Fri, 07 Nov 2008 09:29:38 -0800
6462850: generate biased locking code in C2 ideal graph
Summary: Inline biased locking code in C2 ideal graph during macro nodes expansion
Reviewed-by: never
1 /*
2 * Copyright 2005-2008 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_escape.cpp.incl"
28 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) {
29 uint v = (targIdx << EdgeShift) + ((uint) et);
30 if (_edges == NULL) {
31 Arena *a = Compile::current()->comp_arena();
32 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0);
33 }
34 _edges->append_if_missing(v);
35 }
37 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) {
38 uint v = (targIdx << EdgeShift) + ((uint) et);
40 _edges->remove(v);
41 }
43 #ifndef PRODUCT
44 static const char *node_type_names[] = {
45 "UnknownType",
46 "JavaObject",
47 "LocalVar",
48 "Field"
49 };
51 static const char *esc_names[] = {
52 "UnknownEscape",
53 "NoEscape",
54 "ArgEscape",
55 "GlobalEscape"
56 };
58 static const char *edge_type_suffix[] = {
59 "?", // UnknownEdge
60 "P", // PointsToEdge
61 "D", // DeferredEdge
62 "F" // FieldEdge
63 };
65 void PointsToNode::dump(bool print_state) const {
66 NodeType nt = node_type();
67 tty->print("%s ", node_type_names[(int) nt]);
68 if (print_state) {
69 EscapeState es = escape_state();
70 tty->print("%s %s ", esc_names[(int) es], _scalar_replaceable ? "":"NSR");
71 }
72 tty->print("[[");
73 for (uint i = 0; i < edge_count(); i++) {
74 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]);
75 }
76 tty->print("]] ");
77 if (_node == NULL)
78 tty->print_cr("<null>");
79 else
80 _node->dump();
81 }
82 #endif
84 ConnectionGraph::ConnectionGraph(Compile * C) :
85 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()),
86 _processed(C->comp_arena()),
87 _collecting(true),
88 _compile(C),
89 _node_map(C->comp_arena()) {
91 _phantom_object = C->top()->_idx,
92 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true);
94 // Add ConP(#NULL) and ConN(#NULL) nodes.
95 PhaseGVN* igvn = C->initial_gvn();
96 Node* oop_null = igvn->zerocon(T_OBJECT);
97 _oop_null = oop_null->_idx;
98 assert(_oop_null < C->unique(), "should be created already");
99 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
101 if (UseCompressedOops) {
102 Node* noop_null = igvn->zerocon(T_NARROWOOP);
103 _noop_null = noop_null->_idx;
104 assert(_noop_null < C->unique(), "should be created already");
105 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
106 }
107 }
109 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) {
110 PointsToNode *f = ptnode_adr(from_i);
111 PointsToNode *t = ptnode_adr(to_i);
113 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
114 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge");
115 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge");
116 f->add_edge(to_i, PointsToNode::PointsToEdge);
117 }
119 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) {
120 PointsToNode *f = ptnode_adr(from_i);
121 PointsToNode *t = ptnode_adr(to_i);
123 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
124 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
125 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
126 // don't add a self-referential edge, this can occur during removal of
127 // deferred edges
128 if (from_i != to_i)
129 f->add_edge(to_i, PointsToNode::DeferredEdge);
130 }
132 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
133 const Type *adr_type = phase->type(adr);
134 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
135 adr->in(AddPNode::Address)->is_Proj() &&
136 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
137 // We are computing a raw address for a store captured by an Initialize
138 // compute an appropriate address type. AddP cases #3 and #5 (see below).
139 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
140 assert(offs != Type::OffsetBot ||
141 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
142 "offset must be a constant or it is initialization of array");
143 return offs;
144 }
145 const TypePtr *t_ptr = adr_type->isa_ptr();
146 assert(t_ptr != NULL, "must be a pointer type");
147 return t_ptr->offset();
148 }
150 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) {
151 PointsToNode *f = ptnode_adr(from_i);
152 PointsToNode *t = ptnode_adr(to_i);
154 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
155 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge");
156 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge");
157 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets");
158 t->set_offset(offset);
160 f->add_edge(to_i, PointsToNode::FieldEdge);
161 }
163 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) {
164 PointsToNode *npt = ptnode_adr(ni);
165 PointsToNode::EscapeState old_es = npt->escape_state();
166 if (es > old_es)
167 npt->set_escape_state(es);
168 }
170 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt,
171 PointsToNode::EscapeState es, bool done) {
172 PointsToNode* ptadr = ptnode_adr(n->_idx);
173 ptadr->_node = n;
174 ptadr->set_node_type(nt);
176 // inline set_escape_state(idx, es);
177 PointsToNode::EscapeState old_es = ptadr->escape_state();
178 if (es > old_es)
179 ptadr->set_escape_state(es);
181 if (done)
182 _processed.set(n->_idx);
183 }
185 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n, PhaseTransform *phase) {
186 uint idx = n->_idx;
187 PointsToNode::EscapeState es;
189 // If we are still collecting or there were no non-escaping allocations
190 // we don't know the answer yet
191 if (_collecting)
192 return PointsToNode::UnknownEscape;
194 // if the node was created after the escape computation, return
195 // UnknownEscape
196 if (idx >= nodes_size())
197 return PointsToNode::UnknownEscape;
199 es = ptnode_adr(idx)->escape_state();
201 // if we have already computed a value, return it
202 if (es != PointsToNode::UnknownEscape)
203 return es;
205 // PointsTo() calls n->uncast() which can return a new ideal node.
206 if (n->uncast()->_idx >= nodes_size())
207 return PointsToNode::UnknownEscape;
209 // compute max escape state of anything this node could point to
210 VectorSet ptset(Thread::current()->resource_area());
211 PointsTo(ptset, n, phase);
212 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) {
213 uint pt = i.elem;
214 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
215 if (pes > es)
216 es = pes;
217 }
218 // cache the computed escape state
219 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state");
220 ptnode_adr(idx)->set_escape_state(es);
221 return es;
222 }
224 void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n, PhaseTransform *phase) {
225 VectorSet visited(Thread::current()->resource_area());
226 GrowableArray<uint> worklist;
228 #ifdef ASSERT
229 Node *orig_n = n;
230 #endif
232 n = n->uncast();
233 PointsToNode* npt = ptnode_adr(n->_idx);
235 // If we have a JavaObject, return just that object
236 if (npt->node_type() == PointsToNode::JavaObject) {
237 ptset.set(n->_idx);
238 return;
239 }
240 #ifdef ASSERT
241 if (npt->_node == NULL) {
242 if (orig_n != n)
243 orig_n->dump();
244 n->dump();
245 assert(npt->_node != NULL, "unregistered node");
246 }
247 #endif
248 worklist.push(n->_idx);
249 while(worklist.length() > 0) {
250 int ni = worklist.pop();
251 if (visited.test_set(ni))
252 continue;
254 PointsToNode* pn = ptnode_adr(ni);
255 // ensure that all inputs of a Phi have been processed
256 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
258 int edges_processed = 0;
259 uint e_cnt = pn->edge_count();
260 for (uint e = 0; e < e_cnt; e++) {
261 uint etgt = pn->edge_target(e);
262 PointsToNode::EdgeType et = pn->edge_type(e);
263 if (et == PointsToNode::PointsToEdge) {
264 ptset.set(etgt);
265 edges_processed++;
266 } else if (et == PointsToNode::DeferredEdge) {
267 worklist.push(etgt);
268 edges_processed++;
269 } else {
270 assert(false,"neither PointsToEdge or DeferredEdge");
271 }
272 }
273 if (edges_processed == 0) {
274 // no deferred or pointsto edges found. Assume the value was set
275 // outside this method. Add the phantom object to the pointsto set.
276 ptset.set(_phantom_object);
277 }
278 }
279 }
281 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
282 // This method is most expensive during ConnectionGraph construction.
283 // Reuse vectorSet and an additional growable array for deferred edges.
284 deferred_edges->clear();
285 visited->Clear();
287 visited->set(ni);
288 PointsToNode *ptn = ptnode_adr(ni);
290 // Mark current edges as visited and move deferred edges to separate array.
291 for (uint i = 0; i < ptn->edge_count(); ) {
292 uint t = ptn->edge_target(i);
293 #ifdef ASSERT
294 assert(!visited->test_set(t), "expecting no duplications");
295 #else
296 visited->set(t);
297 #endif
298 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
299 ptn->remove_edge(t, PointsToNode::DeferredEdge);
300 deferred_edges->append(t);
301 } else {
302 i++;
303 }
304 }
305 for (int next = 0; next < deferred_edges->length(); ++next) {
306 uint t = deferred_edges->at(next);
307 PointsToNode *ptt = ptnode_adr(t);
308 uint e_cnt = ptt->edge_count();
309 for (uint e = 0; e < e_cnt; e++) {
310 uint etgt = ptt->edge_target(e);
311 if (visited->test_set(etgt))
312 continue;
314 PointsToNode::EdgeType et = ptt->edge_type(e);
315 if (et == PointsToNode::PointsToEdge) {
316 add_pointsto_edge(ni, etgt);
317 if(etgt == _phantom_object) {
318 // Special case - field set outside (globally escaping).
319 ptn->set_escape_state(PointsToNode::GlobalEscape);
320 }
321 } else if (et == PointsToNode::DeferredEdge) {
322 deferred_edges->append(etgt);
323 } else {
324 assert(false,"invalid connection graph");
325 }
326 }
327 }
328 }
331 // Add an edge to node given by "to_i" from any field of adr_i whose offset
332 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
333 // a pointsto edge is added if it is a JavaObject
335 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
336 PointsToNode* an = ptnode_adr(adr_i);
337 PointsToNode* to = ptnode_adr(to_i);
338 bool deferred = (to->node_type() == PointsToNode::LocalVar);
340 for (uint fe = 0; fe < an->edge_count(); fe++) {
341 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
342 int fi = an->edge_target(fe);
343 PointsToNode* pf = ptnode_adr(fi);
344 int po = pf->offset();
345 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
346 if (deferred)
347 add_deferred_edge(fi, to_i);
348 else
349 add_pointsto_edge(fi, to_i);
350 }
351 }
352 }
354 // Add a deferred edge from node given by "from_i" to any field of adr_i
355 // whose offset matches "offset".
356 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
357 PointsToNode* an = ptnode_adr(adr_i);
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 (pf->edge_count() == 0) {
364 // we have not seen any stores to this field, assume it was set outside this method
365 add_pointsto_edge(fi, _phantom_object);
366 }
367 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
368 add_deferred_edge(from_i, fi);
369 }
370 }
371 }
373 // Helper functions
375 static Node* get_addp_base(Node *addp) {
376 assert(addp->is_AddP(), "must be AddP");
377 //
378 // AddP cases for Base and Address inputs:
379 // case #1. Direct object's field reference:
380 // Allocate
381 // |
382 // Proj #5 ( oop result )
383 // |
384 // CheckCastPP (cast to instance type)
385 // | |
386 // AddP ( base == address )
387 //
388 // case #2. Indirect object's field reference:
389 // Phi
390 // |
391 // CastPP (cast to instance type)
392 // | |
393 // AddP ( base == address )
394 //
395 // case #3. Raw object's field reference for Initialize node:
396 // Allocate
397 // |
398 // Proj #5 ( oop result )
399 // top |
400 // \ |
401 // AddP ( base == top )
402 //
403 // case #4. Array's element reference:
404 // {CheckCastPP | CastPP}
405 // | | |
406 // | AddP ( array's element offset )
407 // | |
408 // AddP ( array's offset )
409 //
410 // case #5. Raw object's field reference for arraycopy stub call:
411 // The inline_native_clone() case when the arraycopy stub is called
412 // after the allocation before Initialize and CheckCastPP nodes.
413 // Allocate
414 // |
415 // Proj #5 ( oop result )
416 // | |
417 // AddP ( base == address )
418 //
419 // case #6. Constant Pool, ThreadLocal, CastX2P or
420 // Raw object's field reference:
421 // {ConP, ThreadLocal, CastX2P, raw Load}
422 // top |
423 // \ |
424 // AddP ( base == top )
425 //
426 // case #7. Klass's field reference.
427 // LoadKlass
428 // | |
429 // AddP ( base == address )
430 //
431 // case #8. narrow Klass's field reference.
432 // LoadNKlass
433 // |
434 // DecodeN
435 // | |
436 // AddP ( base == address )
437 //
438 Node *base = addp->in(AddPNode::Base)->uncast();
439 if (base->is_top()) { // The AddP case #3 and #6.
440 base = addp->in(AddPNode::Address)->uncast();
441 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
442 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
443 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
444 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
445 }
446 return base;
447 }
449 static Node* find_second_addp(Node* addp, Node* n) {
450 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
452 Node* addp2 = addp->raw_out(0);
453 if (addp->outcnt() == 1 && addp2->is_AddP() &&
454 addp2->in(AddPNode::Base) == n &&
455 addp2->in(AddPNode::Address) == addp) {
457 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
458 //
459 // Find array's offset to push it on worklist first and
460 // as result process an array's element offset first (pushed second)
461 // to avoid CastPP for the array's offset.
462 // Otherwise the inserted CastPP (LocalVar) will point to what
463 // the AddP (Field) points to. Which would be wrong since
464 // the algorithm expects the CastPP has the same point as
465 // as AddP's base CheckCastPP (LocalVar).
466 //
467 // ArrayAllocation
468 // |
469 // CheckCastPP
470 // |
471 // memProj (from ArrayAllocation CheckCastPP)
472 // | ||
473 // | || Int (element index)
474 // | || | ConI (log(element size))
475 // | || | /
476 // | || LShift
477 // | || /
478 // | AddP (array's element offset)
479 // | |
480 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
481 // | / /
482 // AddP (array's offset)
483 // |
484 // Load/Store (memory operation on array's element)
485 //
486 return addp2;
487 }
488 return NULL;
489 }
491 //
492 // Adjust the type and inputs of an AddP which computes the
493 // address of a field of an instance
494 //
495 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
496 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
497 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
498 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
499 if (t == NULL) {
500 // We are computing a raw address for a store captured by an Initialize
501 // compute an appropriate address type (cases #3 and #5).
502 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
503 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
504 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
505 assert(offs != Type::OffsetBot, "offset must be a constant");
506 t = base_t->add_offset(offs)->is_oopptr();
507 }
508 int inst_id = base_t->instance_id();
509 assert(!t->is_known_instance() || t->instance_id() == inst_id,
510 "old type must be non-instance or match new type");
512 // The type 't' could be subclass of 'base_t'.
513 // As result t->offset() could be large then base_t's size and it will
514 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
515 // constructor verifies correctness of the offset.
516 //
517 // It could happend on subclass's branch (from the type profiling
518 // inlining) which was not eliminated during parsing since the exactness
519 // of the allocation type was not propagated to the subclass type check.
520 //
521 // Do nothing for such AddP node and don't process its users since
522 // this code branch will go away.
523 //
524 if (!t->is_known_instance() &&
525 !t->klass()->equals(base_t->klass()) &&
526 t->klass()->is_subtype_of(base_t->klass())) {
527 return false; // bail out
528 }
530 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
531 // Do NOT remove the next call: ensure an new alias index is allocated
532 // for the instance type
533 int alias_idx = _compile->get_alias_index(tinst);
534 igvn->set_type(addp, tinst);
535 // record the allocation in the node map
536 set_map(addp->_idx, get_map(base->_idx));
538 // Set addp's Base and Address to 'base'.
539 Node *abase = addp->in(AddPNode::Base);
540 Node *adr = addp->in(AddPNode::Address);
541 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
542 adr->in(0)->_idx == (uint)inst_id) {
543 // Skip AddP cases #3 and #5.
544 } else {
545 assert(!abase->is_top(), "sanity"); // AddP case #3
546 if (abase != base) {
547 igvn->hash_delete(addp);
548 addp->set_req(AddPNode::Base, base);
549 if (abase == adr) {
550 addp->set_req(AddPNode::Address, base);
551 } else {
552 // AddP case #4 (adr is array's element offset AddP node)
553 #ifdef ASSERT
554 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
555 assert(adr->is_AddP() && atype != NULL &&
556 atype->instance_id() == inst_id, "array's element offset should be processed first");
557 #endif
558 }
559 igvn->hash_insert(addp);
560 }
561 }
562 // Put on IGVN worklist since at least addp's type was changed above.
563 record_for_optimizer(addp);
564 return true;
565 }
567 //
568 // Create a new version of orig_phi if necessary. Returns either the newly
569 // created phi or an existing phi. Sets create_new to indicate wheter a new
570 // phi was created. Cache the last newly created phi in the node map.
571 //
572 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
573 Compile *C = _compile;
574 new_created = false;
575 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
576 // nothing to do if orig_phi is bottom memory or matches alias_idx
577 if (phi_alias_idx == alias_idx) {
578 return orig_phi;
579 }
580 // have we already created a Phi for this alias index?
581 PhiNode *result = get_map_phi(orig_phi->_idx);
582 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
583 return result;
584 }
585 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
586 if (C->do_escape_analysis() == true && !C->failing()) {
587 // Retry compilation without escape analysis.
588 // If this is the first failure, the sentinel string will "stick"
589 // to the Compile object, and the C2Compiler will see it and retry.
590 C->record_failure(C2Compiler::retry_no_escape_analysis());
591 }
592 return NULL;
593 }
594 orig_phi_worklist.append_if_missing(orig_phi);
595 const TypePtr *atype = C->get_adr_type(alias_idx);
596 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
597 set_map_phi(orig_phi->_idx, result);
598 igvn->set_type(result, result->bottom_type());
599 record_for_optimizer(result);
600 new_created = true;
601 return result;
602 }
604 //
605 // Return a new version of Memory Phi "orig_phi" with the inputs having the
606 // specified alias index.
607 //
608 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
610 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
611 Compile *C = _compile;
612 bool new_phi_created;
613 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
614 if (!new_phi_created) {
615 return result;
616 }
618 GrowableArray<PhiNode *> phi_list;
619 GrowableArray<uint> cur_input;
621 PhiNode *phi = orig_phi;
622 uint idx = 1;
623 bool finished = false;
624 while(!finished) {
625 while (idx < phi->req()) {
626 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
627 if (mem != NULL && mem->is_Phi()) {
628 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
629 if (new_phi_created) {
630 // found an phi for which we created a new split, push current one on worklist and begin
631 // processing new one
632 phi_list.push(phi);
633 cur_input.push(idx);
634 phi = mem->as_Phi();
635 result = newphi;
636 idx = 1;
637 continue;
638 } else {
639 mem = newphi;
640 }
641 }
642 if (C->failing()) {
643 return NULL;
644 }
645 result->set_req(idx++, mem);
646 }
647 #ifdef ASSERT
648 // verify that the new Phi has an input for each input of the original
649 assert( phi->req() == result->req(), "must have same number of inputs.");
650 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
651 #endif
652 // Check if all new phi's inputs have specified alias index.
653 // Otherwise use old phi.
654 for (uint i = 1; i < phi->req(); i++) {
655 Node* in = result->in(i);
656 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
657 }
658 // we have finished processing a Phi, see if there are any more to do
659 finished = (phi_list.length() == 0 );
660 if (!finished) {
661 phi = phi_list.pop();
662 idx = cur_input.pop();
663 PhiNode *prev_result = get_map_phi(phi->_idx);
664 prev_result->set_req(idx++, result);
665 result = prev_result;
666 }
667 }
668 return result;
669 }
672 //
673 // The next methods are derived from methods in MemNode.
674 //
675 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *tinst) {
676 Node *mem = mmem;
677 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
678 // means an array I have not precisely typed yet. Do not do any
679 // alias stuff with it any time soon.
680 if( tinst->base() != Type::AnyPtr &&
681 !(tinst->klass()->is_java_lang_Object() &&
682 tinst->offset() == Type::OffsetBot) ) {
683 mem = mmem->memory_at(alias_idx);
684 // Update input if it is progress over what we have now
685 }
686 return mem;
687 }
689 //
690 // Search memory chain of "mem" to find a MemNode whose address
691 // is the specified alias index.
692 //
693 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
694 if (orig_mem == NULL)
695 return orig_mem;
696 Compile* C = phase->C;
697 const TypeOopPtr *tinst = C->get_adr_type(alias_idx)->isa_oopptr();
698 bool is_instance = (tinst != NULL) && tinst->is_known_instance();
699 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
700 Node *prev = NULL;
701 Node *result = orig_mem;
702 while (prev != result) {
703 prev = result;
704 if (result == start_mem)
705 break; // hit one of our sentinals
706 if (result->is_Mem()) {
707 const Type *at = phase->type(result->in(MemNode::Address));
708 if (at != Type::TOP) {
709 assert (at->isa_ptr() != NULL, "pointer type required.");
710 int idx = C->get_alias_index(at->is_ptr());
711 if (idx == alias_idx)
712 break;
713 }
714 result = result->in(MemNode::Memory);
715 }
716 if (!is_instance)
717 continue; // don't search further for non-instance types
718 // skip over a call which does not affect this memory slice
719 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
720 Node *proj_in = result->in(0);
721 if (proj_in->is_Allocate() && proj_in->_idx == (uint)tinst->instance_id()) {
722 break; // hit one of our sentinals
723 } else if (proj_in->is_Call()) {
724 CallNode *call = proj_in->as_Call();
725 if (!call->may_modify(tinst, phase)) {
726 result = call->in(TypeFunc::Memory);
727 }
728 } else if (proj_in->is_Initialize()) {
729 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
730 // Stop if this is the initialization for the object instance which
731 // which contains this memory slice, otherwise skip over it.
732 if (alloc == NULL || alloc->_idx != (uint)tinst->instance_id()) {
733 result = proj_in->in(TypeFunc::Memory);
734 }
735 } else if (proj_in->is_MemBar()) {
736 result = proj_in->in(TypeFunc::Memory);
737 }
738 } else if (result->is_MergeMem()) {
739 MergeMemNode *mmem = result->as_MergeMem();
740 result = step_through_mergemem(mmem, alias_idx, tinst);
741 if (result == mmem->base_memory()) {
742 // Didn't find instance memory, search through general slice recursively.
743 result = mmem->memory_at(C->get_general_index(alias_idx));
744 result = find_inst_mem(result, alias_idx, orig_phis, phase);
745 if (C->failing()) {
746 return NULL;
747 }
748 mmem->set_memory_at(alias_idx, result);
749 }
750 } else if (result->is_Phi() &&
751 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
752 Node *un = result->as_Phi()->unique_input(phase);
753 if (un != NULL) {
754 result = un;
755 } else {
756 break;
757 }
758 }
759 }
760 if (result->is_Phi()) {
761 PhiNode *mphi = result->as_Phi();
762 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
763 const TypePtr *t = mphi->adr_type();
764 if (C->get_alias_index(t) != alias_idx) {
765 // Create a new Phi with the specified alias index type.
766 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
767 } else if (!is_instance) {
768 // Push all non-instance Phis on the orig_phis worklist to update inputs
769 // during Phase 4 if needed.
770 orig_phis.append_if_missing(mphi);
771 }
772 }
773 // the result is either MemNode, PhiNode, InitializeNode.
774 return result;
775 }
778 //
779 // Convert the types of unescaped object to instance types where possible,
780 // propagate the new type information through the graph, and update memory
781 // edges and MergeMem inputs to reflect the new type.
782 //
783 // We start with allocations (and calls which may be allocations) on alloc_worklist.
784 // The processing is done in 4 phases:
785 //
786 // Phase 1: Process possible allocations from alloc_worklist. Create instance
787 // types for the CheckCastPP for allocations where possible.
788 // Propagate the the new types through users as follows:
789 // casts and Phi: push users on alloc_worklist
790 // AddP: cast Base and Address inputs to the instance type
791 // push any AddP users on alloc_worklist and push any memnode
792 // users onto memnode_worklist.
793 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
794 // search the Memory chain for a store with the appropriate type
795 // address type. If a Phi is found, create a new version with
796 // the approriate memory slices from each of the Phi inputs.
797 // For stores, process the users as follows:
798 // MemNode: push on memnode_worklist
799 // MergeMem: push on mergemem_worklist
800 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
801 // moving the first node encountered of each instance type to the
802 // the input corresponding to its alias index.
803 // appropriate memory slice.
804 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
805 //
806 // In the following example, the CheckCastPP nodes are the cast of allocation
807 // results and the allocation of node 29 is unescaped and eligible to be an
808 // instance type.
809 //
810 // We start with:
811 //
812 // 7 Parm #memory
813 // 10 ConI "12"
814 // 19 CheckCastPP "Foo"
815 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
816 // 29 CheckCastPP "Foo"
817 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
818 //
819 // 40 StoreP 25 7 20 ... alias_index=4
820 // 50 StoreP 35 40 30 ... alias_index=4
821 // 60 StoreP 45 50 20 ... alias_index=4
822 // 70 LoadP _ 60 30 ... alias_index=4
823 // 80 Phi 75 50 60 Memory alias_index=4
824 // 90 LoadP _ 80 30 ... alias_index=4
825 // 100 LoadP _ 80 20 ... alias_index=4
826 //
827 //
828 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
829 // and creating a new alias index for node 30. This gives:
830 //
831 // 7 Parm #memory
832 // 10 ConI "12"
833 // 19 CheckCastPP "Foo"
834 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
835 // 29 CheckCastPP "Foo" iid=24
836 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
837 //
838 // 40 StoreP 25 7 20 ... alias_index=4
839 // 50 StoreP 35 40 30 ... alias_index=6
840 // 60 StoreP 45 50 20 ... alias_index=4
841 // 70 LoadP _ 60 30 ... alias_index=6
842 // 80 Phi 75 50 60 Memory alias_index=4
843 // 90 LoadP _ 80 30 ... alias_index=6
844 // 100 LoadP _ 80 20 ... alias_index=4
845 //
846 // In phase 2, new memory inputs are computed for the loads and stores,
847 // And a new version of the phi is created. In phase 4, the inputs to
848 // node 80 are updated and then the memory nodes are updated with the
849 // values computed in phase 2. This results in:
850 //
851 // 7 Parm #memory
852 // 10 ConI "12"
853 // 19 CheckCastPP "Foo"
854 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
855 // 29 CheckCastPP "Foo" iid=24
856 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
857 //
858 // 40 StoreP 25 7 20 ... alias_index=4
859 // 50 StoreP 35 7 30 ... alias_index=6
860 // 60 StoreP 45 40 20 ... alias_index=4
861 // 70 LoadP _ 50 30 ... alias_index=6
862 // 80 Phi 75 40 60 Memory alias_index=4
863 // 120 Phi 75 50 50 Memory alias_index=6
864 // 90 LoadP _ 120 30 ... alias_index=6
865 // 100 LoadP _ 80 20 ... alias_index=4
866 //
867 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
868 GrowableArray<Node *> memnode_worklist;
869 GrowableArray<Node *> mergemem_worklist;
870 GrowableArray<PhiNode *> orig_phis;
871 PhaseGVN *igvn = _compile->initial_gvn();
872 uint new_index_start = (uint) _compile->num_alias_types();
873 VectorSet visited(Thread::current()->resource_area());
874 VectorSet ptset(Thread::current()->resource_area());
877 // Phase 1: Process possible allocations from alloc_worklist.
878 // Create instance types for the CheckCastPP for allocations where possible.
879 //
880 // (Note: don't forget to change the order of the second AddP node on
881 // the alloc_worklist if the order of the worklist processing is changed,
882 // see the comment in find_second_addp().)
883 //
884 while (alloc_worklist.length() != 0) {
885 Node *n = alloc_worklist.pop();
886 uint ni = n->_idx;
887 const TypeOopPtr* tinst = NULL;
888 if (n->is_Call()) {
889 CallNode *alloc = n->as_Call();
890 // copy escape information to call node
891 PointsToNode* ptn = ptnode_adr(alloc->_idx);
892 PointsToNode::EscapeState es = escape_state(alloc, igvn);
893 // We have an allocation or call which returns a Java object,
894 // see if it is unescaped.
895 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable)
896 continue;
897 if (alloc->is_Allocate()) {
898 // Set the scalar_replaceable flag before the next check.
899 alloc->as_Allocate()->_is_scalar_replaceable = true;
900 }
901 // find CheckCastPP of call return value
902 n = alloc->result_cast();
903 if (n == NULL || // No uses accept Initialize or
904 !n->is_CheckCastPP()) // not unique CheckCastPP.
905 continue;
906 // The inline code for Object.clone() casts the allocation result to
907 // java.lang.Object and then to the actual type of the allocated
908 // object. Detect this case and use the second cast.
909 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
910 // the allocation result is cast to java.lang.Object and then
911 // to the actual Array type.
912 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
913 && (alloc->is_AllocateArray() ||
914 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
915 Node *cast2 = NULL;
916 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
917 Node *use = n->fast_out(i);
918 if (use->is_CheckCastPP()) {
919 cast2 = use;
920 break;
921 }
922 }
923 if (cast2 != NULL) {
924 n = cast2;
925 } else {
926 continue;
927 }
928 }
929 set_escape_state(n->_idx, es);
930 // in order for an object to be scalar-replaceable, it must be:
931 // - a direct allocation (not a call returning an object)
932 // - non-escaping
933 // - eligible to be a unique type
934 // - not determined to be ineligible by escape analysis
935 set_map(alloc->_idx, n);
936 set_map(n->_idx, alloc);
937 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
938 if (t == NULL)
939 continue; // not a TypeInstPtr
940 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
941 igvn->hash_delete(n);
942 igvn->set_type(n, tinst);
943 n->raise_bottom_type(tinst);
944 igvn->hash_insert(n);
945 record_for_optimizer(n);
946 if (alloc->is_Allocate() && ptn->_scalar_replaceable &&
947 (t->isa_instptr() || t->isa_aryptr())) {
949 // First, put on the worklist all Field edges from Connection Graph
950 // which is more accurate then putting immediate users from Ideal Graph.
951 for (uint e = 0; e < ptn->edge_count(); e++) {
952 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
953 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
954 "only AddP nodes are Field edges in CG");
955 if (use->outcnt() > 0) { // Don't process dead nodes
956 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
957 if (addp2 != NULL) {
958 assert(alloc->is_AllocateArray(),"array allocation was expected");
959 alloc_worklist.append_if_missing(addp2);
960 }
961 alloc_worklist.append_if_missing(use);
962 }
963 }
965 // An allocation may have an Initialize which has raw stores. Scan
966 // the users of the raw allocation result and push AddP users
967 // on alloc_worklist.
968 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
969 assert (raw_result != NULL, "must have an allocation result");
970 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
971 Node *use = raw_result->fast_out(i);
972 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
973 Node* addp2 = find_second_addp(use, raw_result);
974 if (addp2 != NULL) {
975 assert(alloc->is_AllocateArray(),"array allocation was expected");
976 alloc_worklist.append_if_missing(addp2);
977 }
978 alloc_worklist.append_if_missing(use);
979 } else if (use->is_Initialize()) {
980 memnode_worklist.append_if_missing(use);
981 }
982 }
983 }
984 } else if (n->is_AddP()) {
985 ptset.Clear();
986 PointsTo(ptset, get_addp_base(n), igvn);
987 assert(ptset.Size() == 1, "AddP address is unique");
988 uint elem = ptset.getelem(); // Allocation node's index
989 if (elem == _phantom_object)
990 continue; // Assume the value was set outside this method.
991 Node *base = get_map(elem); // CheckCastPP node
992 if (!split_AddP(n, base, igvn)) continue; // wrong type
993 tinst = igvn->type(base)->isa_oopptr();
994 } else if (n->is_Phi() ||
995 n->is_CheckCastPP() ||
996 n->is_EncodeP() ||
997 n->is_DecodeN() ||
998 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
999 if (visited.test_set(n->_idx)) {
1000 assert(n->is_Phi(), "loops only through Phi's");
1001 continue; // already processed
1002 }
1003 ptset.Clear();
1004 PointsTo(ptset, n, igvn);
1005 if (ptset.Size() == 1) {
1006 uint elem = ptset.getelem(); // Allocation node's index
1007 if (elem == _phantom_object)
1008 continue; // Assume the value was set outside this method.
1009 Node *val = get_map(elem); // CheckCastPP node
1010 TypeNode *tn = n->as_Type();
1011 tinst = igvn->type(val)->isa_oopptr();
1012 assert(tinst != NULL && tinst->is_known_instance() &&
1013 (uint)tinst->instance_id() == elem , "instance type expected.");
1015 const Type *tn_type = igvn->type(tn);
1016 const TypeOopPtr *tn_t;
1017 if (tn_type->isa_narrowoop()) {
1018 tn_t = tn_type->make_ptr()->isa_oopptr();
1019 } else {
1020 tn_t = tn_type->isa_oopptr();
1021 }
1023 if (tn_t != NULL &&
1024 tinst->cast_to_instance_id(TypeOopPtr::InstanceBot)->higher_equal(tn_t)) {
1025 if (tn_type->isa_narrowoop()) {
1026 tn_type = tinst->make_narrowoop();
1027 } else {
1028 tn_type = tinst;
1029 }
1030 igvn->hash_delete(tn);
1031 igvn->set_type(tn, tn_type);
1032 tn->set_type(tn_type);
1033 igvn->hash_insert(tn);
1034 record_for_optimizer(n);
1035 } else {
1036 continue; // wrong type
1037 }
1038 }
1039 } else {
1040 continue;
1041 }
1042 // push users on appropriate worklist
1043 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1044 Node *use = n->fast_out(i);
1045 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1046 memnode_worklist.append_if_missing(use);
1047 } else if (use->is_Initialize()) {
1048 memnode_worklist.append_if_missing(use);
1049 } else if (use->is_MergeMem()) {
1050 mergemem_worklist.append_if_missing(use);
1051 } else if (use->is_SafePoint() && tinst != NULL) {
1052 // Look for MergeMem nodes for calls which reference unique allocation
1053 // (through CheckCastPP nodes) even for debug info.
1054 Node* m = use->in(TypeFunc::Memory);
1055 uint iid = tinst->instance_id();
1056 while (m->is_Proj() && m->in(0)->is_SafePoint() &&
1057 m->in(0) != use && !m->in(0)->_idx != iid) {
1058 m = m->in(0)->in(TypeFunc::Memory);
1059 }
1060 if (m->is_MergeMem()) {
1061 mergemem_worklist.append_if_missing(m);
1062 }
1063 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1064 Node* addp2 = find_second_addp(use, n);
1065 if (addp2 != NULL) {
1066 alloc_worklist.append_if_missing(addp2);
1067 }
1068 alloc_worklist.append_if_missing(use);
1069 } else if (use->is_Phi() ||
1070 use->is_CheckCastPP() ||
1071 use->is_EncodeP() ||
1072 use->is_DecodeN() ||
1073 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1074 alloc_worklist.append_if_missing(use);
1075 }
1076 }
1078 }
1079 // New alias types were created in split_AddP().
1080 uint new_index_end = (uint) _compile->num_alias_types();
1082 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1083 // compute new values for Memory inputs (the Memory inputs are not
1084 // actually updated until phase 4.)
1085 if (memnode_worklist.length() == 0)
1086 return; // nothing to do
1088 while (memnode_worklist.length() != 0) {
1089 Node *n = memnode_worklist.pop();
1090 if (visited.test_set(n->_idx))
1091 continue;
1092 if (n->is_Phi()) {
1093 assert(n->as_Phi()->adr_type() != TypePtr::BOTTOM, "narrow memory slice required");
1094 // we don't need to do anything, but the users must be pushed if we haven't processed
1095 // this Phi before
1096 } else if (n->is_Initialize()) {
1097 // we don't need to do anything, but the users of the memory projection must be pushed
1098 n = n->as_Initialize()->proj_out(TypeFunc::Memory);
1099 if (n == NULL)
1100 continue;
1101 } else {
1102 assert(n->is_Mem(), "memory node required.");
1103 Node *addr = n->in(MemNode::Address);
1104 assert(addr->is_AddP(), "AddP required");
1105 const Type *addr_t = igvn->type(addr);
1106 if (addr_t == Type::TOP)
1107 continue;
1108 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1109 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1110 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1111 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1112 if (_compile->failing()) {
1113 return;
1114 }
1115 if (mem != n->in(MemNode::Memory)) {
1116 set_map(n->_idx, mem);
1117 ptnode_adr(n->_idx)->_node = n;
1118 }
1119 if (n->is_Load()) {
1120 continue; // don't push users
1121 } else if (n->is_LoadStore()) {
1122 // get the memory projection
1123 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1124 Node *use = n->fast_out(i);
1125 if (use->Opcode() == Op_SCMemProj) {
1126 n = use;
1127 break;
1128 }
1129 }
1130 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1131 }
1132 }
1133 // push user on appropriate worklist
1134 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1135 Node *use = n->fast_out(i);
1136 if (use->is_Phi()) {
1137 memnode_worklist.append_if_missing(use);
1138 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1139 memnode_worklist.append_if_missing(use);
1140 } else if (use->is_Initialize()) {
1141 memnode_worklist.append_if_missing(use);
1142 } else if (use->is_MergeMem()) {
1143 mergemem_worklist.append_if_missing(use);
1144 }
1145 }
1146 }
1148 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1149 // Walk each memory moving the first node encountered of each
1150 // instance type to the the input corresponding to its alias index.
1151 while (mergemem_worklist.length() != 0) {
1152 Node *n = mergemem_worklist.pop();
1153 assert(n->is_MergeMem(), "MergeMem node required.");
1154 if (visited.test_set(n->_idx))
1155 continue;
1156 MergeMemNode *nmm = n->as_MergeMem();
1157 // Note: we don't want to use MergeMemStream here because we only want to
1158 // scan inputs which exist at the start, not ones we add during processing.
1159 uint nslices = nmm->req();
1160 igvn->hash_delete(nmm);
1161 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1162 Node* mem = nmm->in(i);
1163 Node* cur = NULL;
1164 if (mem == NULL || mem->is_top())
1165 continue;
1166 while (mem->is_Mem()) {
1167 const Type *at = igvn->type(mem->in(MemNode::Address));
1168 if (at != Type::TOP) {
1169 assert (at->isa_ptr() != NULL, "pointer type required.");
1170 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1171 if (idx == i) {
1172 if (cur == NULL)
1173 cur = mem;
1174 } else {
1175 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1176 nmm->set_memory_at(idx, mem);
1177 }
1178 }
1179 }
1180 mem = mem->in(MemNode::Memory);
1181 }
1182 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1183 // Find any instance of the current type if we haven't encountered
1184 // a value of the instance along the chain.
1185 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1186 if((uint)_compile->get_general_index(ni) == i) {
1187 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1188 if (nmm->is_empty_memory(m)) {
1189 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1190 if (_compile->failing()) {
1191 return;
1192 }
1193 nmm->set_memory_at(ni, result);
1194 }
1195 }
1196 }
1197 }
1198 // Find the rest of instances values
1199 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1200 const TypeOopPtr *tinst = igvn->C->get_adr_type(ni)->isa_oopptr();
1201 Node* result = step_through_mergemem(nmm, ni, tinst);
1202 if (result == nmm->base_memory()) {
1203 // Didn't find instance memory, search through general slice recursively.
1204 result = nmm->memory_at(igvn->C->get_general_index(ni));
1205 result = find_inst_mem(result, ni, orig_phis, igvn);
1206 if (_compile->failing()) {
1207 return;
1208 }
1209 nmm->set_memory_at(ni, result);
1210 }
1211 }
1212 igvn->hash_insert(nmm);
1213 record_for_optimizer(nmm);
1215 // Propagate new memory slices to following MergeMem nodes.
1216 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1217 Node *use = n->fast_out(i);
1218 if (use->is_Call()) {
1219 CallNode* in = use->as_Call();
1220 if (in->proj_out(TypeFunc::Memory) != NULL) {
1221 Node* m = in->proj_out(TypeFunc::Memory);
1222 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
1223 Node* mm = m->fast_out(j);
1224 if (mm->is_MergeMem()) {
1225 mergemem_worklist.append_if_missing(mm);
1226 }
1227 }
1228 }
1229 if (use->is_Allocate()) {
1230 use = use->as_Allocate()->initialization();
1231 if (use == NULL) {
1232 continue;
1233 }
1234 }
1235 }
1236 if (use->is_Initialize()) {
1237 InitializeNode* in = use->as_Initialize();
1238 if (in->proj_out(TypeFunc::Memory) != NULL) {
1239 Node* m = in->proj_out(TypeFunc::Memory);
1240 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
1241 Node* mm = m->fast_out(j);
1242 if (mm->is_MergeMem()) {
1243 mergemem_worklist.append_if_missing(mm);
1244 }
1245 }
1246 }
1247 }
1248 }
1249 }
1251 // Phase 4: Update the inputs of non-instance memory Phis and
1252 // the Memory input of memnodes
1253 // First update the inputs of any non-instance Phi's from
1254 // which we split out an instance Phi. Note we don't have
1255 // to recursively process Phi's encounted on the input memory
1256 // chains as is done in split_memory_phi() since they will
1257 // also be processed here.
1258 for (int j = 0; j < orig_phis.length(); j++) {
1259 PhiNode *phi = orig_phis.at(j);
1260 int alias_idx = _compile->get_alias_index(phi->adr_type());
1261 igvn->hash_delete(phi);
1262 for (uint i = 1; i < phi->req(); i++) {
1263 Node *mem = phi->in(i);
1264 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1265 if (_compile->failing()) {
1266 return;
1267 }
1268 if (mem != new_mem) {
1269 phi->set_req(i, new_mem);
1270 }
1271 }
1272 igvn->hash_insert(phi);
1273 record_for_optimizer(phi);
1274 }
1276 // Update the memory inputs of MemNodes with the value we computed
1277 // in Phase 2.
1278 for (uint i = 0; i < nodes_size(); i++) {
1279 Node *nmem = get_map(i);
1280 if (nmem != NULL) {
1281 Node *n = ptnode_adr(i)->_node;
1282 if (n != NULL && n->is_Mem()) {
1283 igvn->hash_delete(n);
1284 n->set_req(MemNode::Memory, nmem);
1285 igvn->hash_insert(n);
1286 record_for_optimizer(n);
1287 }
1288 }
1289 }
1290 }
1292 bool ConnectionGraph::has_candidates(Compile *C) {
1293 // EA brings benefits only when the code has allocations and/or locks which
1294 // are represented by ideal Macro nodes.
1295 int cnt = C->macro_count();
1296 for( int i=0; i < cnt; i++ ) {
1297 Node *n = C->macro_node(i);
1298 if ( n->is_Allocate() )
1299 return true;
1300 if( n->is_Lock() ) {
1301 Node* obj = n->as_Lock()->obj_node()->uncast();
1302 if( !(obj->is_Parm() || obj->is_Con()) )
1303 return true;
1304 }
1305 }
1306 return false;
1307 }
1309 bool ConnectionGraph::compute_escape() {
1310 Compile* C = _compile;
1312 // 1. Populate Connection Graph (CG) with Ideal nodes.
1314 Unique_Node_List worklist_init;
1315 worklist_init.map(C->unique(), NULL); // preallocate space
1317 // Initialize worklist
1318 if (C->root() != NULL) {
1319 worklist_init.push(C->root());
1320 }
1322 GrowableArray<int> cg_worklist;
1323 PhaseGVN* igvn = C->initial_gvn();
1324 bool has_allocations = false;
1326 // Push all useful nodes onto CG list and set their type.
1327 for( uint next = 0; next < worklist_init.size(); ++next ) {
1328 Node* n = worklist_init.at(next);
1329 record_for_escape_analysis(n, igvn);
1330 // Only allocations and java static calls results are checked
1331 // for an escape status. See process_call_result() below.
1332 if (n->is_Allocate() || n->is_CallStaticJava() &&
1333 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1334 has_allocations = true;
1335 }
1336 if(n->is_AddP())
1337 cg_worklist.append(n->_idx);
1338 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1339 Node* m = n->fast_out(i); // Get user
1340 worklist_init.push(m);
1341 }
1342 }
1344 if (!has_allocations) {
1345 _collecting = false;
1346 return false; // Nothing to do.
1347 }
1349 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1350 uint delayed_size = _delayed_worklist.size();
1351 for( uint next = 0; next < delayed_size; ++next ) {
1352 Node* n = _delayed_worklist.at(next);
1353 build_connection_graph(n, igvn);
1354 }
1356 // 3. Pass to create fields edges (Allocate -F-> AddP).
1357 uint cg_length = cg_worklist.length();
1358 for( uint next = 0; next < cg_length; ++next ) {
1359 int ni = cg_worklist.at(next);
1360 build_connection_graph(ptnode_adr(ni)->_node, igvn);
1361 }
1363 cg_worklist.clear();
1364 cg_worklist.append(_phantom_object);
1366 // 4. Build Connection Graph which need
1367 // to walk the connection graph.
1368 for (uint ni = 0; ni < nodes_size(); ni++) {
1369 PointsToNode* ptn = ptnode_adr(ni);
1370 Node *n = ptn->_node;
1371 if (n != NULL) { // Call, AddP, LoadP, StoreP
1372 build_connection_graph(n, igvn);
1373 if (ptn->node_type() != PointsToNode::UnknownType)
1374 cg_worklist.append(n->_idx); // Collect CG nodes
1375 }
1376 }
1378 VectorSet ptset(Thread::current()->resource_area());
1379 GrowableArray<uint> deferred_edges;
1380 VectorSet visited(Thread::current()->resource_area());
1382 // 5. Remove deferred edges from the graph and collect
1383 // information needed for type splitting.
1384 cg_length = cg_worklist.length();
1385 for( uint next = 0; next < cg_length; ++next ) {
1386 int ni = cg_worklist.at(next);
1387 PointsToNode* ptn = ptnode_adr(ni);
1388 PointsToNode::NodeType nt = ptn->node_type();
1389 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1390 remove_deferred(ni, &deferred_edges, &visited);
1391 Node *n = ptn->_node;
1392 if (n->is_AddP()) {
1393 // Search for objects which are not scalar replaceable.
1394 // Mark their escape state as ArgEscape to propagate the state
1395 // to referenced objects.
1396 // Note: currently there are no difference in compiler optimizations
1397 // for ArgEscape objects and NoEscape objects which are not
1398 // scalar replaceable.
1400 int offset = ptn->offset();
1401 Node *base = get_addp_base(n);
1402 ptset.Clear();
1403 PointsTo(ptset, base, igvn);
1404 int ptset_size = ptset.Size();
1406 // Check if a field's initializing value is recorded and add
1407 // a corresponding NULL field's value if it is not recorded.
1408 // Connection Graph does not record a default initialization by NULL
1409 // captured by Initialize node.
1410 //
1411 // Note: it will disable scalar replacement in some cases:
1412 //
1413 // Point p[] = new Point[1];
1414 // p[0] = new Point(); // Will be not scalar replaced
1415 //
1416 // but it will save us from incorrect optimizations in next cases:
1417 //
1418 // Point p[] = new Point[1];
1419 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1420 //
1421 // Without a control flow analysis we can't distinguish above cases.
1422 //
1423 if (offset != Type::OffsetBot && ptset_size == 1) {
1424 uint elem = ptset.getelem(); // Allocation node's index
1425 // It does not matter if it is not Allocation node since
1426 // only non-escaping allocations are scalar replaced.
1427 if (ptnode_adr(elem)->_node->is_Allocate() &&
1428 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) {
1429 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate();
1430 InitializeNode* ini = alloc->initialization();
1431 Node* value = NULL;
1432 if (ini != NULL) {
1433 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
1434 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), igvn);
1435 if (store != NULL && store->is_Store())
1436 value = store->in(MemNode::ValueIn);
1437 }
1438 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
1439 // A field's initializing value was not recorded. Add NULL.
1440 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1441 add_pointsto_edge(ni, null_idx);
1442 }
1443 }
1444 }
1446 // An object is not scalar replaceable if the field which may point
1447 // to it has unknown offset (unknown element of an array of objects).
1448 //
1449 if (offset == Type::OffsetBot) {
1450 uint e_cnt = ptn->edge_count();
1451 for (uint ei = 0; ei < e_cnt; ei++) {
1452 uint npi = ptn->edge_target(ei);
1453 set_escape_state(npi, PointsToNode::ArgEscape);
1454 ptnode_adr(npi)->_scalar_replaceable = false;
1455 }
1456 }
1458 // Currently an object is not scalar replaceable if a LoadStore node
1459 // access its field since the field value is unknown after it.
1460 //
1461 bool has_LoadStore = false;
1462 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1463 Node *use = n->fast_out(i);
1464 if (use->is_LoadStore()) {
1465 has_LoadStore = true;
1466 break;
1467 }
1468 }
1469 // An object is not scalar replaceable if the address points
1470 // to unknown field (unknown element for arrays, offset is OffsetBot).
1471 //
1472 // Or the address may point to more then one object. This may produce
1473 // the false positive result (set scalar_replaceable to false)
1474 // since the flow-insensitive escape analysis can't separate
1475 // the case when stores overwrite the field's value from the case
1476 // when stores happened on different control branches.
1477 //
1478 if (ptset_size > 1 || ptset_size != 0 &&
1479 (has_LoadStore || offset == Type::OffsetBot)) {
1480 for( VectorSetI j(&ptset); j.test(); ++j ) {
1481 set_escape_state(j.elem, PointsToNode::ArgEscape);
1482 ptnode_adr(j.elem)->_scalar_replaceable = false;
1483 }
1484 }
1485 }
1486 }
1487 }
1489 // 6. Propagate escape states.
1490 GrowableArray<int> worklist;
1491 bool has_non_escaping_obj = false;
1493 // push all GlobalEscape nodes on the worklist
1494 for( uint next = 0; next < cg_length; ++next ) {
1495 int nk = cg_worklist.at(next);
1496 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape)
1497 worklist.push(nk);
1498 }
1499 // mark all nodes reachable from GlobalEscape nodes
1500 while(worklist.length() > 0) {
1501 PointsToNode* ptn = ptnode_adr(worklist.pop());
1502 uint e_cnt = ptn->edge_count();
1503 for (uint ei = 0; ei < e_cnt; ei++) {
1504 uint npi = ptn->edge_target(ei);
1505 PointsToNode *np = ptnode_adr(npi);
1506 if (np->escape_state() < PointsToNode::GlobalEscape) {
1507 np->set_escape_state(PointsToNode::GlobalEscape);
1508 worklist.push(npi);
1509 }
1510 }
1511 }
1513 // push all ArgEscape nodes on the worklist
1514 for( uint next = 0; next < cg_length; ++next ) {
1515 int nk = cg_worklist.at(next);
1516 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape)
1517 worklist.push(nk);
1518 }
1519 // mark all nodes reachable from ArgEscape nodes
1520 while(worklist.length() > 0) {
1521 PointsToNode* ptn = ptnode_adr(worklist.pop());
1522 if (ptn->node_type() == PointsToNode::JavaObject)
1523 has_non_escaping_obj = true; // Non GlobalEscape
1524 uint e_cnt = ptn->edge_count();
1525 for (uint ei = 0; ei < e_cnt; ei++) {
1526 uint npi = ptn->edge_target(ei);
1527 PointsToNode *np = ptnode_adr(npi);
1528 if (np->escape_state() < PointsToNode::ArgEscape) {
1529 np->set_escape_state(PointsToNode::ArgEscape);
1530 worklist.push(npi);
1531 }
1532 }
1533 }
1535 GrowableArray<Node*> alloc_worklist;
1537 // push all NoEscape nodes on the worklist
1538 for( uint next = 0; next < cg_length; ++next ) {
1539 int nk = cg_worklist.at(next);
1540 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape)
1541 worklist.push(nk);
1542 }
1543 // mark all nodes reachable from NoEscape nodes
1544 while(worklist.length() > 0) {
1545 PointsToNode* ptn = ptnode_adr(worklist.pop());
1546 if (ptn->node_type() == PointsToNode::JavaObject)
1547 has_non_escaping_obj = true; // Non GlobalEscape
1548 Node* n = ptn->_node;
1549 if (n->is_Allocate() && ptn->_scalar_replaceable ) {
1550 // Push scalar replaceable alocations on alloc_worklist
1551 // for processing in split_unique_types().
1552 alloc_worklist.append(n);
1553 }
1554 uint e_cnt = ptn->edge_count();
1555 for (uint ei = 0; ei < e_cnt; ei++) {
1556 uint npi = ptn->edge_target(ei);
1557 PointsToNode *np = ptnode_adr(npi);
1558 if (np->escape_state() < PointsToNode::NoEscape) {
1559 np->set_escape_state(PointsToNode::NoEscape);
1560 worklist.push(npi);
1561 }
1562 }
1563 }
1565 _collecting = false;
1566 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1568 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0;
1569 if ( has_scalar_replaceable_candidates &&
1570 C->AliasLevel() >= 3 && EliminateAllocations ) {
1572 // Now use the escape information to create unique types for
1573 // scalar replaceable objects.
1574 split_unique_types(alloc_worklist);
1576 if (C->failing()) return false;
1578 // Clean up after split unique types.
1579 ResourceMark rm;
1580 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn());
1582 C->print_method("After Escape Analysis", 2);
1584 #ifdef ASSERT
1585 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1586 tty->print("=== No allocations eliminated for ");
1587 C->method()->print_short_name();
1588 if(!EliminateAllocations) {
1589 tty->print(" since EliminateAllocations is off ===");
1590 } else if(!has_scalar_replaceable_candidates) {
1591 tty->print(" since there are no scalar replaceable candidates ===");
1592 } else if(C->AliasLevel() < 3) {
1593 tty->print(" since AliasLevel < 3 ===");
1594 }
1595 tty->cr();
1596 #endif
1597 }
1598 return has_non_escaping_obj;
1599 }
1601 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
1603 switch (call->Opcode()) {
1604 #ifdef ASSERT
1605 case Op_Allocate:
1606 case Op_AllocateArray:
1607 case Op_Lock:
1608 case Op_Unlock:
1609 assert(false, "should be done already");
1610 break;
1611 #endif
1612 case Op_CallLeafNoFP:
1613 {
1614 // Stub calls, objects do not escape but they are not scale replaceable.
1615 // Adjust escape state for outgoing arguments.
1616 const TypeTuple * d = call->tf()->domain();
1617 VectorSet ptset(Thread::current()->resource_area());
1618 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1619 const Type* at = d->field_at(i);
1620 Node *arg = call->in(i)->uncast();
1621 const Type *aat = phase->type(arg);
1622 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr()) {
1623 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
1624 aat->isa_ptr() != NULL, "expecting an Ptr");
1625 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1626 if (arg->is_AddP()) {
1627 //
1628 // The inline_native_clone() case when the arraycopy stub is called
1629 // after the allocation before Initialize and CheckCastPP nodes.
1630 //
1631 // Set AddP's base (Allocate) as not scalar replaceable since
1632 // pointer to the base (with offset) is passed as argument.
1633 //
1634 arg = get_addp_base(arg);
1635 }
1636 ptset.Clear();
1637 PointsTo(ptset, arg, phase);
1638 for( VectorSetI j(&ptset); j.test(); ++j ) {
1639 uint pt = j.elem;
1640 set_escape_state(pt, PointsToNode::ArgEscape);
1641 }
1642 }
1643 }
1644 break;
1645 }
1647 case Op_CallStaticJava:
1648 // For a static call, we know exactly what method is being called.
1649 // Use bytecode estimator to record the call's escape affects
1650 {
1651 ciMethod *meth = call->as_CallJava()->method();
1652 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1653 // fall-through if not a Java method or no analyzer information
1654 if (call_analyzer != NULL) {
1655 const TypeTuple * d = call->tf()->domain();
1656 VectorSet ptset(Thread::current()->resource_area());
1657 bool copy_dependencies = false;
1658 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1659 const Type* at = d->field_at(i);
1660 int k = i - TypeFunc::Parms;
1662 if (at->isa_oopptr() != NULL) {
1663 Node *arg = call->in(i)->uncast();
1665 bool global_escapes = false;
1666 bool fields_escapes = false;
1667 if (!call_analyzer->is_arg_stack(k)) {
1668 // The argument global escapes, mark everything it could point to
1669 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
1670 global_escapes = true;
1671 } else {
1672 if (!call_analyzer->is_arg_local(k)) {
1673 // The argument itself doesn't escape, but any fields might
1674 fields_escapes = true;
1675 }
1676 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1677 copy_dependencies = true;
1678 }
1680 ptset.Clear();
1681 PointsTo(ptset, arg, phase);
1682 for( VectorSetI j(&ptset); j.test(); ++j ) {
1683 uint pt = j.elem;
1684 if (global_escapes) {
1685 //The argument global escapes, mark everything it could point to
1686 set_escape_state(pt, PointsToNode::GlobalEscape);
1687 } else {
1688 if (fields_escapes) {
1689 // The argument itself doesn't escape, but any fields might
1690 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
1691 }
1692 set_escape_state(pt, PointsToNode::ArgEscape);
1693 }
1694 }
1695 }
1696 }
1697 if (copy_dependencies)
1698 call_analyzer->copy_dependencies(_compile->dependencies());
1699 break;
1700 }
1701 }
1703 default:
1704 // Fall-through here if not a Java method or no analyzer information
1705 // or some other type of call, assume the worst case: all arguments
1706 // globally escape.
1707 {
1708 // adjust escape state for outgoing arguments
1709 const TypeTuple * d = call->tf()->domain();
1710 VectorSet ptset(Thread::current()->resource_area());
1711 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1712 const Type* at = d->field_at(i);
1713 if (at->isa_oopptr() != NULL) {
1714 Node *arg = call->in(i)->uncast();
1715 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
1716 ptset.Clear();
1717 PointsTo(ptset, arg, phase);
1718 for( VectorSetI j(&ptset); j.test(); ++j ) {
1719 uint pt = j.elem;
1720 set_escape_state(pt, PointsToNode::GlobalEscape);
1721 }
1722 }
1723 }
1724 }
1725 }
1726 }
1727 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
1728 CallNode *call = resproj->in(0)->as_Call();
1729 uint call_idx = call->_idx;
1730 uint resproj_idx = resproj->_idx;
1732 switch (call->Opcode()) {
1733 case Op_Allocate:
1734 {
1735 Node *k = call->in(AllocateNode::KlassNode);
1736 const TypeKlassPtr *kt;
1737 if (k->Opcode() == Op_LoadKlass) {
1738 kt = k->as_Load()->type()->isa_klassptr();
1739 } else {
1740 // Also works for DecodeN(LoadNKlass).
1741 kt = k->as_Type()->type()->isa_klassptr();
1742 }
1743 assert(kt != NULL, "TypeKlassPtr required.");
1744 ciKlass* cik = kt->klass();
1745 ciInstanceKlass* ciik = cik->as_instance_klass();
1747 PointsToNode::EscapeState es;
1748 uint edge_to;
1749 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || ciik->has_finalizer()) {
1750 es = PointsToNode::GlobalEscape;
1751 edge_to = _phantom_object; // Could not be worse
1752 } else {
1753 es = PointsToNode::NoEscape;
1754 edge_to = call_idx;
1755 }
1756 set_escape_state(call_idx, es);
1757 add_pointsto_edge(resproj_idx, edge_to);
1758 _processed.set(resproj_idx);
1759 break;
1760 }
1762 case Op_AllocateArray:
1763 {
1764 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
1765 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
1766 // Not scalar replaceable if the length is not constant or too big.
1767 ptnode_adr(call_idx)->_scalar_replaceable = false;
1768 }
1769 set_escape_state(call_idx, PointsToNode::NoEscape);
1770 add_pointsto_edge(resproj_idx, call_idx);
1771 _processed.set(resproj_idx);
1772 break;
1773 }
1775 case Op_CallStaticJava:
1776 // For a static call, we know exactly what method is being called.
1777 // Use bytecode estimator to record whether the call's return value escapes
1778 {
1779 bool done = true;
1780 const TypeTuple *r = call->tf()->range();
1781 const Type* ret_type = NULL;
1783 if (r->cnt() > TypeFunc::Parms)
1784 ret_type = r->field_at(TypeFunc::Parms);
1786 // Note: we use isa_ptr() instead of isa_oopptr() here because the
1787 // _multianewarray functions return a TypeRawPtr.
1788 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
1789 _processed.set(resproj_idx);
1790 break; // doesn't return a pointer type
1791 }
1792 ciMethod *meth = call->as_CallJava()->method();
1793 const TypeTuple * d = call->tf()->domain();
1794 if (meth == NULL) {
1795 // not a Java method, assume global escape
1796 set_escape_state(call_idx, PointsToNode::GlobalEscape);
1797 add_pointsto_edge(resproj_idx, _phantom_object);
1798 } else {
1799 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
1800 bool copy_dependencies = false;
1802 if (call_analyzer->is_return_allocated()) {
1803 // Returns a newly allocated unescaped object, simply
1804 // update dependency information.
1805 // Mark it as NoEscape so that objects referenced by
1806 // it's fields will be marked as NoEscape at least.
1807 set_escape_state(call_idx, PointsToNode::NoEscape);
1808 add_pointsto_edge(resproj_idx, call_idx);
1809 copy_dependencies = true;
1810 } else if (call_analyzer->is_return_local()) {
1811 // determine whether any arguments are returned
1812 set_escape_state(call_idx, PointsToNode::NoEscape);
1813 bool ret_arg = false;
1814 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1815 const Type* at = d->field_at(i);
1817 if (at->isa_oopptr() != NULL) {
1818 Node *arg = call->in(i)->uncast();
1820 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
1821 ret_arg = true;
1822 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
1823 if (arg_esp->node_type() == PointsToNode::UnknownType)
1824 done = false;
1825 else if (arg_esp->node_type() == PointsToNode::JavaObject)
1826 add_pointsto_edge(resproj_idx, arg->_idx);
1827 else
1828 add_deferred_edge(resproj_idx, arg->_idx);
1829 arg_esp->_hidden_alias = true;
1830 }
1831 }
1832 }
1833 if (done && !ret_arg) {
1834 // Returns unknown object.
1835 set_escape_state(call_idx, PointsToNode::GlobalEscape);
1836 add_pointsto_edge(resproj_idx, _phantom_object);
1837 }
1838 copy_dependencies = true;
1839 } else {
1840 set_escape_state(call_idx, PointsToNode::GlobalEscape);
1841 add_pointsto_edge(resproj_idx, _phantom_object);
1842 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1843 const Type* at = d->field_at(i);
1844 if (at->isa_oopptr() != NULL) {
1845 Node *arg = call->in(i)->uncast();
1846 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
1847 arg_esp->_hidden_alias = true;
1848 }
1849 }
1850 }
1851 if (copy_dependencies)
1852 call_analyzer->copy_dependencies(_compile->dependencies());
1853 }
1854 if (done)
1855 _processed.set(resproj_idx);
1856 break;
1857 }
1859 default:
1860 // Some other type of call, assume the worst case that the
1861 // returned value, if any, globally escapes.
1862 {
1863 const TypeTuple *r = call->tf()->range();
1864 if (r->cnt() > TypeFunc::Parms) {
1865 const Type* ret_type = r->field_at(TypeFunc::Parms);
1867 // Note: we use isa_ptr() instead of isa_oopptr() here because the
1868 // _multianewarray functions return a TypeRawPtr.
1869 if (ret_type->isa_ptr() != NULL) {
1870 set_escape_state(call_idx, PointsToNode::GlobalEscape);
1871 add_pointsto_edge(resproj_idx, _phantom_object);
1872 }
1873 }
1874 _processed.set(resproj_idx);
1875 }
1876 }
1877 }
1879 // Populate Connection Graph with Ideal nodes and create simple
1880 // connection graph edges (do not need to check the node_type of inputs
1881 // or to call PointsTo() to walk the connection graph).
1882 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
1883 if (_processed.test(n->_idx))
1884 return; // No need to redefine node's state.
1886 if (n->is_Call()) {
1887 // Arguments to allocation and locking don't escape.
1888 if (n->is_Allocate()) {
1889 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
1890 record_for_optimizer(n);
1891 } else if (n->is_Lock() || n->is_Unlock()) {
1892 // Put Lock and Unlock nodes on IGVN worklist to process them during
1893 // the first IGVN optimization when escape information is still available.
1894 record_for_optimizer(n);
1895 _processed.set(n->_idx);
1896 } else {
1897 // Have to process call's arguments first.
1898 PointsToNode::NodeType nt = PointsToNode::UnknownType;
1900 // Check if a call returns an object.
1901 const TypeTuple *r = n->as_Call()->tf()->range();
1902 if (n->is_CallStaticJava() && r->cnt() > TypeFunc::Parms &&
1903 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
1904 // Note: use isa_ptr() instead of isa_oopptr() here because
1905 // the _multianewarray functions return a TypeRawPtr.
1906 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
1907 nt = PointsToNode::JavaObject;
1908 }
1909 }
1910 add_node(n, nt, PointsToNode::UnknownEscape, false);
1911 }
1912 return;
1913 }
1915 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1916 // ThreadLocal has RawPrt type.
1917 switch (n->Opcode()) {
1918 case Op_AddP:
1919 {
1920 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
1921 break;
1922 }
1923 case Op_CastX2P:
1924 { // "Unsafe" memory access.
1925 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
1926 break;
1927 }
1928 case Op_CastPP:
1929 case Op_CheckCastPP:
1930 case Op_EncodeP:
1931 case Op_DecodeN:
1932 {
1933 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
1934 int ti = n->in(1)->_idx;
1935 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
1936 if (nt == PointsToNode::UnknownType) {
1937 _delayed_worklist.push(n); // Process it later.
1938 break;
1939 } else if (nt == PointsToNode::JavaObject) {
1940 add_pointsto_edge(n->_idx, ti);
1941 } else {
1942 add_deferred_edge(n->_idx, ti);
1943 }
1944 _processed.set(n->_idx);
1945 break;
1946 }
1947 case Op_ConP:
1948 {
1949 // assume all pointer constants globally escape except for null
1950 PointsToNode::EscapeState es;
1951 if (phase->type(n) == TypePtr::NULL_PTR)
1952 es = PointsToNode::NoEscape;
1953 else
1954 es = PointsToNode::GlobalEscape;
1956 add_node(n, PointsToNode::JavaObject, es, true);
1957 break;
1958 }
1959 case Op_ConN:
1960 {
1961 // assume all narrow oop constants globally escape except for null
1962 PointsToNode::EscapeState es;
1963 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
1964 es = PointsToNode::NoEscape;
1965 else
1966 es = PointsToNode::GlobalEscape;
1968 add_node(n, PointsToNode::JavaObject, es, true);
1969 break;
1970 }
1971 case Op_CreateEx:
1972 {
1973 // assume that all exception objects globally escape
1974 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
1975 break;
1976 }
1977 case Op_LoadKlass:
1978 case Op_LoadNKlass:
1979 {
1980 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
1981 break;
1982 }
1983 case Op_LoadP:
1984 case Op_LoadN:
1985 {
1986 const Type *t = phase->type(n);
1987 if (t->make_ptr() == NULL) {
1988 _processed.set(n->_idx);
1989 return;
1990 }
1991 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
1992 break;
1993 }
1994 case Op_Parm:
1995 {
1996 _processed.set(n->_idx); // No need to redefine it state.
1997 uint con = n->as_Proj()->_con;
1998 if (con < TypeFunc::Parms)
1999 return;
2000 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2001 if (t->isa_ptr() == NULL)
2002 return;
2003 // We have to assume all input parameters globally escape
2004 // (Note: passing 'false' since _processed is already set).
2005 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2006 break;
2007 }
2008 case Op_Phi:
2009 {
2010 const Type *t = n->as_Phi()->type();
2011 if (t->make_ptr() == NULL) {
2012 // nothing to do if not an oop or narrow oop
2013 _processed.set(n->_idx);
2014 return;
2015 }
2016 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2017 uint i;
2018 for (i = 1; i < n->req() ; i++) {
2019 Node* in = n->in(i);
2020 if (in == NULL)
2021 continue; // ignore NULL
2022 in = in->uncast();
2023 if (in->is_top() || in == n)
2024 continue; // ignore top or inputs which go back this node
2025 int ti = in->_idx;
2026 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2027 if (nt == PointsToNode::UnknownType) {
2028 break;
2029 } else if (nt == PointsToNode::JavaObject) {
2030 add_pointsto_edge(n->_idx, ti);
2031 } else {
2032 add_deferred_edge(n->_idx, ti);
2033 }
2034 }
2035 if (i >= n->req())
2036 _processed.set(n->_idx);
2037 else
2038 _delayed_worklist.push(n);
2039 break;
2040 }
2041 case Op_Proj:
2042 {
2043 // we are only interested in the result projection from a call
2044 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2045 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2046 process_call_result(n->as_Proj(), phase);
2047 if (!_processed.test(n->_idx)) {
2048 // The call's result may need to be processed later if the call
2049 // returns it's argument and the argument is not processed yet.
2050 _delayed_worklist.push(n);
2051 }
2052 } else {
2053 _processed.set(n->_idx);
2054 }
2055 break;
2056 }
2057 case Op_Return:
2058 {
2059 if( n->req() > TypeFunc::Parms &&
2060 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2061 // Treat Return value as LocalVar with GlobalEscape escape state.
2062 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2063 int ti = n->in(TypeFunc::Parms)->_idx;
2064 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2065 if (nt == PointsToNode::UnknownType) {
2066 _delayed_worklist.push(n); // Process it later.
2067 break;
2068 } else if (nt == PointsToNode::JavaObject) {
2069 add_pointsto_edge(n->_idx, ti);
2070 } else {
2071 add_deferred_edge(n->_idx, ti);
2072 }
2073 }
2074 _processed.set(n->_idx);
2075 break;
2076 }
2077 case Op_StoreP:
2078 case Op_StoreN:
2079 {
2080 const Type *adr_type = phase->type(n->in(MemNode::Address));
2081 adr_type = adr_type->make_ptr();
2082 if (adr_type->isa_oopptr()) {
2083 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2084 } else {
2085 Node* adr = n->in(MemNode::Address);
2086 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2087 adr->in(AddPNode::Address)->is_Proj() &&
2088 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2089 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2090 // We are computing a raw address for a store captured
2091 // by an Initialize compute an appropriate address type.
2092 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2093 assert(offs != Type::OffsetBot, "offset must be a constant");
2094 } else {
2095 _processed.set(n->_idx);
2096 return;
2097 }
2098 }
2099 break;
2100 }
2101 case Op_StorePConditional:
2102 case Op_CompareAndSwapP:
2103 case Op_CompareAndSwapN:
2104 {
2105 const Type *adr_type = phase->type(n->in(MemNode::Address));
2106 adr_type = adr_type->make_ptr();
2107 if (adr_type->isa_oopptr()) {
2108 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2109 } else {
2110 _processed.set(n->_idx);
2111 return;
2112 }
2113 break;
2114 }
2115 case Op_ThreadLocal:
2116 {
2117 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2118 break;
2119 }
2120 default:
2121 ;
2122 // nothing to do
2123 }
2124 return;
2125 }
2127 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2128 uint n_idx = n->_idx;
2130 // Don't set processed bit for AddP, LoadP, StoreP since
2131 // they may need more then one pass to process.
2132 if (_processed.test(n_idx))
2133 return; // No need to redefine node's state.
2135 if (n->is_Call()) {
2136 CallNode *call = n->as_Call();
2137 process_call_arguments(call, phase);
2138 _processed.set(n_idx);
2139 return;
2140 }
2142 switch (n->Opcode()) {
2143 case Op_AddP:
2144 {
2145 Node *base = get_addp_base(n);
2146 // Create a field edge to this node from everything base could point to.
2147 VectorSet ptset(Thread::current()->resource_area());
2148 PointsTo(ptset, base, phase);
2149 for( VectorSetI i(&ptset); i.test(); ++i ) {
2150 uint pt = i.elem;
2151 add_field_edge(pt, n_idx, address_offset(n, phase));
2152 }
2153 break;
2154 }
2155 case Op_CastX2P:
2156 {
2157 assert(false, "Op_CastX2P");
2158 break;
2159 }
2160 case Op_CastPP:
2161 case Op_CheckCastPP:
2162 case Op_EncodeP:
2163 case Op_DecodeN:
2164 {
2165 int ti = n->in(1)->_idx;
2166 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2167 add_pointsto_edge(n_idx, ti);
2168 } else {
2169 add_deferred_edge(n_idx, ti);
2170 }
2171 _processed.set(n_idx);
2172 break;
2173 }
2174 case Op_ConP:
2175 {
2176 assert(false, "Op_ConP");
2177 break;
2178 }
2179 case Op_ConN:
2180 {
2181 assert(false, "Op_ConN");
2182 break;
2183 }
2184 case Op_CreateEx:
2185 {
2186 assert(false, "Op_CreateEx");
2187 break;
2188 }
2189 case Op_LoadKlass:
2190 case Op_LoadNKlass:
2191 {
2192 assert(false, "Op_LoadKlass");
2193 break;
2194 }
2195 case Op_LoadP:
2196 case Op_LoadN:
2197 {
2198 const Type *t = phase->type(n);
2199 #ifdef ASSERT
2200 if (t->make_ptr() == NULL)
2201 assert(false, "Op_LoadP");
2202 #endif
2204 Node* adr = n->in(MemNode::Address)->uncast();
2205 const Type *adr_type = phase->type(adr);
2206 Node* adr_base;
2207 if (adr->is_AddP()) {
2208 adr_base = get_addp_base(adr);
2209 } else {
2210 adr_base = adr;
2211 }
2213 // For everything "adr_base" could point to, create a deferred edge from
2214 // this node to each field with the same offset.
2215 VectorSet ptset(Thread::current()->resource_area());
2216 PointsTo(ptset, adr_base, phase);
2217 int offset = address_offset(adr, phase);
2218 for( VectorSetI i(&ptset); i.test(); ++i ) {
2219 uint pt = i.elem;
2220 add_deferred_edge_to_fields(n_idx, pt, offset);
2221 }
2222 break;
2223 }
2224 case Op_Parm:
2225 {
2226 assert(false, "Op_Parm");
2227 break;
2228 }
2229 case Op_Phi:
2230 {
2231 #ifdef ASSERT
2232 const Type *t = n->as_Phi()->type();
2233 if (t->make_ptr() == NULL)
2234 assert(false, "Op_Phi");
2235 #endif
2236 for (uint i = 1; i < n->req() ; i++) {
2237 Node* in = n->in(i);
2238 if (in == NULL)
2239 continue; // ignore NULL
2240 in = in->uncast();
2241 if (in->is_top() || in == n)
2242 continue; // ignore top or inputs which go back this node
2243 int ti = in->_idx;
2244 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2245 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
2246 if (nt == PointsToNode::JavaObject) {
2247 add_pointsto_edge(n_idx, ti);
2248 } else {
2249 add_deferred_edge(n_idx, ti);
2250 }
2251 }
2252 _processed.set(n_idx);
2253 break;
2254 }
2255 case Op_Proj:
2256 {
2257 // we are only interested in the result projection from a call
2258 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2259 process_call_result(n->as_Proj(), phase);
2260 assert(_processed.test(n_idx), "all call results should be processed");
2261 } else {
2262 assert(false, "Op_Proj");
2263 }
2264 break;
2265 }
2266 case Op_Return:
2267 {
2268 #ifdef ASSERT
2269 if( n->req() <= TypeFunc::Parms ||
2270 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2271 assert(false, "Op_Return");
2272 }
2273 #endif
2274 int ti = n->in(TypeFunc::Parms)->_idx;
2275 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2276 add_pointsto_edge(n_idx, ti);
2277 } else {
2278 add_deferred_edge(n_idx, ti);
2279 }
2280 _processed.set(n_idx);
2281 break;
2282 }
2283 case Op_StoreP:
2284 case Op_StoreN:
2285 case Op_StorePConditional:
2286 case Op_CompareAndSwapP:
2287 case Op_CompareAndSwapN:
2288 {
2289 Node *adr = n->in(MemNode::Address);
2290 const Type *adr_type = phase->type(adr)->make_ptr();
2291 #ifdef ASSERT
2292 if (!adr_type->isa_oopptr())
2293 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
2294 #endif
2296 assert(adr->is_AddP(), "expecting an AddP");
2297 Node *adr_base = get_addp_base(adr);
2298 Node *val = n->in(MemNode::ValueIn)->uncast();
2299 // For everything "adr_base" could point to, create a deferred edge
2300 // to "val" from each field with the same offset.
2301 VectorSet ptset(Thread::current()->resource_area());
2302 PointsTo(ptset, adr_base, phase);
2303 for( VectorSetI i(&ptset); i.test(); ++i ) {
2304 uint pt = i.elem;
2305 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase));
2306 }
2307 break;
2308 }
2309 case Op_ThreadLocal:
2310 {
2311 assert(false, "Op_ThreadLocal");
2312 break;
2313 }
2314 default:
2315 ;
2316 // nothing to do
2317 }
2318 }
2320 #ifndef PRODUCT
2321 void ConnectionGraph::dump() {
2322 PhaseGVN *igvn = _compile->initial_gvn();
2323 bool first = true;
2325 uint size = nodes_size();
2326 for (uint ni = 0; ni < size; ni++) {
2327 PointsToNode *ptn = ptnode_adr(ni);
2328 PointsToNode::NodeType ptn_type = ptn->node_type();
2330 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
2331 continue;
2332 PointsToNode::EscapeState es = escape_state(ptn->_node, igvn);
2333 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
2334 if (first) {
2335 tty->cr();
2336 tty->print("======== Connection graph for ");
2337 _compile->method()->print_short_name();
2338 tty->cr();
2339 first = false;
2340 }
2341 tty->print("%6d ", ni);
2342 ptn->dump();
2343 // Print all locals which reference this allocation
2344 for (uint li = ni; li < size; li++) {
2345 PointsToNode *ptn_loc = ptnode_adr(li);
2346 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
2347 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
2348 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
2349 ptnode_adr(li)->dump(false);
2350 }
2351 }
2352 if (Verbose) {
2353 // Print all fields which reference this allocation
2354 for (uint i = 0; i < ptn->edge_count(); i++) {
2355 uint ei = ptn->edge_target(i);
2356 ptnode_adr(ei)->dump(false);
2357 }
2358 }
2359 tty->cr();
2360 }
2361 }
2362 }
2363 #endif