type.
953 //
954 // We start with:
955 //
956 // 7 Parm #memory
957 // 10 ConI "12"
958 // 19 CheckCastPP "Foo"
959 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
960 // 29 CheckCastPP "Foo"
961 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
962 //
963 // 40 StoreP 25 7 20 ... alias_index=4
964 // 50 StoreP 35 40 30 ... alias_index=4
965 // 60 StoreP 45 50 20 ... alias_index=4
966 // 70 LoadP _ 60 30 ... alias_index=4
967 // 80 Phi 75 50 60 Memory alias_index=4
968 // 90 LoadP _ 80 30 ... alias_index=4
969 // 100 LoadP _ 80 20 ... alias_index=4
970 //
971 //
972 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
973 // and creating a new alias index for node 30. This gives:
974 //
975 // 7 Parm #memory
976 // 10 ConI "12"
977 // 19 CheckCastPP "Foo"
978 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
979 // 29 CheckCastPP "Foo" iid=24
980 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
981 //
982 // 40 StoreP 25 7 20 ... alias_index=4
983 // 50 StoreP 35 40 30 ... alias_index=6
984 // 60 StoreP 45 50 20 ... alias_index=4
985 // 70 LoadP _ 60 30 ... alias_index=6
986 // 80 Phi 75 50 60 Memory alias_index=4
987 // 90 LoadP _ 80 30 ... alias_index=6
988 // 100 LoadP _ 80 20 ... alias_index=4
989 //
990 // In phase 2, new memory inputs are computed for the loads and stores,
991 // And a new version of the phi is created. In phase 4, the inputs to
992 // node 80 are updated and then the memory nodes are updated with the
993 // values computed in phase 2. This results in:
994 //
995 // 7 Parm #memory
996 // 10 ConI "12"
997 // 19 CheckCastPP "Foo"
998 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
999 // 29 CheckCastPP "Foo" iid=24
1000 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
1001 //
1002 // 40 StoreP 25 7 20 ... alias_index=4
1003 // 50 StoreP 35 7 30 ... alias_index=6
1004 // 60 StoreP 45 40 20 ... alias_index=4
1005 // 70 LoadP _ 50 30 ... alias_index=6
1006 // 80 Phi 75 40 60 Memory alias_index=4
1007 // 120 Phi 75 50 50 Memory alias_index=6
1008 // 90 LoadP _ 120 30 ... alias_index=6
1009 // 100 LoadP _ 80 20 ... alias_index=4
1010 //
1011 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
1012 GrowableArray<Node *> memnode_worklist;
1013 GrowableArray<PhiNode *> orig_phis;
1015 PhaseIterGVN *igvn = _igvn;
1016 uint new_index_start = (uint) _compile->num_alias_types();
1017 Arena* arena = Thread::current()->resource_area();
1018 VectorSet visited(arena);
1021 // Phase 1: Process possible allocations from alloc_worklist.
1022 // Create instance types for the CheckCastPP for allocations where possible.
1023 //
1024 // (Note: don't forget to change the order of the second AddP node on
1025 // the alloc_worklist if the order of the worklist processing is changed,
1026 // see the comment in find_second_addp().)
1027 //
1028 while (alloc_worklist.length() != 0) {
1029 Node *n = alloc_worklist.pop();
1030 uint ni = n->_idx;
1031 const TypeOopPtr* tinst = NULL;
1032 if (n->is_Call()) {
1033 CallNode *alloc = n->as_Call();
1034 // copy escape information to call node
1035 PointsToNode* ptn = ptnode_adr(alloc->_idx);
1036 PointsToNode::EscapeState es = escape_state(alloc);
1037 // We have an allocation or call which returns a Java object,
1038 // see if it is unescaped.
1039 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable)
1040 continue;
1042 // Find CheckCastPP for the allocate or for the return value of a call
1043 n = alloc->result_cast();
1044 if (n == NULL) { // No uses except Initialize node
1045 if (alloc->is_Allocate()) {
1046 // Set the scalar_replaceable flag for allocation
1047 // so it could be eliminated if it has no uses.
1048 alloc->as_Allocate()->_is_scalar_replaceable = true;
1049 }
1050 continue;
1051 }
1052 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
1053 assert(!alloc->is_Allocate(), "allocation should have unique type");
1054 continue;
1055 }
1057 // The inline code for Object.clone() casts the allocation result to
1058 // java.lang.Object and then to the actual type of the allocated
1059 // object. Detect this case and use the second cast.
1060 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
1061 // the allocation result is cast to java.lang.Object and then
1062 // to the actual Array type.
1063 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
1064 && (alloc->is_AllocateArray() ||
1065 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
1066 Node *cast2 = NULL;
1067 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1068 Node *use = n->fast_out(i);
1069 if (use->is_CheckCastPP()) {
1070 cast2 = use;
1071 break;
1072 }
1073 }
1074 if (cast2 != NULL) {
1075 n = cast2;
1076 } else {
1077 // Non-scalar replaceable if the allocation type is unknown statically
1078 // (reflection allocation), the object can't be restored during
1079 // deoptimization without precise type.
1080 continue;
1081 }
1082 }
1083 if (alloc->is_Allocate()) {
1084 // Set the scalar_replaceable flag for allocation
1085 // so it could be eliminated.
1086 alloc->as_Allocate()->_is_scalar_replaceable = true;
1087 }
1088 set_escape_state(n->_idx, es);
1089 // in order for an object to be scalar-replaceable, it must be:
1090 // - a direct allocation (not a call returning an object)
1091 // - non-escaping
1092 // - eligible to be a unique type
1093 // - not determined to be ineligible by escape analysis
1094 assert(ptnode_adr(alloc->_idx)->_node != NULL &&
1095 ptnode_adr(n->_idx)->_node != NULL, "should be registered");
1096 set_map(alloc->_idx, n);
1097 set_map(n->_idx, alloc);
1098 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
1099 if (t == NULL)
1100 continue; // not a TypeInstPtr
1101 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
1102 igvn->hash_delete(n);
1103 igvn->set_type(n, tinst);
1104 n->raise_bottom_type(tinst);
1105 igvn->hash_insert(n);
1106 record_for_optimizer(n);
1107 if (alloc->is_Allocate() && ptn->_scalar_replaceable &&
1108 (t->isa_instptr() || t->isa_aryptr())) {
1110 // First, put on the worklist all Field edges from Connection Graph
1111 // which is more accurate then putting immediate users from Ideal Graph.
1112 for (uint e = 0; e < ptn->edge_count(); e++) {
1113 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
1114 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
1115 "only AddP nodes are Field edges in CG");
1116 if (use->outcnt() > 0) { // Don't process dead nodes
1117 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
1118 if (addp2 != NULL) {
1119 assert(alloc->is_AllocateArray(),"array allocation was expected");
1120 alloc_worklist.append_if_missing(addp2);
1121 }
1122 alloc_worklist.append_if_missing(use);
1123 }
1124 }
1126 // An allocation may have an Initialize which has raw stores. Scan
1127 // the users of the raw allocation result and push AddP users
1128 // on alloc_worklist.
1129 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
1130 assert (raw_result != NULL, "must have an allocation result");
1131 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
1132 Node *use = raw_result->fast_out(i);
1133 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
1134 Node* addp2 = find_second_addp(use, raw_result);
1135 if (addp2 != NULL) {
1136 assert(alloc->is_AllocateArray(),"array allocation was expected");
1137 alloc_worklist.append_if_missing(addp2);
1138 }
1139 alloc_worklist.append_if_missing(use);
1140 } else if (use->is_MemBar()) {
1141 memnode_worklist.append_if_missing(use);
1142 }
1143 }
1144 }
1145 } else if (n->is_AddP()) {
1146 VectorSet* ptset = PointsTo(get_addp_base(n));
1147 assert(ptset->Size() == 1, "AddP address is unique");
1148 uint elem = ptset->getelem(); // Allocation node's index
1149 if (elem == _phantom_object) {
1150 assert(false, "escaped allocation");
1151 continue; // Assume the value was set outside this method.
1152 }
1153 Node *base = get_map(elem); // CheckCastPP node
1154 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path
1155 tinst = igvn->type(base)->isa_oopptr();
1156 } else if (n->is_Phi() ||
1157 n->is_CheckCastPP() ||
1158 n->is_EncodeP() ||
1159 n->is_DecodeN() ||
1160 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
1161 if (visited.test_set(n->_idx)) {
1162 assert(n->is_Phi(), "loops only through Phi's");
1163 continue; // already processed
1164 }
1165 VectorSet* ptset = PointsTo(n);
1166 if (ptset->Size() == 1) {
1167 uint elem = ptset->getelem(); // Allocation node's index
1168 if (elem == _phantom_object) {
1169 assert(false, "escaped allocation");
1170 continue; // Assume the value was set outside this method.
1171 }
1172 Node *val = get_map(elem); // CheckCastPP node
1173 TypeNode *tn = n->as_Type();
1174 tinst = igvn->type(val)->isa_oopptr();
1175 assert(tinst != NULL && tinst->is_known_instance() &&
1176 (uint)tinst->instance_id() == elem , "instance type expected.");
1178 const Type *tn_type = igvn->type(tn);
1179 const TypeOopPtr *tn_t;
1180 if (tn_type->isa_narrowoop()) {
1181 tn_t = tn_type->make_ptr()->isa_oopptr();
1182 } else {
1183 tn_t = tn_type->isa_oopptr();
1184 }
1186 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
1187 if (tn_type->isa_narrowoop()) {
1188 tn_type = tinst->make_narrowoop();
1189 } else {
1190 tn_type = tinst;
1191 }
1192 igvn->hash_delete(tn);
1193 igvn->set_type(tn, tn_type);
1194 tn->set_type(tn_type);
1195 igvn->hash_insert(tn);
1196 record_for_optimizer(n);
1197 } else {
1198 assert(tn_type == TypePtr::NULL_PTR ||
1199 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
1200 "unexpected type");
1201 continue; // Skip dead path with different type
1202 }
1203 }
1204 } else {
1205 debug_only(n->dump();)
1206 assert(false, "EA: unexpected node");
1207 continue;
1208 }
1209 // push allocation's users on appropriate worklist
1210 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1211 Node *use = n->fast_out(i);
1212 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1213 // Load/store to instance's field
1214 memnode_worklist.append_if_missing(use);
1215 } else if (use->is_MemBar()) {
1216 memnode_worklist.append_if_missing(use);
1217 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1218 Node* addp2 = find_second_addp(use, n);
1219 if (addp2 != NULL) {
1220 alloc_worklist.append_if_missing(addp2);
1221 }
1222 alloc_worklist.append_if_missing(use);
1223 } else if (use->is_Phi() ||
1224 use->is_CheckCastPP() ||
1225 use->is_EncodeP() ||
1226 use->is_DecodeN() ||
1227 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1228 alloc_worklist.append_if_missing(use);
1229 #ifdef ASSERT
1230 } else if (use->is_Mem()) {
1231 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
1232 } else if (use->is_MergeMem()) {
1233 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1234 } else if (use->is_SafePoint()) {
1235 // Look for MergeMem nodes for calls which reference unique allocation
1236 // (through CheckCastPP nodes) even for debug info.
1237 Node* m = use->in(TypeFunc::Memory);
1238 if (m->is_MergeMem()) {
1239 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1240 }
1241 } else {
1242 uint op = use->Opcode();
1243 if (!(op == Op_CmpP || op == Op_Conv2B ||
1244 op == Op_CastP2X || op == Op_StoreCM ||
1245 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
1246 op == Op_StrEquals || op == Op_StrIndexOf)) {
1247 n->dump();
1248 use->dump();
1249 assert(false, "EA: missing allocation reference path");
1250 }
1251 #endif
1252 }
1253 }
1255 }
1256 // New alias types were created in split_AddP().
1257 uint new_index_end = (uint) _compile->num_alias_types();
1259 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1260 // compute new values for Memory inputs (the Memory inputs are not
1261 // actually updated until phase 4.)
1262 if (memnode_worklist.length() == 0)
1263 return; // nothing to do
1265 while (memnode_worklist.length() != 0) {
1266 Node *n = memnode_worklist.pop();
1267 if (visited.test_set(n->_idx))
1268 continue;
1269 if (n->is_Phi() || n->is_ClearArray()) {
1270 // we don't need to do anything, but the users must be pushed
1271 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
1272 // we don't need to do anything, but the users must be pushed
1273 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
1274 if (n == NULL)
1275 continue;
1276 } else {
1277 assert(n->is_Mem(), "memory node required.");
1278 Node *addr = n->in(MemNode::Address);
1279 const Type *addr_t = igvn->type(addr);
1280 if (addr_t == Type::TOP)
1281 continue;
1282 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1283 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1284 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1285 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1286 if (_compile->failing()) {
1287 return;
1288 }
1289 if (mem != n->in(MemNode::Memory)) {
1290 // We delay the memory edge update since we need old one in
1291 // MergeMem code below when instances memory slices are separated.
1292 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;)
1293 assert(pn == NULL || pn == n, "wrong node");
1294 set_map(n->_idx, mem);
1295 ptnode_adr(n->_idx)->_node = n;
1296 }
1297 if (n->is_Load()) {
1298 continue; // don't push users
1299 } else if (n->is_LoadStore()) {
1300 // get the memory projection
1301 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1302 Node *use = n->fast_out(i);
1303 if (use->Opcode() == Op_SCMemProj) {
1304 n = use;
1305 break;
1306 }
1307 }
1308 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1309 }
1310 }
1311 // push user on appropriate worklist
1312 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1313 Node *use = n->fast_out(i);
1314 if (use->is_Phi() || use->is_ClearArray()) {
1315 memnode_worklist.append_if_missing(use);
1316 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1317 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
1318 continue;
1319 memnode_worklist.append_if_missing(use);
1320 } else if (use->is_MemBar()) {
1321 memnode_worklist.append_if_missing(use);
1322 #ifdef ASSERT
1323 } else if(use->is_Mem()) {
1324 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
1325 } else if (use->is_MergeMem()) {
1326 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1327 } else {
1328 uint op = use->Opcode();
1329 if (!(op == Op_StoreCM ||
1330 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
1331 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
1332 op == Op_AryEq || op == Op_StrComp ||
1333 op == Op_StrEquals || op == Op_StrIndexOf)) {
1334 n->dump();
1335 use->dump();
1336 assert(false, "EA: missing memory path");
1337 }
1338 #endif
1339 }
1340 }
1341 }
1343 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1344 // Walk each memory slice moving the first node encountered of each
1345 // instance type to the the input corresponding to its alias index.
1346 uint length = _mergemem_worklist.length();
1347 for( uint next = 0; next < length; ++next ) {
1348 MergeMemNode* nmm = _mergemem_worklist.at(next);
1349 assert(!visited.test_set(nmm->_idx), "should not be visited before");
1350 // Note: we don't want to use MergeMemStream here because we only want to
1351 // scan inputs which exist at the start, not ones we add during processing.
1352 // Note 2: MergeMem may already contains instance memory slices added
1353 // during find_inst_mem() call when memory nodes were processed above.
1354 igvn->hash_delete(nmm);
1355 uint nslices = nmm->req();
1356 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1357 Node* mem = nmm->in(i);
1358 Node* cur = NULL;
1359 if (mem == NULL || mem->is_top())
1360 continue;
1361 // First, update mergemem by moving memory nodes to corresponding slices
1362 // if their type became more precise since this mergemem was created.
1363 while (mem->is_Mem()) {
1364 const Type *at = igvn->type(mem->in(MemNode::Address));
1365 if (at != Type::TOP) {
1366 assert (at->isa_ptr() != NULL, "pointer type required.");
1367 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1368 if (idx == i) {
1369 if (cur == NULL)
1370 cur = mem;
1371 } else {
1372 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1373 nmm->set_memory_at(idx, mem);
1374 }
1375 }
1376 }
1377 mem = mem->in(MemNode::Memory);
1378 }
1379 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1380 // Find any instance of the current type if we haven't encountered
1381 // already a memory slice of the instance along the memory chain.
1382 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1383 if((uint)_compile->get_general_index(ni) == i) {
1384 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1385 if (nmm->is_empty_memory(m)) {
1386 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1387 if (_compile->failing()) {
1388 return;
1389 }
1390 nmm->set_memory_at(ni, result);
1391 }
1392 }
1393 }
1394 }
1395 // Find the rest of instances values
1396 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1397 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
1398 Node* result = step_through_mergemem(nmm, ni, tinst);
1399 if (result == nmm->base_memory()) {
1400 // Didn't find instance memory, search through general slice recursively.
1401 result = nmm->memory_at(_compile->get_general_index(ni));
1402 result = find_inst_mem(result, ni, orig_phis, igvn);
1403 if (_compile->failing()) {
1404 return;
1405 }
1406 nmm->set_memory_at(ni, result);
1407 }
1408 }
1409 igvn->hash_insert(nmm);
1410 record_for_optimizer(nmm);
1411 }
1413 // Phase 4: Update the inputs of non-instance memory Phis and
1414 // the Memory input of memnodes
1415 // First update the inputs of any non-instance Phi's from
1416 // which we split out an instance Phi. Note we don't have
1417 // to recursively process Phi's encounted on the input memory
1418 // chains as is done in split_memory_phi() since they will
1419 // also be processed here.
1420 for (int j = 0; j < orig_phis.length(); j++) {
1421 PhiNode *phi = orig_phis.at(j);
1422 int alias_idx = _compile->get_alias_index(phi->adr_type());
1423 igvn->hash_delete(phi);
1424 for (uint i = 1; i < phi->req(); i++) {
1425 Node *mem = phi->in(i);
1426 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1427 if (_compile->failing()) {
1428 return;
1429 }
1430 if (mem != new_mem) {
1431 phi->set_req(i, new_mem);
1432 }
1433 }
1434 igvn->hash_insert(phi);
1435 record_for_optimizer(phi);
1436 }
1438 // Update the memory inputs of MemNodes with the value we computed
1439 // in Phase 2 and move stores memory users to corresponding memory slices.
1441 // Disable memory split verification code until the fix for 6984348.
1442 // Currently it produces false negative results since it does not cover all cases.
1443 #if 0 // ifdef ASSERT
1444 visited.Reset();
1445 Node_Stack old_mems(arena, _compile->unique() >> 2);
1446 #endif
1447 for (uint i = 0; i < nodes_size(); i++) {
1448 Node *nmem = get_map(i);
1449 if (nmem != NULL) {
1450 Node *n = ptnode_adr(i)->_node;
1451 assert(n != NULL, "sanity");
1452 if (n->is_Mem()) {
1453 #if 0 // ifdef ASSERT
1454 Node* old_mem = n->in(MemNode::Memory);
1455 if (!visited.test_set(old_mem->_idx)) {
1456 old_mems.push(old_mem, old_mem->outcnt());
1457 }
1458 #endif
1459 assert(n->in(MemNode::Memory) != nmem, "sanity");
1460 if (!n->is_Load()) {
1461 // Move memory users of a store first.
1462 move_inst_mem(n, orig_phis, igvn);
1463 }
1464 // Now update memory input
1465 igvn->hash_delete(n);
1466 n->set_req(MemNode::Memory, nmem);
1467 igvn->hash_insert(n);
1468 record_for_optimizer(n);
1469 } else {
1470 assert(n->is_Allocate() || n->is_CheckCastPP() ||
1471 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
1472 }
1473 }
1474 }
1475 #if 0 // ifdef ASSERT
1476 // Verify that memory was split correctly
1477 while (old_mems.is_nonempty()) {
1478 Node* old_mem = old_mems.node();
1479 uint old_cnt = old_mems.index();
1480 old_mems.pop();
1481 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
1482 }
1483 #endif
1484 }
1486 bool ConnectionGraph::has_candidates(Compile *C) {
1487 // EA brings benefits only when the code has allocations and/or locks which
1488 // are represented by ideal Macro nodes.
1489 int cnt = C->macro_count();
1490 for( int i=0; i < cnt; i++ ) {
1491 Node *n = C->macro_node(i);
1492 if ( n->is_Allocate() )
1493 return true;
1494 if( n->is_Lock() ) {
1495 Node* obj = n->as_Lock()->obj_node()->uncast();
1496 if( !(obj->is_Parm() || obj->is_Con()) )
1497 return true;
1498 }
1499 }
1500 return false;
1501 }
1503 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
1504 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
1505 // to create space for them in ConnectionGraph::_nodes[].
1506 Node* oop_null = igvn->zerocon(T_OBJECT);
1507 Node* noop_null = igvn->zerocon(T_NARROWOOP);
1509 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
1510 // Perform escape analysis
1511 if (congraph->compute_escape()) {
1512 // There are non escaping objects.
1513 C->set_congraph(congraph);
1514 }
1516 // Cleanup.
1517 if (oop_null->outcnt() == 0)
1518 igvn->hash_delete(oop_null);
1519 if (noop_null->outcnt() == 0)
1520 igvn->hash_delete(noop_null);
1521 }
1523 bool ConnectionGraph::compute_escape() {
1524 Compile* C = _compile;
1526 // 1. Populate Connection Graph (CG) with Ideal nodes.
1528 Unique_Node_List worklist_init;
1529 worklist_init.map(C->unique(), NULL); // preallocate space
1531 // Initialize worklist
1532 if (C->root() != NULL) {
1533 worklist_init.push(C->root());
1534 }
1536 GrowableArray<int> cg_worklist;
1537 PhaseGVN* igvn = _igvn;
1538 bool has_allocations = false;
1540 // Push all useful nodes onto CG list and set their type.
1541 for( uint next = 0; next < worklist_init.size(); ++next ) {
1542 Node* n = worklist_init.at(next);
1543 record_for_escape_analysis(n, igvn);
1544 // Only allocations and java static calls results are checked
1545 // for an escape status. See process_call_result() below.
1546 if (n->is_Allocate() || n->is_CallStaticJava() &&
1547 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1548 has_allocations = true;
1549 }
1550 if(n->is_AddP()) {
1551 // Collect address nodes. Use them during stage 3 below
1552 // to build initial connection graph field edges.
1553 cg_worklist.append(n->_idx);
1554 } else if (n->is_MergeMem()) {
1555 // Collect all MergeMem nodes to add memory slices for
1556 // scalar replaceable objects in split_unique_types().
1557 _mergemem_worklist.append(n->as_MergeMem());
1558 }
1559 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1560 Node* m = n->fast_out(i); // Get user
1561 worklist_init.push(m);
1562 }
1563 }
1565 if (!has_allocations) {
1566 _collecting = false;
1567 return false; // Nothing to do.
1568 }
1570 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1571 uint delayed_size = _delayed_worklist.size();
1572 for( uint next = 0; next < delayed_size; ++next ) {
1573 Node* n = _delayed_worklist.at(next);
1574 build_connection_graph(n, igvn);
1575 }
1577 // 3. Pass to create initial fields edges (JavaObject -F-> AddP)
1578 // to reduce number of iterations during stage 4 below.
1579 uint cg_length = cg_worklist.length();
1580 for( uint next = 0; next < cg_length; ++next ) {
1581 int ni = cg_worklist.at(next);
1582 Node* n = ptnode_adr(ni)->_node;
1583 Node* base = get_addp_base(n);
1584 if (base->is_Proj())
1585 base = base->in(0);
1586 PointsToNode::NodeType nt = ptnode_adr(base->_idx)->node_type();
1587 if (nt == PointsToNode::JavaObject) {
1588 build_connection_graph(n, igvn);
1589 }
1590 }
1592 cg_worklist.clear();
1593 cg_worklist.append(_phantom_object);
1594 GrowableArray<uint> worklist;
1596 // 4. Build Connection Graph which need
1597 // to walk the connection graph.
1598 _progress = false;
1599 for (uint ni = 0; ni < nodes_size(); ni++) {
1600 PointsToNode* ptn = ptnode_adr(ni);
1601 Node *n = ptn->_node;
1602 if (n != NULL) { // Call, AddP, LoadP, StoreP
1603 build_connection_graph(n, igvn);
1604 if (ptn->node_type() != PointsToNode::UnknownType)
1605 cg_worklist.append(n->_idx); // Collect CG nodes
1606 if (!_processed.test(n->_idx))
1607 worklist.append(n->_idx); // Collect C/A/L/S nodes
1608 }
1609 }
1611 // After IGVN user nodes may have smaller _idx than
1612 // their inputs so they will be processed first in
1613 // previous loop. Because of that not all Graph
1614 // edges will be created. Walk over interesting
1615 // nodes again until no new edges are created.
1616 //
1617 // Normally only 1-3 passes needed to build
1618 // Connection Graph depending on graph complexity.
1619 // Observed 8 passes in jvm2008 compiler.compiler.
1620 // Set limit to 20 to catch situation when something
1621 // did go wrong and recompile the method without EA.
1623 #define CG_BUILD_ITER_LIMIT 20
1625 uint length = worklist.length();
1626 int iterations = 0;
1627 while(_progress && (iterations++ < CG_BUILD_ITER_LIMIT)) {
1628 _progress = false;
1629 for( uint next = 0; next < length; ++next ) {
1630 int ni = worklist.at(next);
1631 PointsToNode* ptn = ptnode_adr(ni);
1632 Node* n = ptn->_node;
1633 assert(n != NULL, "should be known node");
1634 build_connection_graph(n, igvn);
1635 }
1636 }
1637 if (iterations >= CG_BUILD_ITER_LIMIT) {
1638 assert(iterations < CG_BUILD_ITER_LIMIT,
1639 err_msg("infinite EA connection graph build with %d nodes and worklist size %d",
1640 nodes_size(), length));
1641 // Possible infinite build_connection_graph loop,
1642 // retry compilation without escape analysis.
1643 C->record_failure(C2Compiler::retry_no_escape_analysis());
1644 _collecting = false;
1645 return false;
1646 }
1647 #undef CG_BUILD_ITER_LIMIT
1649 Arena* arena = Thread::current()->resource_area();
1650 VectorSet visited(arena);
1651 worklist.clear();
1653 // 5. Remove deferred edges from the graph and adjust
1654 // escape state of nonescaping objects.
1655 cg_length = cg_worklist.length();
1656 for( uint next = 0; next < cg_length; ++next ) {
1657 int ni = cg_worklist.at(next);
1658 PointsToNode* ptn = ptnode_adr(ni);
1659 PointsToNode::NodeType nt = ptn->node_type();
1660 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1661 remove_deferred(ni, &worklist, &visited);
1662 Node *n = ptn->_node;
1663 if (n->is_AddP()) {
1664 // Search for objects which are not scalar replaceable
1665 // and adjust their escape state.
1666 adjust_escape_state(ni, igvn);
1667 }
1668 }
1669 }
1671 // 6. Propagate escape states.
1672 worklist.clear();
1673 bool has_non_escaping_obj = false;
1675 // push all GlobalEscape nodes on the worklist
1676 for( uint next = 0; next < cg_length; ++next ) {
1677 int nk = cg_worklist.at(next);
1678 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape)
1679 worklist.push(nk);
1680 }
1681 // mark all nodes reachable from GlobalEscape nodes
1682 while(worklist.length() > 0) {
1683 PointsToNode* ptn = ptnode_adr(worklist.pop());
1684 uint e_cnt = ptn->edge_count();
1685 for (uint ei = 0; ei < e_cnt; ei++) {
1686 uint npi = ptn->edge_target(ei);
1687 PointsToNode *np = ptnode_adr(npi);
1688 if (np->escape_state() < PointsToNode::GlobalEscape) {
1689 np->set_escape_state(PointsToNode::GlobalEscape);
1690 worklist.push(npi);
1691 }
1692 }
1693 }
1695 // push all ArgEscape nodes on the worklist
1696 for( uint next = 0; next < cg_length; ++next ) {
1697 int nk = cg_worklist.at(next);
1698 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape)
1699 worklist.push(nk);
1700 }
1701 // mark all nodes reachable from ArgEscape nodes
1702 while(worklist.length() > 0) {
1703 PointsToNode* ptn = ptnode_adr(worklist.pop());
1704 if (ptn->node_type() == PointsToNode::JavaObject)
1705 has_non_escaping_obj = true; // Non GlobalEscape
1706 uint e_cnt = ptn->edge_count();
1707 for (uint ei = 0; ei < e_cnt; ei++) {
1708 uint npi = ptn->edge_target(ei);
1709 PointsToNode *np = ptnode_adr(npi);
1710 if (np->escape_state() < PointsToNode::ArgEscape) {
1711 np->set_escape_state(PointsToNode::ArgEscape);
1712 worklist.push(npi);
1713 }
1714 }
1715 }
1717 GrowableArray<Node*> alloc_worklist;
1719 // push all NoEscape nodes on the worklist
1720 for( uint next = 0; next < cg_length; ++next ) {
1721 int nk = cg_worklist.at(next);
1722 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape)
1723 worklist.push(nk);
1724 }
1725 // mark all nodes reachable from NoEscape nodes
1726 while(worklist.length() > 0) {
1727 PointsToNode* ptn = ptnode_adr(worklist.pop());
1728 if (ptn->node_type() == PointsToNode::JavaObject)
1729 has_non_escaping_obj = true; // Non GlobalEscape
1730 Node* n = ptn->_node;
1731 if (n->is_Allocate() && ptn->_scalar_replaceable ) {
1732 // Push scalar replaceable allocations on alloc_worklist
1733 // for processing in split_unique_types().
1734 alloc_worklist.append(n);
1735 }
1736 uint e_cnt = ptn->edge_count();
1737 for (uint ei = 0; ei < e_cnt; ei++) {
1738 uint npi = ptn->edge_target(ei);
1739 PointsToNode *np = ptnode_adr(npi);
1740 if (np->escape_state() < PointsToNode::NoEscape) {
1741 np->set_escape_state(PointsToNode::NoEscape);
1742 worklist.push(npi);
1743 }
1744 }
1745 }
1747 _collecting = false;
1748 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1750 if (EliminateLocks) {
1751 // Mark locks before changing ideal graph.
1752 int cnt = C->macro_count();
1753 for( int i=0; i < cnt; i++ ) {
1754 Node *n = C->macro_node(i);
1755 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1756 AbstractLockNode* alock = n->as_AbstractLock();
1757 if (!alock->is_eliminated()) {
1758 PointsToNode::EscapeState es = escape_state(alock->obj_node());
1759 assert(es != PointsToNode::UnknownEscape, "should know");
1760 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1761 // Mark it eliminated
1762 alock->set_eliminated();
1763 }
1764 }
1765 }
1766 }
1767 }
1769 #ifndef PRODUCT
1770 if (PrintEscapeAnalysis) {
1771 dump(); // Dump ConnectionGraph
1772 }
1773 #endif
1775 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0;
1776 if ( has_scalar_replaceable_candidates &&
1777 C->AliasLevel() >= 3 && EliminateAllocations ) {
1779 // Now use the escape information to create unique types for
1780 // scalar replaceable objects.
1781 split_unique_types(alloc_worklist);
1783 if (C->failing()) return false;
1785 C->print_method("After Escape Analysis", 2);
1787 #ifdef ASSERT
1788 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1789 tty->print("=== No allocations eliminated for ");
1790 C->method()->print_short_name();
1791 if(!EliminateAllocations) {
1792 tty->print(" since EliminateAllocations is off ===");
1793 } else if(!has_scalar_replaceable_candidates) {
1794 tty->print(" since there are no scalar replaceable candidates ===");
1795 } else if(C->AliasLevel() < 3) {
1796 tty->print(" since AliasLevel < 3 ===");
1797 }
1798 tty->cr();
1799 #endif
1800 }
1801 return has_non_escaping_obj;
1802 }
1804 // Adjust escape state after Connection Graph is built.
1805 void ConnectionGraph::adjust_escape_state(int nidx, PhaseTransform* phase) {
1806 PointsToNode* ptn = ptnode_adr(nidx);
1807 Node* n = ptn->_node;
1808 assert(n->is_AddP(), "Should be called for AddP nodes only");
1809 // Search for objects which are not scalar replaceable.
1810 // Mark their escape state as ArgEscape to propagate the state
1811 // to referenced objects.
1812 // Note: currently there are no difference in compiler optimizations
1813 // for ArgEscape objects and NoEscape objects which are not
1814 // scalar replaceable.
1816 Compile* C = _compile;
1818 int offset = ptn->offset();
1819 Node* base = get_addp_base(n);
1820 VectorSet* ptset = PointsTo(base);
1821 int ptset_size = ptset->Size();
1823 // Check if a oop field's initializing value is recorded and add
1824 // a corresponding NULL field's value if it is not recorded.
1825 // Connection Graph does not record a default initialization by NULL
1826 // captured by Initialize node.
1827 //
1828 // Note: it will disable scalar replacement in some cases:
1829 //
1830 // Point p[] = new Point[1];
1831 // p[0] = new Point(); // Will be not scalar replaced
1832 //
1833 // but it will save us from incorrect optimizations in next cases:
1834 //
1835 // Point p[] = new Point[1];
1836 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1837 //
1838 // Do a simple control flow analysis to distinguish above cases.
1839 //
1840 if (offset != Type::OffsetBot && ptset_size == 1) {
1841 uint elem = ptset->getelem(); // Allocation node's index
1842 // It does not matter if it is not Allocation node since
1843 // only non-escaping allocations are scalar replaced.
1844 if (ptnode_adr(elem)->_node->is_Allocate() &&
1845 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) {
1846 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate();
1847 InitializeNode* ini = alloc->initialization();
1849 // Check only oop fields.
1850 const Type* adr_type = n->as_AddP()->bottom_type();
1851 BasicType basic_field_type = T_INT;
1852 if (adr_type->isa_instptr()) {
1853 ciField* field = C->alias_type(adr_type->isa_instptr())->field();
1854 if (field != NULL) {
1855 basic_field_type = field->layout_type();
1856 } else {
1857 // Ignore non field load (for example, klass load)
1858 }
1859 } else if (adr_type->isa_aryptr()) {
1860 const Type* elemtype = adr_type->isa_aryptr()->elem();
1861 basic_field_type = elemtype->array_element_basic_type();
1862 } else {
1863 // Raw pointers are used for initializing stores so skip it.
1864 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1865 (base->in(0) == alloc),"unexpected pointer type");
1866 }
1867 if (basic_field_type == T_OBJECT ||
1868 basic_field_type == T_NARROWOOP ||
1869 basic_field_type == T_ARRAY) {
1870 Node* value = NULL;
1871 if (ini != NULL) {
1872 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
1873 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase);
1874 if (store != NULL && store->is_Store()) {
1875 value = store->in(MemNode::ValueIn);
1876 } else if (ptn->edge_count() > 0) { // Are there oop stores?
1877 // Check for a store which follows allocation without branches.
1878 // For example, a volatile field store is not collected
1879 // by Initialize node. TODO: it would be nice to use idom() here.
1880 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1881 store = n->fast_out(i);
1882 if (store->is_Store() && store->in(0) != NULL) {
1883 Node* ctrl = store->in(0);
1884 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL ||
1885 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() ||
1886 ctrl->is_IfTrue() || ctrl->is_IfFalse())) {
1887 ctrl = ctrl->in(0);
1888 }
1889 if (ctrl == ini || ctrl == alloc) {
1890 value = store->in(MemNode::ValueIn);
1891 break;
1892 }
1893 }
1894 }
1895 }
1896 }
1897 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
1898 // A field's initializing value was not recorded. Add NULL.
1899 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1900 add_pointsto_edge(nidx, null_idx);
1901 }
1902 }
1903 }
1904 }
1906 // An object is not scalar replaceable if the field which may point
1907 // to it has unknown offset (unknown element of an array of objects).
1908 //
1909 if (offset == Type::OffsetBot) {
1910 uint e_cnt = ptn->edge_count();
1911 for (uint ei = 0; ei < e_cnt; ei++) {
1912 uint npi = ptn->edge_target(ei);
1913 set_escape_state(npi, PointsToNode::ArgEscape);
1914 ptnode_adr(npi)->_scalar_replaceable = false;
1915 }
1916 }
1918 // Currently an object is not scalar replaceable if a LoadStore node
1919 // access its field since the field value is unknown after it.
1920 //
1921 bool has_LoadStore = false;
1922 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1923 Node *use = n->fast_out(i);
1924 if (use->is_LoadStore()) {
1925 has_LoadStore = true;
1926 break;
1927 }
1928 }
1929 // An object is not scalar replaceable if the address points
1930 // to unknown field (unknown element for arrays, offset is OffsetBot).
1931 //
1932 // Or the address may point to more then one object. This may produce
1933 // the false positive result (set scalar_replaceable to false)
1934 // since the flow-insensitive escape analysis can't separate
1935 // the case when stores overwrite the field's value from the case
1936 // when stores happened on different control branches.
1937 //
1938 if (ptset_size > 1 || ptset_size != 0 &&
1939 (has_LoadStore || offset == Type::OffsetBot)) {
1940 for( VectorSetI j(ptset); j.test(); ++j ) {
1941 set_escape_state(j.elem, PointsToNode::ArgEscape);
1942 ptnode_adr(j.elem)->_scalar_replaceable = false;
1943 }
1944 }
1945 }
1947 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
1949 switch (call->Opcode()) {
1950 #ifdef ASSERT
1951 case Op_Allocate:
1952 case Op_AllocateArray:
1953 case Op_Lock:
1954 case Op_Unlock:
1955 assert(false, "should be done already");
1956 break;
1957 #endif
1958 case Op_CallLeaf:
1959 case Op_CallLeafNoFP:
1960 {
1961 // Stub calls, objects do not escape but they are not scale replaceable.
1962 // Adjust escape state for outgoing arguments.
1963 const TypeTuple * d = call->tf()->domain();
1964 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1965 const Type* at = d->field_at(i);
1966 Node *arg = call->in(i)->uncast();
1967 const Type *aat = phase->type(arg);
1968 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() &&
1969 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) {
1971 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
1972 aat->isa_ptr() != NULL, "expecting an Ptr");
1973 #ifdef ASSERT
1974 if (!(call->Opcode() == Op_CallLeafNoFP &&
1975 call->as_CallLeaf()->_name != NULL &&
1976 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) ||
1977 call->as_CallLeaf()->_name != NULL &&
1978 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
1979 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ))
1980 ) {
1981 call->dump();
1982 assert(false, "EA: unexpected CallLeaf");
1983 }
1984 #endif
1985 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1986 if (arg->is_AddP()) {
1987 //
1988 // The inline_native_clone() case when the arraycopy stub is called
1989 // after the allocation before Initialize and CheckCastPP nodes.
1990 //
1991 // Set AddP's base (Allocate) as not scalar replaceable since
1992 // pointer to the base (with offset) is passed as argument.
1993 //
1994 arg = get_addp_base(arg);
1995 }
1996 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
1997 uint pt = j.elem;
1998 set_escape_state(pt, PointsToNode::ArgEscape);
1999 }
2000 }
2001 }
2002 break;
2003 }
2005 case Op_CallStaticJava:
2006 // For a static call, we know exactly what method is being called.
2007 // Use bytecode estimator to record the call's escape affects
2008 {
2009 ciMethod *meth = call->as_CallJava()->method();
2010 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
2011 // fall-through if not a Java method or no analyzer information
2012 if (call_analyzer != NULL) {
2013 const TypeTuple * d = call->tf()->domain();
2014 bool copy_dependencies = false;
2015 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2016 const Type* at = d->field_at(i);
2017 int k = i - TypeFunc::Parms;
2018 Node *arg = call->in(i)->uncast();
2020 if (at->isa_oopptr() != NULL &&
2021 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) {
2023 bool global_escapes = false;
2024 bool fields_escapes = false;
2025 if (!call_analyzer->is_arg_stack(k)) {
2026 // The argument global escapes, mark everything it could point to
2027 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2028 global_escapes = true;
2029 } else {
2030 if (!call_analyzer->is_arg_local(k)) {
2031 // The argument itself doesn't escape, but any fields might
2032 fields_escapes = true;
2033 }
2034 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
2035 copy_dependencies = true;
2036 }
2038 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2039 uint pt = j.elem;
2040 if (global_escapes) {
2041 //The argument global escapes, mark everything it could point to
2042 set_escape_state(pt, PointsToNode::GlobalEscape);
2043 } else {
2044 if (fields_escapes) {
2045 // The argument itself doesn't escape, but any fields might
2046 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
2047 }
2048 set_escape_state(pt, PointsToNode::ArgEscape);
2049 }
2050 }
2051 }
2052 }
2053 if (copy_dependencies)
2054 call_analyzer->copy_dependencies(_compile->dependencies());
2055 break;
2056 }
2057 }
2059 default:
2060 // Fall-through here if not a Java method or no analyzer information
2061 // or some other type of call, assume the worst case: all arguments
2062 // globally escape.
2063 {
2064 // adjust escape state for outgoing arguments
2065 const TypeTuple * d = call->tf()->domain();
2066 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2067 const Type* at = d->field_at(i);
2068 if (at->isa_oopptr() != NULL) {
2069 Node *arg = call->in(i)->uncast();
2070 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2071 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2072 uint pt = j.elem;
2073 set_escape_state(pt, PointsToNode::GlobalEscape);
2074 }
2075 }
2076 }
2077 }
2078 }
2079 }
2080 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
2081 CallNode *call = resproj->in(0)->as_Call();
2082 uint call_idx = call->_idx;
2083 uint resproj_idx = resproj->_idx;
2085 switch (call->Opcode()) {
2086 case Op_Allocate:
2087 {
2088 Node *k = call->in(AllocateNode::KlassNode);
2089 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2090 assert(kt != NULL, "TypeKlassPtr required.");
2091 ciKlass* cik = kt->klass();
2093 PointsToNode::EscapeState es;
2094 uint edge_to;
2095 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
2096 !cik->is_instance_klass() || // StressReflectiveCode
2097 cik->as_instance_klass()->has_finalizer()) {
2098 es = PointsToNode::GlobalEscape;
2099 edge_to = _phantom_object; // Could not be worse
2100 } else {
2101 es = PointsToNode::NoEscape;
2102 edge_to = call_idx;
2103 }
2104 set_escape_state(call_idx, es);
2105 add_pointsto_edge(resproj_idx, edge_to);
2106 _processed.set(resproj_idx);
2107 break;
2108 }
2110 case Op_AllocateArray:
2111 {
2113 Node *k = call->in(AllocateNode::KlassNode);
2114 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2115 assert(kt != NULL, "TypeKlassPtr required.");
2116 ciKlass* cik = kt->klass();
2118 PointsToNode::EscapeState es;
2119 uint edge_to;
2120 if (!cik->is_array_klass()) { // StressReflectiveCode
2121 es = PointsToNode::GlobalEscape;
2122 edge_to = _phantom_object;
2123 } else {
2124 es = PointsToNode::NoEscape;
2125 edge_to = call_idx;
2126 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2127 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
2128 // Not scalar replaceable if the length is not constant or too big.
2129 ptnode_adr(call_idx)->_scalar_replaceable = false;
2130 }
2131 }
2132 set_escape_state(call_idx, es);
2133 add_pointsto_edge(resproj_idx, edge_to);
2134 _processed.set(resproj_idx);
2135 break;
2136 }
2138 case Op_CallStaticJava:
2139 // For a static call, we know exactly what method is being called.
2140 // Use bytecode estimator to record whether the call's return value escapes
2141 {
2142 bool done = true;
2143 const TypeTuple *r = call->tf()->range();
2144 const Type* ret_type = NULL;
2146 if (r->cnt() > TypeFunc::Parms)
2147 ret_type = r->field_at(TypeFunc::Parms);
2149 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2150 // _multianewarray functions return a TypeRawPtr.
2151 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
2152 _processed.set(resproj_idx);
2153 break; // doesn't return a pointer type
2154 }
2155 ciMethod *meth = call->as_CallJava()->method();
2156 const TypeTuple * d = call->tf()->domain();
2157 if (meth == NULL) {
2158 // not a Java method, assume global escape
2159 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2160 add_pointsto_edge(resproj_idx, _phantom_object);
2161 } else {
2162 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
2163 bool copy_dependencies = false;
2165 if (call_analyzer->is_return_allocated()) {
2166 // Returns a newly allocated unescaped object, simply
2167 // update dependency information.
2168 // Mark it as NoEscape so that objects referenced by
2169 // it's fields will be marked as NoEscape at least.
2170 set_escape_state(call_idx, PointsToNode::NoEscape);
2171 add_pointsto_edge(resproj_idx, call_idx);
2172 copy_dependencies = true;
2173 } else if (call_analyzer->is_return_local()) {
2174 // determine whether any arguments are returned
2175 set_escape_state(call_idx, PointsToNode::NoEscape);
2176 bool ret_arg = false;
2177 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2178 const Type* at = d->field_at(i);
2180 if (at->isa_oopptr() != NULL) {
2181 Node *arg = call->in(i)->uncast();
2183 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2184 ret_arg = true;
2185 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2186 if (arg_esp->node_type() == PointsToNode::UnknownType)
2187 done = false;
2188 else if (arg_esp->node_type() == PointsToNode::JavaObject)
2189 add_pointsto_edge(resproj_idx, arg->_idx);
2190 else
2191 add_deferred_edge(resproj_idx, arg->_idx);
2192 arg_esp->_hidden_alias = true;
2193 }
2194 }
2195 }
2196 if (done && !ret_arg) {
2197 // Returns unknown object.
2198 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2199 add_pointsto_edge(resproj_idx, _phantom_object);
2200 }
2201 copy_dependencies = true;
2202 } else {
2203 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2204 add_pointsto_edge(resproj_idx, _phantom_object);
2205 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2206 const Type* at = d->field_at(i);
2207 if (at->isa_oopptr() != NULL) {
2208 Node *arg = call->in(i)->uncast();
2209 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2210 arg_esp->_hidden_alias = true;
2211 }
2212 }
2213 }
2214 if (copy_dependencies)
2215 call_analyzer->copy_dependencies(_compile->dependencies());
2216 }
2217 if (done)
2218 _processed.set(resproj_idx);
2219 break;
2220 }
2222 default:
2223 // Some other type of call, assume the worst case that the
2224 // returned value, if any, globally escapes.
2225 {
2226 const TypeTuple *r = call->tf()->range();
2227 if (r->cnt() > TypeFunc::Parms) {
2228 const Type* ret_type = r->field_at(TypeFunc::Parms);
2230 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2231 // _multianewarray functions return a TypeRawPtr.
2232 if (ret_type->isa_ptr() != NULL) {
2233 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2234 add_pointsto_edge(resproj_idx, _phantom_object);
2235 }
2236 }
2237 _processed.set(resproj_idx);
2238 }
2239 }
2240 }
2242 // Populate Connection Graph with Ideal nodes and create simple
2243 // connection graph edges (do not need to check the node_type of inputs
2244 // or to call PointsTo() to walk the connection graph).
2245 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
2246 if (_processed.test(n->_idx))
2247 return; // No need to redefine node's state.
2249 if (n->is_Call()) {
2250 // Arguments to allocation and locking don't escape.
2251 if (n->is_Allocate()) {
2252 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
2253 record_for_optimizer(n);
2254 } else if (n->is_Lock() || n->is_Unlock()) {
2255 // Put Lock and Unlock nodes on IGVN worklist to process them during
2256 // the first IGVN optimization when escape information is still available.
2257 record_for_optimizer(n);
2258 _processed.set(n->_idx);
2259 } else {
2260 // Don't mark as processed since call's arguments have to be processed.
2261 PointsToNode::NodeType nt = PointsToNode::UnknownType;
2262 PointsToNode::EscapeState es = PointsToNode::UnknownEscape;
2264 // Check if a call returns an object.
2265 const TypeTuple *r = n->as_Call()->tf()->range();
2266 if (r->cnt() > TypeFunc::Parms &&
2267 r->field_at(TypeFunc::Parms)->isa_ptr() &&
2268 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
2269 nt = PointsToNode::JavaObject;
2270 if (!n->is_CallStaticJava()) {
2271 // Since the called mathod is statically unknown assume
2272 // the worst case that the returned value globally escapes.
2273 es = PointsToNode::GlobalEscape;
2274 }
2275 }
2276 add_node(n, nt, es, false);
2277 }
2278 return;
2279 }
2281 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
2282 // ThreadLocal has RawPrt type.
2283 switch (n->Opcode()) {
2284 case Op_AddP:
2285 {
2286 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
2287 break;
2288 }
2289 case Op_CastX2P:
2290 { // "Unsafe" memory access.
2291 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2292 break;
2293 }
2294 case Op_CastPP:
2295 case Op_CheckCastPP:
2296 case Op_EncodeP:
2297 case Op_DecodeN:
2298 {
2299 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2300 int ti = n->in(1)->_idx;
2301 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2302 if (nt == PointsToNode::UnknownType) {
2303 _delayed_worklist.push(n); // Process it later.
2304 break;
2305 } else if (nt == PointsToNode::JavaObject) {
2306 add_pointsto_edge(n->_idx, ti);
2307 } else {
2308 add_deferred_edge(n->_idx, ti);
2309 }
2310 _processed.set(n->_idx);
2311 break;
2312 }
2313 case Op_ConP:
2314 {
2315 // assume all pointer constants globally escape except for null
2316 PointsToNode::EscapeState es;
2317 if (phase->type(n) == TypePtr::NULL_PTR)
2318 es = PointsToNode::NoEscape;
2319 else
2320 es = PointsToNode::GlobalEscape;
2322 add_node(n, PointsToNode::JavaObject, es, true);
2323 break;
2324 }
2325 case Op_ConN:
2326 {
2327 // assume all narrow oop constants globally escape except for null
2328 PointsToNode::EscapeState es;
2329 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
2330 es = PointsToNode::NoEscape;
2331 else
2332 es = PointsToNode::GlobalEscape;
2334 add_node(n, PointsToNode::JavaObject, es, true);
2335 break;
2336 }
2337 case Op_CreateEx:
2338 {
2339 // assume that all exception objects globally escape
2340 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2341 break;
2342 }
2343 case Op_LoadKlass:
2344 case Op_LoadNKlass:
2345 {
2346 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2347 break;
2348 }
2349 case Op_LoadP:
2350 case Op_LoadN:
2351 {
2352 const Type *t = phase->type(n);
2353 if (t->make_ptr() == NULL) {
2354 _processed.set(n->_idx);
2355 return;
2356 }
2357 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2358 break;
2359 }
2360 case Op_Parm:
2361 {
2362 _processed.set(n->_idx); // No need to redefine it state.
2363 uint con = n->as_Proj()->_con;
2364 if (con < TypeFunc::Parms)
2365 return;
2366 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2367 if (t->isa_ptr() == NULL)
2368 return;
2369 // We have to assume all input parameters globally escape
2370 // (Note: passing 'false' since _processed is already set).
2371 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2372 break;
2373 }
2374 case Op_Phi:
2375 {
2376 const Type *t = n->as_Phi()->type();
2377 if (t->make_ptr() == NULL) {
2378 // nothing to do if not an oop or narrow oop
2379 _processed.set(n->_idx);
2380 return;
2381 }
2382 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2383 uint i;
2384 for (i = 1; i < n->req() ; i++) {
2385 Node* in = n->in(i);
2386 if (in == NULL)
2387 continue; // ignore NULL
2388 in = in->uncast();
2389 if (in->is_top() || in == n)
2390 continue; // ignore top or inputs which go back this node
2391 int ti = in->_idx;
2392 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2393 if (nt == PointsToNode::UnknownType) {
2394 break;
2395 } else if (nt == PointsToNode::JavaObject) {
2396 add_pointsto_edge(n->_idx, ti);
2397 } else {
2398 add_deferred_edge(n->_idx, ti);
2399 }
2400 }
2401 if (i >= n->req())
2402 _processed.set(n->_idx);
2403 else
2404 _delayed_worklist.push(n);
2405 break;
2406 }
2407 case Op_Proj:
2408 {
2409 // we are only interested in the oop result projection from a call
2410 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2411 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2412 assert(r->cnt() > TypeFunc::Parms, "sanity");
2413 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2414 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2415 int ti = n->in(0)->_idx;
2416 // The call may not be registered yet (since not all its inputs are registered)
2417 // if this is the projection from backbranch edge of Phi.
2418 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) {
2419 process_call_result(n->as_Proj(), phase);
2420 }
2421 if (!_processed.test(n->_idx)) {
2422 // The call's result may need to be processed later if the call
2423 // returns it's argument and the argument is not processed yet.
2424 _delayed_worklist.push(n);
2425 }
2426 break;
2427 }
2428 }
2429 _processed.set(n->_idx);
2430 break;
2431 }
2432 case Op_Return:
2433 {
2434 if( n->req() > TypeFunc::Parms &&
2435 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2436 // Treat Return value as LocalVar with GlobalEscape escape state.
2437 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2438 int ti = n->in(TypeFunc::Parms)->_idx;
2439 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2440 if (nt == PointsToNode::UnknownType) {
2441 _delayed_worklist.push(n); // Process it later.
2442 break;
2443 } else if (nt == PointsToNode::JavaObject) {
2444 add_pointsto_edge(n->_idx, ti);
2445 } else {
2446 add_deferred_edge(n->_idx, ti);
2447 }
2448 }
2449 _processed.set(n->_idx);
2450 break;
2451 }
2452 case Op_StoreP:
2453 case Op_StoreN:
2454 {
2455 const Type *adr_type = phase->type(n->in(MemNode::Address));
2456 adr_type = adr_type->make_ptr();
2457 if (adr_type->isa_oopptr()) {
2458 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2459 } else {
2460 Node* adr = n->in(MemNode::Address);
2461 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2462 adr->in(AddPNode::Address)->is_Proj() &&
2463 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2464 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2465 // We are computing a raw address for a store captured
2466 // by an Initialize compute an appropriate address type.
2467 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2468 assert(offs != Type::OffsetBot, "offset must be a constant");
2469 } else {
2470 _processed.set(n->_idx);
2471 return;
2472 }
2473 }
2474 break;
2475 }
2476 case Op_StorePConditional:
2477 case Op_CompareAndSwapP:
2478 case Op_CompareAndSwapN:
2479 {
2480 const Type *adr_type = phase->type(n->in(MemNode::Address));
2481 adr_type = adr_type->make_ptr();
2482 if (adr_type->isa_oopptr()) {
2483 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2484 } else {
2485 _processed.set(n->_idx);
2486 return;
2487 }
2488 break;
2489 }
2490 case Op_AryEq:
2491 case Op_StrComp:
2492 case Op_StrEquals:
2493 case Op_StrIndexOf:
2494 {
2495 // char[] arrays passed to string intrinsics are not scalar replaceable.
2496 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2497 break;
2498 }
2499 case Op_ThreadLocal:
2500 {
2501 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2502 break;
2503 }
2504 default:
2505 ;
2506 // nothing to do
2507 }
2508 return;
2509 }
2511 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2512 uint n_idx = n->_idx;
2513 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered");
2515 // Don't set processed bit for AddP, LoadP, StoreP since
2516 // they may need more then one pass to process.
2517 // Also don't mark as processed Call nodes since their
2518 // arguments may need more then one pass to process.
2519 if (_processed.test(n_idx))
2520 return; // No need to redefine node's state.
2522 if (n->is_Call()) {
2523 CallNode *call = n->as_Call();
2524 process_call_arguments(call, phase);
2525 return;
2526 }
2528 switch (n->Opcode()) {
2529 case Op_AddP:
2530 {
2531 Node *base = get_addp_base(n);
2532 // Create a field edge to this node from everything base could point to.
2533 for( VectorSetI i(PointsTo(base)); i.test(); ++i ) {
2534 uint pt = i.elem;
2535 add_field_edge(pt, n_idx, address_offset(n, phase));
2536 }
2537 break;
2538 }
2539 case Op_CastX2P:
2540 {
2541 assert(false, "Op_CastX2P");
2542 break;
2543 }
2544 case Op_CastPP:
2545 case Op_CheckCastPP:
2546 case Op_EncodeP:
2547 case Op_DecodeN:
2548 {
2549 int ti = n->in(1)->_idx;
2550 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered");
2551 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2552 add_pointsto_edge(n_idx, ti);
2553 } else {
2554 add_deferred_edge(n_idx, ti);
2555 }
2556 _processed.set(n_idx);
2557 break;
2558 }
2559 case Op_ConP:
2560 {
2561 assert(false, "Op_ConP");
2562 break;
2563 }
2564 case Op_ConN:
2565 {
2566 assert(false, "Op_ConN");
2567 break;
2568 }
2569 case Op_CreateEx:
2570 {
2571 assert(false, "Op_CreateEx");
2572 break;
2573 }
2574 case Op_LoadKlass:
2575 case Op_LoadNKlass:
2576 {
2577 assert(false, "Op_LoadKlass");
2578 break;
2579 }
2580 case Op_LoadP:
2581 case Op_LoadN:
2582 {
2583 const Type *t = phase->type(n);
2584 #ifdef ASSERT
2585 if (t->make_ptr() == NULL)
2586 assert(false, "Op_LoadP");
2587 #endif
2589 Node* adr = n->in(MemNode::Address)->uncast();
2590 Node* adr_base;
2591 if (adr->is_AddP()) {
2592 adr_base = get_addp_base(adr);
2593 } else {
2594 adr_base = adr;
2595 }
2597 // For everything "adr_base" could point to, create a deferred edge from
2598 // this node to each field with the same offset.
2599 int offset = address_offset(adr, phase);
2600 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2601 uint pt = i.elem;
2602 add_deferred_edge_to_fields(n_idx, pt, offset);
2603 }
2604 break;
2605 }
2606 case Op_Parm:
2607 {
2608 assert(false, "Op_Parm");
2609 break;
2610 }
2611 case Op_Phi:
2612 {
2613 #ifdef ASSERT
2614 const Type *t = n->as_Phi()->type();
2615 if (t->make_ptr() == NULL)
2616 assert(false, "Op_Phi");
2617 #endif
2618 for (uint i = 1; i < n->req() ; i++) {
2619 Node* in = n->in(i);
2620 if (in == NULL)
2621 continue; // ignore NULL
2622 in = in->uncast();
2623 if (in->is_top() || in == n)
2624 continue; // ignore top or inputs which go back this node
2625 int ti = in->_idx;
2626 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2627 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
2628 if (nt == PointsToNode::JavaObject) {
2629 add_pointsto_edge(n_idx, ti);
2630 } else {
2631 add_deferred_edge(n_idx, ti);
2632 }
2633 }
2634 _processed.set(n_idx);
2635 break;
2636 }
2637 case Op_Proj:
2638 {
2639 // we are only interested in the oop result projection from a call
2640 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2641 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType,
2642 "all nodes should be registered");
2643 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2644 assert(r->cnt() > TypeFunc::Parms, "sanity");
2645 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2646 process_call_result(n->as_Proj(), phase);
2647 assert(_processed.test(n_idx), "all call results should be processed");
2648 break;
2649 }
2650 }
2651 assert(false, "Op_Proj");
2652 break;
2653 }
2654 case Op_Return:
2655 {
2656 #ifdef ASSERT
2657 if( n->req() <= TypeFunc::Parms ||
2658 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2659 assert(false, "Op_Return");
2660 }
2661 #endif
2662 int ti = n->in(TypeFunc::Parms)->_idx;
2663 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered");
2664 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2665 add_pointsto_edge(n_idx, ti);
2666 } else {
2667 add_deferred_edge(n_idx, ti);
2668 }
2669 _processed.set(n_idx);
2670 break;
2671 }
2672 case Op_StoreP:
2673 case Op_StoreN:
2674 case Op_StorePConditional:
2675 case Op_CompareAndSwapP:
2676 case Op_CompareAndSwapN:
2677 {
2678 Node *adr = n->in(MemNode::Address);
2679 const Type *adr_type = phase->type(adr)->make_ptr();
2680 #ifdef ASSERT
2681 if (!adr_type->isa_oopptr())
2682 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
2683 #endif
2685 assert(adr->is_AddP(), "expecting an AddP");
2686 Node *adr_base = get_addp_base(adr);
2687 Node *val = n->in(MemNode::ValueIn)->uncast();
2688 // For everything "adr_base" could point to, create a deferred edge
2689 // to "val" from each field with the same offset.
2690 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2691 uint pt = i.elem;
2692 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase));
2693 }
2694 break;
2695 }
2696 case Op_AryEq:
2697 case Op_StrComp:
2698 case Op_StrEquals:
2699 case Op_StrIndexOf:
2700 {
2701 // char[] arrays passed to string intrinsic do not escape but
2702 // they are not scalar replaceable. Adjust escape state for them.
2703 // Start from in(2) edge since in(1) is memory edge.
2704 for (uint i = 2; i < n->req(); i++) {
2705 Node* adr = n->in(i)->uncast();
2706 const Type *at = phase->type(adr);
2707 if (!adr->is_top() && at->isa_ptr()) {
2708 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
2709 at->isa_ptr() != NULL, "expecting an Ptr");
2710 if (adr->is_AddP()) {
2711 adr = get_addp_base(adr);
2712 }
2713 // Mark as ArgEscape everything "adr" could point to.
2714 set_escape_state(adr->_idx, PointsToNode::ArgEscape);
2715 }
2716 }
2717 _processed.set(n_idx);
2718 break;
2719 }
2720 case Op_ThreadLocal:
2721 {
2722 assert(false, "Op_ThreadLocal");
2723 break;
2724 }
2725 default:
2726 // This method should be called only for EA specific nodes.
2727 ShouldNotReachHere();
2728 }
2729 }
2731 #ifndef PRODUCT
2732 void ConnectionGraph::dump() {
2733 bool first = true;
2735 uint size = nodes_size();
2736 for (uint ni = 0; ni < size; ni++) {
2737 PointsToNode *ptn = ptnode_adr(ni);
2738 PointsToNode::NodeType ptn_type = ptn->node_type();
2740 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
2741 continue;
2742 PointsToNode::EscapeState es = escape_state(ptn->_node);
2743 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
2744 if (first) {
2745 tty->cr();
2746 tty->print("======== Connection graph for ");
2747 _compile->method()->print_short_name();
2748 tty->cr();
2749 first = false;
2750 }
2751 tty->print("%6d ", ni);
2752 ptn->dump();
2753 // Print all locals which reference this allocation
2754 for (uint li = ni; li < size; li++) {
2755 PointsToNode *ptn_loc = ptnode_adr(li);
2756 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
2757 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
2758 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
2759 ptnode_adr(li)->dump(false);
2760 }
2761 }
2762 if (Verbose) {
2763 // Print all fields which reference this allocation
2764 for (uint i = 0; i < ptn->edge_count(); i++) {
2765 uint ei = ptn->edge_target(i);
2766 ptnode_adr(ei)->dump(false);
2767 }
2768 }
2769 tty->cr();
2770 }
2771 }
2772 }
2773 #endif