Tue, 23 Nov 2010 13:22:55 -0800
6989984: Use standard include model for Hospot
Summary: Replaced MakeDeps and the includeDB files with more standardized solutions.
Reviewed-by: coleenp, kvn, kamg
1 /*
2 * Copyright (c) 2005, 2010, 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 _collecting(true),
97 _progress(false),
98 _compile(C),
99 _igvn(igvn),
100 _node_map(C->comp_arena()) {
102 _phantom_object = C->top()->_idx,
103 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true);
105 // Add ConP(#NULL) and ConN(#NULL) nodes.
106 Node* oop_null = igvn->zerocon(T_OBJECT);
107 _oop_null = oop_null->_idx;
108 assert(_oop_null < C->unique(), "should be created already");
109 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
111 if (UseCompressedOops) {
112 Node* noop_null = igvn->zerocon(T_NARROWOOP);
113 _noop_null = noop_null->_idx;
114 assert(_noop_null < C->unique(), "should be created already");
115 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
116 }
117 }
119 void ConnectionGraph::add_pointsto_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 PointsTo edge");
125 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge");
126 add_edge(f, to_i, PointsToNode::PointsToEdge);
127 }
129 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) {
130 PointsToNode *f = ptnode_adr(from_i);
131 PointsToNode *t = ptnode_adr(to_i);
133 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
134 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
135 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
136 // don't add a self-referential edge, this can occur during removal of
137 // deferred edges
138 if (from_i != to_i)
139 add_edge(f, to_i, PointsToNode::DeferredEdge);
140 }
142 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
143 const Type *adr_type = phase->type(adr);
144 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
145 adr->in(AddPNode::Address)->is_Proj() &&
146 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
147 // We are computing a raw address for a store captured by an Initialize
148 // compute an appropriate address type. AddP cases #3 and #5 (see below).
149 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
150 assert(offs != Type::OffsetBot ||
151 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
152 "offset must be a constant or it is initialization of array");
153 return offs;
154 }
155 const TypePtr *t_ptr = adr_type->isa_ptr();
156 assert(t_ptr != NULL, "must be a pointer type");
157 return t_ptr->offset();
158 }
160 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) {
161 PointsToNode *f = ptnode_adr(from_i);
162 PointsToNode *t = ptnode_adr(to_i);
164 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
165 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge");
166 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge");
167 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets");
168 t->set_offset(offset);
170 add_edge(f, to_i, PointsToNode::FieldEdge);
171 }
173 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) {
174 PointsToNode *npt = ptnode_adr(ni);
175 PointsToNode::EscapeState old_es = npt->escape_state();
176 if (es > old_es)
177 npt->set_escape_state(es);
178 }
180 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt,
181 PointsToNode::EscapeState es, bool done) {
182 PointsToNode* ptadr = ptnode_adr(n->_idx);
183 ptadr->_node = n;
184 ptadr->set_node_type(nt);
186 // inline set_escape_state(idx, es);
187 PointsToNode::EscapeState old_es = ptadr->escape_state();
188 if (es > old_es)
189 ptadr->set_escape_state(es);
191 if (done)
192 _processed.set(n->_idx);
193 }
195 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) {
196 uint idx = n->_idx;
197 PointsToNode::EscapeState es;
199 // If we are still collecting or there were no non-escaping allocations
200 // we don't know the answer yet
201 if (_collecting)
202 return PointsToNode::UnknownEscape;
204 // if the node was created after the escape computation, return
205 // UnknownEscape
206 if (idx >= nodes_size())
207 return PointsToNode::UnknownEscape;
209 es = ptnode_adr(idx)->escape_state();
211 // if we have already computed a value, return it
212 if (es != PointsToNode::UnknownEscape &&
213 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject)
214 return es;
216 // PointsTo() calls n->uncast() which can return a new ideal node.
217 if (n->uncast()->_idx >= nodes_size())
218 return PointsToNode::UnknownEscape;
220 PointsToNode::EscapeState orig_es = es;
222 // compute max escape state of anything this node could point to
223 VectorSet ptset(Thread::current()->resource_area());
224 PointsTo(ptset, n);
225 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) {
226 uint pt = i.elem;
227 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
228 if (pes > es)
229 es = pes;
230 }
231 if (orig_es != es) {
232 // cache the computed escape state
233 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state");
234 ptnode_adr(idx)->set_escape_state(es);
235 } // orig_es could be PointsToNode::UnknownEscape
236 return es;
237 }
239 void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n) {
240 VectorSet visited(Thread::current()->resource_area());
241 GrowableArray<uint> worklist;
243 #ifdef ASSERT
244 Node *orig_n = n;
245 #endif
247 n = n->uncast();
248 PointsToNode* npt = ptnode_adr(n->_idx);
250 // If we have a JavaObject, return just that object
251 if (npt->node_type() == PointsToNode::JavaObject) {
252 ptset.set(n->_idx);
253 return;
254 }
255 #ifdef ASSERT
256 if (npt->_node == NULL) {
257 if (orig_n != n)
258 orig_n->dump();
259 n->dump();
260 assert(npt->_node != NULL, "unregistered node");
261 }
262 #endif
263 worklist.push(n->_idx);
264 while(worklist.length() > 0) {
265 int ni = worklist.pop();
266 if (visited.test_set(ni))
267 continue;
269 PointsToNode* pn = ptnode_adr(ni);
270 // ensure that all inputs of a Phi have been processed
271 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
273 int edges_processed = 0;
274 uint e_cnt = pn->edge_count();
275 for (uint e = 0; e < e_cnt; e++) {
276 uint etgt = pn->edge_target(e);
277 PointsToNode::EdgeType et = pn->edge_type(e);
278 if (et == PointsToNode::PointsToEdge) {
279 ptset.set(etgt);
280 edges_processed++;
281 } else if (et == PointsToNode::DeferredEdge) {
282 worklist.push(etgt);
283 edges_processed++;
284 } else {
285 assert(false,"neither PointsToEdge or DeferredEdge");
286 }
287 }
288 if (edges_processed == 0) {
289 // no deferred or pointsto edges found. Assume the value was set
290 // outside this method. Add the phantom object to the pointsto set.
291 ptset.set(_phantom_object);
292 }
293 }
294 }
296 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
297 // This method is most expensive during ConnectionGraph construction.
298 // Reuse vectorSet and an additional growable array for deferred edges.
299 deferred_edges->clear();
300 visited->Clear();
302 visited->set(ni);
303 PointsToNode *ptn = ptnode_adr(ni);
305 // Mark current edges as visited and move deferred edges to separate array.
306 for (uint i = 0; i < ptn->edge_count(); ) {
307 uint t = ptn->edge_target(i);
308 #ifdef ASSERT
309 assert(!visited->test_set(t), "expecting no duplications");
310 #else
311 visited->set(t);
312 #endif
313 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
314 ptn->remove_edge(t, PointsToNode::DeferredEdge);
315 deferred_edges->append(t);
316 } else {
317 i++;
318 }
319 }
320 for (int next = 0; next < deferred_edges->length(); ++next) {
321 uint t = deferred_edges->at(next);
322 PointsToNode *ptt = ptnode_adr(t);
323 uint e_cnt = ptt->edge_count();
324 for (uint e = 0; e < e_cnt; e++) {
325 uint etgt = ptt->edge_target(e);
326 if (visited->test_set(etgt))
327 continue;
329 PointsToNode::EdgeType et = ptt->edge_type(e);
330 if (et == PointsToNode::PointsToEdge) {
331 add_pointsto_edge(ni, etgt);
332 if(etgt == _phantom_object) {
333 // Special case - field set outside (globally escaping).
334 ptn->set_escape_state(PointsToNode::GlobalEscape);
335 }
336 } else if (et == PointsToNode::DeferredEdge) {
337 deferred_edges->append(etgt);
338 } else {
339 assert(false,"invalid connection graph");
340 }
341 }
342 }
343 }
346 // Add an edge to node given by "to_i" from any field of adr_i whose offset
347 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
348 // a pointsto edge is added if it is a JavaObject
350 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
351 PointsToNode* an = ptnode_adr(adr_i);
352 PointsToNode* to = ptnode_adr(to_i);
353 bool deferred = (to->node_type() == PointsToNode::LocalVar);
355 for (uint fe = 0; fe < an->edge_count(); fe++) {
356 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
357 int fi = an->edge_target(fe);
358 PointsToNode* pf = ptnode_adr(fi);
359 int po = pf->offset();
360 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
361 if (deferred)
362 add_deferred_edge(fi, to_i);
363 else
364 add_pointsto_edge(fi, to_i);
365 }
366 }
367 }
369 // Add a deferred edge from node given by "from_i" to any field of adr_i
370 // whose offset matches "offset".
371 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
372 PointsToNode* an = ptnode_adr(adr_i);
373 for (uint fe = 0; fe < an->edge_count(); fe++) {
374 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
375 int fi = an->edge_target(fe);
376 PointsToNode* pf = ptnode_adr(fi);
377 int po = pf->offset();
378 if (pf->edge_count() == 0) {
379 // we have not seen any stores to this field, assume it was set outside this method
380 add_pointsto_edge(fi, _phantom_object);
381 }
382 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
383 add_deferred_edge(from_i, fi);
384 }
385 }
386 }
388 // Helper functions
390 static Node* get_addp_base(Node *addp) {
391 assert(addp->is_AddP(), "must be AddP");
392 //
393 // AddP cases for Base and Address inputs:
394 // case #1. Direct object's field reference:
395 // Allocate
396 // |
397 // Proj #5 ( oop result )
398 // |
399 // CheckCastPP (cast to instance type)
400 // | |
401 // AddP ( base == address )
402 //
403 // case #2. Indirect object's field reference:
404 // Phi
405 // |
406 // CastPP (cast to instance type)
407 // | |
408 // AddP ( base == address )
409 //
410 // case #3. Raw object's field reference for Initialize node:
411 // Allocate
412 // |
413 // Proj #5 ( oop result )
414 // top |
415 // \ |
416 // AddP ( base == top )
417 //
418 // case #4. Array's element reference:
419 // {CheckCastPP | CastPP}
420 // | | |
421 // | AddP ( array's element offset )
422 // | |
423 // AddP ( array's offset )
424 //
425 // case #5. Raw object's field reference for arraycopy stub call:
426 // The inline_native_clone() case when the arraycopy stub is called
427 // after the allocation before Initialize and CheckCastPP nodes.
428 // Allocate
429 // |
430 // Proj #5 ( oop result )
431 // | |
432 // AddP ( base == address )
433 //
434 // case #6. Constant Pool, ThreadLocal, CastX2P or
435 // Raw object's field reference:
436 // {ConP, ThreadLocal, CastX2P, raw Load}
437 // top |
438 // \ |
439 // AddP ( base == top )
440 //
441 // case #7. Klass's field reference.
442 // LoadKlass
443 // | |
444 // AddP ( base == address )
445 //
446 // case #8. narrow Klass's field reference.
447 // LoadNKlass
448 // |
449 // DecodeN
450 // | |
451 // AddP ( base == address )
452 //
453 Node *base = addp->in(AddPNode::Base)->uncast();
454 if (base->is_top()) { // The AddP case #3 and #6.
455 base = addp->in(AddPNode::Address)->uncast();
456 while (base->is_AddP()) {
457 // Case #6 (unsafe access) may have several chained AddP nodes.
458 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only");
459 base = base->in(AddPNode::Address)->uncast();
460 }
461 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
462 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
463 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
464 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
465 }
466 return base;
467 }
469 static Node* find_second_addp(Node* addp, Node* n) {
470 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
472 Node* addp2 = addp->raw_out(0);
473 if (addp->outcnt() == 1 && addp2->is_AddP() &&
474 addp2->in(AddPNode::Base) == n &&
475 addp2->in(AddPNode::Address) == addp) {
477 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
478 //
479 // Find array's offset to push it on worklist first and
480 // as result process an array's element offset first (pushed second)
481 // to avoid CastPP for the array's offset.
482 // Otherwise the inserted CastPP (LocalVar) will point to what
483 // the AddP (Field) points to. Which would be wrong since
484 // the algorithm expects the CastPP has the same point as
485 // as AddP's base CheckCastPP (LocalVar).
486 //
487 // ArrayAllocation
488 // |
489 // CheckCastPP
490 // |
491 // memProj (from ArrayAllocation CheckCastPP)
492 // | ||
493 // | || Int (element index)
494 // | || | ConI (log(element size))
495 // | || | /
496 // | || LShift
497 // | || /
498 // | AddP (array's element offset)
499 // | |
500 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
501 // | / /
502 // AddP (array's offset)
503 // |
504 // Load/Store (memory operation on array's element)
505 //
506 return addp2;
507 }
508 return NULL;
509 }
511 //
512 // Adjust the type and inputs of an AddP which computes the
513 // address of a field of an instance
514 //
515 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
516 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
517 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
518 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
519 if (t == NULL) {
520 // We are computing a raw address for a store captured by an Initialize
521 // compute an appropriate address type (cases #3 and #5).
522 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
523 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
524 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
525 assert(offs != Type::OffsetBot, "offset must be a constant");
526 t = base_t->add_offset(offs)->is_oopptr();
527 }
528 int inst_id = base_t->instance_id();
529 assert(!t->is_known_instance() || t->instance_id() == inst_id,
530 "old type must be non-instance or match new type");
532 // The type 't' could be subclass of 'base_t'.
533 // As result t->offset() could be large then base_t's size and it will
534 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
535 // constructor verifies correctness of the offset.
536 //
537 // It could happened on subclass's branch (from the type profiling
538 // inlining) which was not eliminated during parsing since the exactness
539 // of the allocation type was not propagated to the subclass type check.
540 //
541 // Or the type 't' could be not related to 'base_t' at all.
542 // It could happened when CHA type is different from MDO type on a dead path
543 // (for example, from instanceof check) which is not collapsed during parsing.
544 //
545 // Do nothing for such AddP node and don't process its users since
546 // this code branch will go away.
547 //
548 if (!t->is_known_instance() &&
549 !base_t->klass()->is_subtype_of(t->klass())) {
550 return false; // bail out
551 }
553 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
554 // Do NOT remove the next line: ensure a new alias index is allocated
555 // for the instance type. Note: C++ will not remove it since the call
556 // has side effect.
557 int alias_idx = _compile->get_alias_index(tinst);
558 igvn->set_type(addp, tinst);
559 // record the allocation in the node map
560 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered");
561 set_map(addp->_idx, get_map(base->_idx));
563 // Set addp's Base and Address to 'base'.
564 Node *abase = addp->in(AddPNode::Base);
565 Node *adr = addp->in(AddPNode::Address);
566 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
567 adr->in(0)->_idx == (uint)inst_id) {
568 // Skip AddP cases #3 and #5.
569 } else {
570 assert(!abase->is_top(), "sanity"); // AddP case #3
571 if (abase != base) {
572 igvn->hash_delete(addp);
573 addp->set_req(AddPNode::Base, base);
574 if (abase == adr) {
575 addp->set_req(AddPNode::Address, base);
576 } else {
577 // AddP case #4 (adr is array's element offset AddP node)
578 #ifdef ASSERT
579 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
580 assert(adr->is_AddP() && atype != NULL &&
581 atype->instance_id() == inst_id, "array's element offset should be processed first");
582 #endif
583 }
584 igvn->hash_insert(addp);
585 }
586 }
587 // Put on IGVN worklist since at least addp's type was changed above.
588 record_for_optimizer(addp);
589 return true;
590 }
592 //
593 // Create a new version of orig_phi if necessary. Returns either the newly
594 // created phi or an existing phi. Sets create_new to indicate wheter a new
595 // phi was created. Cache the last newly created phi in the node map.
596 //
597 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
598 Compile *C = _compile;
599 new_created = false;
600 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
601 // nothing to do if orig_phi is bottom memory or matches alias_idx
602 if (phi_alias_idx == alias_idx) {
603 return orig_phi;
604 }
605 // Have we recently created a Phi for this alias index?
606 PhiNode *result = get_map_phi(orig_phi->_idx);
607 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
608 return result;
609 }
610 // Previous check may fail when the same wide memory Phi was split into Phis
611 // for different memory slices. Search all Phis for this region.
612 if (result != NULL) {
613 Node* region = orig_phi->in(0);
614 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
615 Node* phi = region->fast_out(i);
616 if (phi->is_Phi() &&
617 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
618 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
619 return phi->as_Phi();
620 }
621 }
622 }
623 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
624 if (C->do_escape_analysis() == true && !C->failing()) {
625 // Retry compilation without escape analysis.
626 // If this is the first failure, the sentinel string will "stick"
627 // to the Compile object, and the C2Compiler will see it and retry.
628 C->record_failure(C2Compiler::retry_no_escape_analysis());
629 }
630 return NULL;
631 }
632 orig_phi_worklist.append_if_missing(orig_phi);
633 const TypePtr *atype = C->get_adr_type(alias_idx);
634 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
635 C->copy_node_notes_to(result, orig_phi);
636 igvn->set_type(result, result->bottom_type());
637 record_for_optimizer(result);
639 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;)
640 assert(pn == NULL || pn == orig_phi, "wrong node");
641 set_map(orig_phi->_idx, result);
642 ptnode_adr(orig_phi->_idx)->_node = orig_phi;
644 new_created = true;
645 return result;
646 }
648 //
649 // Return a new version of Memory Phi "orig_phi" with the inputs having the
650 // specified alias index.
651 //
652 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
654 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
655 Compile *C = _compile;
656 bool new_phi_created;
657 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
658 if (!new_phi_created) {
659 return result;
660 }
662 GrowableArray<PhiNode *> phi_list;
663 GrowableArray<uint> cur_input;
665 PhiNode *phi = orig_phi;
666 uint idx = 1;
667 bool finished = false;
668 while(!finished) {
669 while (idx < phi->req()) {
670 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
671 if (mem != NULL && mem->is_Phi()) {
672 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
673 if (new_phi_created) {
674 // found an phi for which we created a new split, push current one on worklist and begin
675 // processing new one
676 phi_list.push(phi);
677 cur_input.push(idx);
678 phi = mem->as_Phi();
679 result = newphi;
680 idx = 1;
681 continue;
682 } else {
683 mem = newphi;
684 }
685 }
686 if (C->failing()) {
687 return NULL;
688 }
689 result->set_req(idx++, mem);
690 }
691 #ifdef ASSERT
692 // verify that the new Phi has an input for each input of the original
693 assert( phi->req() == result->req(), "must have same number of inputs.");
694 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
695 #endif
696 // Check if all new phi's inputs have specified alias index.
697 // Otherwise use old phi.
698 for (uint i = 1; i < phi->req(); i++) {
699 Node* in = result->in(i);
700 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
701 }
702 // we have finished processing a Phi, see if there are any more to do
703 finished = (phi_list.length() == 0 );
704 if (!finished) {
705 phi = phi_list.pop();
706 idx = cur_input.pop();
707 PhiNode *prev_result = get_map_phi(phi->_idx);
708 prev_result->set_req(idx++, result);
709 result = prev_result;
710 }
711 }
712 return result;
713 }
716 //
717 // The next methods are derived from methods in MemNode.
718 //
719 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
720 Node *mem = mmem;
721 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
722 // means an array I have not precisely typed yet. Do not do any
723 // alias stuff with it any time soon.
724 if( toop->base() != Type::AnyPtr &&
725 !(toop->klass() != NULL &&
726 toop->klass()->is_java_lang_Object() &&
727 toop->offset() == Type::OffsetBot) ) {
728 mem = mmem->memory_at(alias_idx);
729 // Update input if it is progress over what we have now
730 }
731 return mem;
732 }
734 //
735 // Move memory users to their memory slices.
736 //
737 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) {
738 Compile* C = _compile;
740 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
741 assert(tp != NULL, "ptr type");
742 int alias_idx = C->get_alias_index(tp);
743 int general_idx = C->get_general_index(alias_idx);
745 // Move users first
746 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
747 Node* use = n->fast_out(i);
748 if (use->is_MergeMem()) {
749 MergeMemNode* mmem = use->as_MergeMem();
750 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
751 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
752 continue; // Nothing to do
753 }
754 // Replace previous general reference to mem node.
755 uint orig_uniq = C->unique();
756 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
757 assert(orig_uniq == C->unique(), "no new nodes");
758 mmem->set_memory_at(general_idx, m);
759 --imax;
760 --i;
761 } else if (use->is_MemBar()) {
762 assert(!use->is_Initialize(), "initializing stores should not be moved");
763 if (use->req() > MemBarNode::Precedent &&
764 use->in(MemBarNode::Precedent) == n) {
765 // Don't move related membars.
766 record_for_optimizer(use);
767 continue;
768 }
769 tp = use->as_MemBar()->adr_type()->isa_ptr();
770 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
771 alias_idx == general_idx) {
772 continue; // Nothing to do
773 }
774 // Move to general memory slice.
775 uint orig_uniq = C->unique();
776 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
777 assert(orig_uniq == C->unique(), "no new nodes");
778 igvn->hash_delete(use);
779 imax -= use->replace_edge(n, m);
780 igvn->hash_insert(use);
781 record_for_optimizer(use);
782 --i;
783 #ifdef ASSERT
784 } else if (use->is_Mem()) {
785 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
786 // Don't move related cardmark.
787 continue;
788 }
789 // Memory nodes should have new memory input.
790 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
791 assert(tp != NULL, "ptr type");
792 int idx = C->get_alias_index(tp);
793 assert(get_map(use->_idx) != NULL || idx == alias_idx,
794 "Following memory nodes should have new memory input or be on the same memory slice");
795 } else if (use->is_Phi()) {
796 // Phi nodes should be split and moved already.
797 tp = use->as_Phi()->adr_type()->isa_ptr();
798 assert(tp != NULL, "ptr type");
799 int idx = C->get_alias_index(tp);
800 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
801 } else {
802 use->dump();
803 assert(false, "should not be here");
804 #endif
805 }
806 }
807 }
809 //
810 // Search memory chain of "mem" to find a MemNode whose address
811 // is the specified alias index.
812 //
813 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
814 if (orig_mem == NULL)
815 return orig_mem;
816 Compile* C = phase->C;
817 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
818 bool is_instance = (toop != NULL) && toop->is_known_instance();
819 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
820 Node *prev = NULL;
821 Node *result = orig_mem;
822 while (prev != result) {
823 prev = result;
824 if (result == start_mem)
825 break; // hit one of our sentinels
826 if (result->is_Mem()) {
827 const Type *at = phase->type(result->in(MemNode::Address));
828 if (at != Type::TOP) {
829 assert (at->isa_ptr() != NULL, "pointer type required.");
830 int idx = C->get_alias_index(at->is_ptr());
831 if (idx == alias_idx)
832 break;
833 }
834 result = result->in(MemNode::Memory);
835 }
836 if (!is_instance)
837 continue; // don't search further for non-instance types
838 // skip over a call which does not affect this memory slice
839 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
840 Node *proj_in = result->in(0);
841 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
842 break; // hit one of our sentinels
843 } else if (proj_in->is_Call()) {
844 CallNode *call = proj_in->as_Call();
845 if (!call->may_modify(toop, phase)) {
846 result = call->in(TypeFunc::Memory);
847 }
848 } else if (proj_in->is_Initialize()) {
849 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
850 // Stop if this is the initialization for the object instance which
851 // which contains this memory slice, otherwise skip over it.
852 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
853 result = proj_in->in(TypeFunc::Memory);
854 }
855 } else if (proj_in->is_MemBar()) {
856 result = proj_in->in(TypeFunc::Memory);
857 }
858 } else if (result->is_MergeMem()) {
859 MergeMemNode *mmem = result->as_MergeMem();
860 result = step_through_mergemem(mmem, alias_idx, toop);
861 if (result == mmem->base_memory()) {
862 // Didn't find instance memory, search through general slice recursively.
863 result = mmem->memory_at(C->get_general_index(alias_idx));
864 result = find_inst_mem(result, alias_idx, orig_phis, phase);
865 if (C->failing()) {
866 return NULL;
867 }
868 mmem->set_memory_at(alias_idx, result);
869 }
870 } else if (result->is_Phi() &&
871 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
872 Node *un = result->as_Phi()->unique_input(phase);
873 if (un != NULL) {
874 orig_phis.append_if_missing(result->as_Phi());
875 result = un;
876 } else {
877 break;
878 }
879 } else if (result->is_ClearArray()) {
880 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) {
881 // Can not bypass initialization of the instance
882 // we are looking for.
883 break;
884 }
885 // Otherwise skip it (the call updated 'result' value).
886 } else if (result->Opcode() == Op_SCMemProj) {
887 assert(result->in(0)->is_LoadStore(), "sanity");
888 const Type *at = phase->type(result->in(0)->in(MemNode::Address));
889 if (at != Type::TOP) {
890 assert (at->isa_ptr() != NULL, "pointer type required.");
891 int idx = C->get_alias_index(at->is_ptr());
892 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
893 break;
894 }
895 result = result->in(0)->in(MemNode::Memory);
896 }
897 }
898 if (result->is_Phi()) {
899 PhiNode *mphi = result->as_Phi();
900 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
901 const TypePtr *t = mphi->adr_type();
902 if (C->get_alias_index(t) != alias_idx) {
903 // Create a new Phi with the specified alias index type.
904 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
905 } else if (!is_instance) {
906 // Push all non-instance Phis on the orig_phis worklist to update inputs
907 // during Phase 4 if needed.
908 orig_phis.append_if_missing(mphi);
909 }
910 }
911 // the result is either MemNode, PhiNode, InitializeNode.
912 return result;
913 }
915 //
916 // Convert the types of unescaped object to instance types where possible,
917 // propagate the new type information through the graph, and update memory
918 // edges and MergeMem inputs to reflect the new type.
919 //
920 // We start with allocations (and calls which may be allocations) on alloc_worklist.
921 // The processing is done in 4 phases:
922 //
923 // Phase 1: Process possible allocations from alloc_worklist. Create instance
924 // types for the CheckCastPP for allocations where possible.
925 // Propagate the the new types through users as follows:
926 // casts and Phi: push users on alloc_worklist
927 // AddP: cast Base and Address inputs to the instance type
928 // push any AddP users on alloc_worklist and push any memnode
929 // users onto memnode_worklist.
930 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
931 // search the Memory chain for a store with the appropriate type
932 // address type. If a Phi is found, create a new version with
933 // the appropriate memory slices from each of the Phi inputs.
934 // For stores, process the users as follows:
935 // MemNode: push on memnode_worklist
936 // MergeMem: push on mergemem_worklist
937 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
938 // moving the first node encountered of each instance type to the
939 // the input corresponding to its alias index.
940 // appropriate memory slice.
941 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
942 //
943 // In the following example, the CheckCastPP nodes are the cast of allocation
944 // results and the allocation of node 29 is unescaped and eligible to be an
945 // instance type.
946 //
947 // We start with:
948 //
949 // 7 Parm #memory
950 // 10 ConI "12"
951 // 19 CheckCastPP "Foo"
952 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
953 // 29 CheckCastPP "Foo"
954 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
955 //
956 // 40 StoreP 25 7 20 ... alias_index=4
957 // 50 StoreP 35 40 30 ... alias_index=4
958 // 60 StoreP 45 50 20 ... alias_index=4
959 // 70 LoadP _ 60 30 ... alias_index=4
960 // 80 Phi 75 50 60 Memory alias_index=4
961 // 90 LoadP _ 80 30 ... alias_index=4
962 // 100 LoadP _ 80 20 ... alias_index=4
963 //
964 //
965 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
966 // and creating a new alias index for node 30. This gives:
967 //
968 // 7 Parm #memory
969 // 10 ConI "12"
970 // 19 CheckCastPP "Foo"
971 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
972 // 29 CheckCastPP "Foo" iid=24
973 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
974 //
975 // 40 StoreP 25 7 20 ... alias_index=4
976 // 50 StoreP 35 40 30 ... alias_index=6
977 // 60 StoreP 45 50 20 ... alias_index=4
978 // 70 LoadP _ 60 30 ... alias_index=6
979 // 80 Phi 75 50 60 Memory alias_index=4
980 // 90 LoadP _ 80 30 ... alias_index=6
981 // 100 LoadP _ 80 20 ... alias_index=4
982 //
983 // In phase 2, new memory inputs are computed for the loads and stores,
984 // And a new version of the phi is created. In phase 4, the inputs to
985 // node 80 are updated and then the memory nodes are updated with the
986 // values computed in phase 2. This results in:
987 //
988 // 7 Parm #memory
989 // 10 ConI "12"
990 // 19 CheckCastPP "Foo"
991 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
992 // 29 CheckCastPP "Foo" iid=24
993 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
994 //
995 // 40 StoreP 25 7 20 ... alias_index=4
996 // 50 StoreP 35 7 30 ... alias_index=6
997 // 60 StoreP 45 40 20 ... alias_index=4
998 // 70 LoadP _ 50 30 ... alias_index=6
999 // 80 Phi 75 40 60 Memory alias_index=4
1000 // 120 Phi 75 50 50 Memory alias_index=6
1001 // 90 LoadP _ 120 30 ... alias_index=6
1002 // 100 LoadP _ 80 20 ... alias_index=4
1003 //
1004 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
1005 GrowableArray<Node *> memnode_worklist;
1006 GrowableArray<PhiNode *> orig_phis;
1008 PhaseIterGVN *igvn = _igvn;
1009 uint new_index_start = (uint) _compile->num_alias_types();
1010 Arena* arena = Thread::current()->resource_area();
1011 VectorSet visited(arena);
1012 VectorSet ptset(arena);
1015 // Phase 1: Process possible allocations from alloc_worklist.
1016 // Create instance types for the CheckCastPP for allocations where possible.
1017 //
1018 // (Note: don't forget to change the order of the second AddP node on
1019 // the alloc_worklist if the order of the worklist processing is changed,
1020 // see the comment in find_second_addp().)
1021 //
1022 while (alloc_worklist.length() != 0) {
1023 Node *n = alloc_worklist.pop();
1024 uint ni = n->_idx;
1025 const TypeOopPtr* tinst = NULL;
1026 if (n->is_Call()) {
1027 CallNode *alloc = n->as_Call();
1028 // copy escape information to call node
1029 PointsToNode* ptn = ptnode_adr(alloc->_idx);
1030 PointsToNode::EscapeState es = escape_state(alloc);
1031 // We have an allocation or call which returns a Java object,
1032 // see if it is unescaped.
1033 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable)
1034 continue;
1036 // Find CheckCastPP for the allocate or for the return value of a call
1037 n = alloc->result_cast();
1038 if (n == NULL) { // No uses except Initialize node
1039 if (alloc->is_Allocate()) {
1040 // Set the scalar_replaceable flag for allocation
1041 // so it could be eliminated if it has no uses.
1042 alloc->as_Allocate()->_is_scalar_replaceable = true;
1043 }
1044 continue;
1045 }
1046 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
1047 assert(!alloc->is_Allocate(), "allocation should have unique type");
1048 continue;
1049 }
1051 // The inline code for Object.clone() casts the allocation result to
1052 // java.lang.Object and then to the actual type of the allocated
1053 // object. Detect this case and use the second cast.
1054 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
1055 // the allocation result is cast to java.lang.Object and then
1056 // to the actual Array type.
1057 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
1058 && (alloc->is_AllocateArray() ||
1059 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
1060 Node *cast2 = NULL;
1061 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1062 Node *use = n->fast_out(i);
1063 if (use->is_CheckCastPP()) {
1064 cast2 = use;
1065 break;
1066 }
1067 }
1068 if (cast2 != NULL) {
1069 n = cast2;
1070 } else {
1071 // Non-scalar replaceable if the allocation type is unknown statically
1072 // (reflection allocation), the object can't be restored during
1073 // deoptimization without precise type.
1074 continue;
1075 }
1076 }
1077 if (alloc->is_Allocate()) {
1078 // Set the scalar_replaceable flag for allocation
1079 // so it could be eliminated.
1080 alloc->as_Allocate()->_is_scalar_replaceable = true;
1081 }
1082 set_escape_state(n->_idx, es);
1083 // in order for an object to be scalar-replaceable, it must be:
1084 // - a direct allocation (not a call returning an object)
1085 // - non-escaping
1086 // - eligible to be a unique type
1087 // - not determined to be ineligible by escape analysis
1088 assert(ptnode_adr(alloc->_idx)->_node != NULL &&
1089 ptnode_adr(n->_idx)->_node != NULL, "should be registered");
1090 set_map(alloc->_idx, n);
1091 set_map(n->_idx, alloc);
1092 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
1093 if (t == NULL)
1094 continue; // not a TypeInstPtr
1095 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
1096 igvn->hash_delete(n);
1097 igvn->set_type(n, tinst);
1098 n->raise_bottom_type(tinst);
1099 igvn->hash_insert(n);
1100 record_for_optimizer(n);
1101 if (alloc->is_Allocate() && ptn->_scalar_replaceable &&
1102 (t->isa_instptr() || t->isa_aryptr())) {
1104 // First, put on the worklist all Field edges from Connection Graph
1105 // which is more accurate then putting immediate users from Ideal Graph.
1106 for (uint e = 0; e < ptn->edge_count(); e++) {
1107 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
1108 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
1109 "only AddP nodes are Field edges in CG");
1110 if (use->outcnt() > 0) { // Don't process dead nodes
1111 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
1112 if (addp2 != NULL) {
1113 assert(alloc->is_AllocateArray(),"array allocation was expected");
1114 alloc_worklist.append_if_missing(addp2);
1115 }
1116 alloc_worklist.append_if_missing(use);
1117 }
1118 }
1120 // An allocation may have an Initialize which has raw stores. Scan
1121 // the users of the raw allocation result and push AddP users
1122 // on alloc_worklist.
1123 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
1124 assert (raw_result != NULL, "must have an allocation result");
1125 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
1126 Node *use = raw_result->fast_out(i);
1127 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
1128 Node* addp2 = find_second_addp(use, raw_result);
1129 if (addp2 != NULL) {
1130 assert(alloc->is_AllocateArray(),"array allocation was expected");
1131 alloc_worklist.append_if_missing(addp2);
1132 }
1133 alloc_worklist.append_if_missing(use);
1134 } else if (use->is_MemBar()) {
1135 memnode_worklist.append_if_missing(use);
1136 }
1137 }
1138 }
1139 } else if (n->is_AddP()) {
1140 ptset.Clear();
1141 PointsTo(ptset, get_addp_base(n));
1142 assert(ptset.Size() == 1, "AddP address is unique");
1143 uint elem = ptset.getelem(); // Allocation node's index
1144 if (elem == _phantom_object) {
1145 assert(false, "escaped allocation");
1146 continue; // Assume the value was set outside this method.
1147 }
1148 Node *base = get_map(elem); // CheckCastPP node
1149 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path
1150 tinst = igvn->type(base)->isa_oopptr();
1151 } else if (n->is_Phi() ||
1152 n->is_CheckCastPP() ||
1153 n->is_EncodeP() ||
1154 n->is_DecodeN() ||
1155 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
1156 if (visited.test_set(n->_idx)) {
1157 assert(n->is_Phi(), "loops only through Phi's");
1158 continue; // already processed
1159 }
1160 ptset.Clear();
1161 PointsTo(ptset, n);
1162 if (ptset.Size() == 1) {
1163 uint elem = ptset.getelem(); // Allocation node's index
1164 if (elem == _phantom_object) {
1165 assert(false, "escaped allocation");
1166 continue; // Assume the value was set outside this method.
1167 }
1168 Node *val = get_map(elem); // CheckCastPP node
1169 TypeNode *tn = n->as_Type();
1170 tinst = igvn->type(val)->isa_oopptr();
1171 assert(tinst != NULL && tinst->is_known_instance() &&
1172 (uint)tinst->instance_id() == elem , "instance type expected.");
1174 const Type *tn_type = igvn->type(tn);
1175 const TypeOopPtr *tn_t;
1176 if (tn_type->isa_narrowoop()) {
1177 tn_t = tn_type->make_ptr()->isa_oopptr();
1178 } else {
1179 tn_t = tn_type->isa_oopptr();
1180 }
1182 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
1183 if (tn_type->isa_narrowoop()) {
1184 tn_type = tinst->make_narrowoop();
1185 } else {
1186 tn_type = tinst;
1187 }
1188 igvn->hash_delete(tn);
1189 igvn->set_type(tn, tn_type);
1190 tn->set_type(tn_type);
1191 igvn->hash_insert(tn);
1192 record_for_optimizer(n);
1193 } else {
1194 assert(tn_type == TypePtr::NULL_PTR ||
1195 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
1196 "unexpected type");
1197 continue; // Skip dead path with different type
1198 }
1199 }
1200 } else {
1201 debug_only(n->dump();)
1202 assert(false, "EA: unexpected node");
1203 continue;
1204 }
1205 // push allocation's users on appropriate worklist
1206 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1207 Node *use = n->fast_out(i);
1208 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1209 // Load/store to instance's field
1210 memnode_worklist.append_if_missing(use);
1211 } else if (use->is_MemBar()) {
1212 memnode_worklist.append_if_missing(use);
1213 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1214 Node* addp2 = find_second_addp(use, n);
1215 if (addp2 != NULL) {
1216 alloc_worklist.append_if_missing(addp2);
1217 }
1218 alloc_worklist.append_if_missing(use);
1219 } else if (use->is_Phi() ||
1220 use->is_CheckCastPP() ||
1221 use->is_EncodeP() ||
1222 use->is_DecodeN() ||
1223 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1224 alloc_worklist.append_if_missing(use);
1225 #ifdef ASSERT
1226 } else if (use->is_Mem()) {
1227 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
1228 } else if (use->is_MergeMem()) {
1229 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1230 } else if (use->is_SafePoint()) {
1231 // Look for MergeMem nodes for calls which reference unique allocation
1232 // (through CheckCastPP nodes) even for debug info.
1233 Node* m = use->in(TypeFunc::Memory);
1234 if (m->is_MergeMem()) {
1235 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1236 }
1237 } else {
1238 uint op = use->Opcode();
1239 if (!(op == Op_CmpP || op == Op_Conv2B ||
1240 op == Op_CastP2X || op == Op_StoreCM ||
1241 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
1242 op == Op_StrEquals || op == Op_StrIndexOf)) {
1243 n->dump();
1244 use->dump();
1245 assert(false, "EA: missing allocation reference path");
1246 }
1247 #endif
1248 }
1249 }
1251 }
1252 // New alias types were created in split_AddP().
1253 uint new_index_end = (uint) _compile->num_alias_types();
1255 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1256 // compute new values for Memory inputs (the Memory inputs are not
1257 // actually updated until phase 4.)
1258 if (memnode_worklist.length() == 0)
1259 return; // nothing to do
1261 while (memnode_worklist.length() != 0) {
1262 Node *n = memnode_worklist.pop();
1263 if (visited.test_set(n->_idx))
1264 continue;
1265 if (n->is_Phi() || n->is_ClearArray()) {
1266 // we don't need to do anything, but the users must be pushed
1267 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
1268 // we don't need to do anything, but the users must be pushed
1269 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
1270 if (n == NULL)
1271 continue;
1272 } else {
1273 assert(n->is_Mem(), "memory node required.");
1274 Node *addr = n->in(MemNode::Address);
1275 const Type *addr_t = igvn->type(addr);
1276 if (addr_t == Type::TOP)
1277 continue;
1278 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1279 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1280 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1281 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1282 if (_compile->failing()) {
1283 return;
1284 }
1285 if (mem != n->in(MemNode::Memory)) {
1286 // We delay the memory edge update since we need old one in
1287 // MergeMem code below when instances memory slices are separated.
1288 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;)
1289 assert(pn == NULL || pn == n, "wrong node");
1290 set_map(n->_idx, mem);
1291 ptnode_adr(n->_idx)->_node = n;
1292 }
1293 if (n->is_Load()) {
1294 continue; // don't push users
1295 } else if (n->is_LoadStore()) {
1296 // get the memory projection
1297 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1298 Node *use = n->fast_out(i);
1299 if (use->Opcode() == Op_SCMemProj) {
1300 n = use;
1301 break;
1302 }
1303 }
1304 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1305 }
1306 }
1307 // push user on appropriate worklist
1308 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1309 Node *use = n->fast_out(i);
1310 if (use->is_Phi() || use->is_ClearArray()) {
1311 memnode_worklist.append_if_missing(use);
1312 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1313 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
1314 continue;
1315 memnode_worklist.append_if_missing(use);
1316 } else if (use->is_MemBar()) {
1317 memnode_worklist.append_if_missing(use);
1318 #ifdef ASSERT
1319 } else if(use->is_Mem()) {
1320 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
1321 } else if (use->is_MergeMem()) {
1322 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1323 } else {
1324 uint op = use->Opcode();
1325 if (!(op == Op_StoreCM ||
1326 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
1327 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
1328 op == Op_AryEq || op == Op_StrComp ||
1329 op == Op_StrEquals || op == Op_StrIndexOf)) {
1330 n->dump();
1331 use->dump();
1332 assert(false, "EA: missing memory path");
1333 }
1334 #endif
1335 }
1336 }
1337 }
1339 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1340 // Walk each memory slice moving the first node encountered of each
1341 // instance type to the the input corresponding to its alias index.
1342 uint length = _mergemem_worklist.length();
1343 for( uint next = 0; next < length; ++next ) {
1344 MergeMemNode* nmm = _mergemem_worklist.at(next);
1345 assert(!visited.test_set(nmm->_idx), "should not be visited before");
1346 // Note: we don't want to use MergeMemStream here because we only want to
1347 // scan inputs which exist at the start, not ones we add during processing.
1348 // Note 2: MergeMem may already contains instance memory slices added
1349 // during find_inst_mem() call when memory nodes were processed above.
1350 igvn->hash_delete(nmm);
1351 uint nslices = nmm->req();
1352 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1353 Node* mem = nmm->in(i);
1354 Node* cur = NULL;
1355 if (mem == NULL || mem->is_top())
1356 continue;
1357 // First, update mergemem by moving memory nodes to corresponding slices
1358 // if their type became more precise since this mergemem was created.
1359 while (mem->is_Mem()) {
1360 const Type *at = igvn->type(mem->in(MemNode::Address));
1361 if (at != Type::TOP) {
1362 assert (at->isa_ptr() != NULL, "pointer type required.");
1363 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1364 if (idx == i) {
1365 if (cur == NULL)
1366 cur = mem;
1367 } else {
1368 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1369 nmm->set_memory_at(idx, mem);
1370 }
1371 }
1372 }
1373 mem = mem->in(MemNode::Memory);
1374 }
1375 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1376 // Find any instance of the current type if we haven't encountered
1377 // already a memory slice of the instance along the memory chain.
1378 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1379 if((uint)_compile->get_general_index(ni) == i) {
1380 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1381 if (nmm->is_empty_memory(m)) {
1382 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1383 if (_compile->failing()) {
1384 return;
1385 }
1386 nmm->set_memory_at(ni, result);
1387 }
1388 }
1389 }
1390 }
1391 // Find the rest of instances values
1392 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1393 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
1394 Node* result = step_through_mergemem(nmm, ni, tinst);
1395 if (result == nmm->base_memory()) {
1396 // Didn't find instance memory, search through general slice recursively.
1397 result = nmm->memory_at(_compile->get_general_index(ni));
1398 result = find_inst_mem(result, ni, orig_phis, igvn);
1399 if (_compile->failing()) {
1400 return;
1401 }
1402 nmm->set_memory_at(ni, result);
1403 }
1404 }
1405 igvn->hash_insert(nmm);
1406 record_for_optimizer(nmm);
1407 }
1409 // Phase 4: Update the inputs of non-instance memory Phis and
1410 // the Memory input of memnodes
1411 // First update the inputs of any non-instance Phi's from
1412 // which we split out an instance Phi. Note we don't have
1413 // to recursively process Phi's encounted on the input memory
1414 // chains as is done in split_memory_phi() since they will
1415 // also be processed here.
1416 for (int j = 0; j < orig_phis.length(); j++) {
1417 PhiNode *phi = orig_phis.at(j);
1418 int alias_idx = _compile->get_alias_index(phi->adr_type());
1419 igvn->hash_delete(phi);
1420 for (uint i = 1; i < phi->req(); i++) {
1421 Node *mem = phi->in(i);
1422 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1423 if (_compile->failing()) {
1424 return;
1425 }
1426 if (mem != new_mem) {
1427 phi->set_req(i, new_mem);
1428 }
1429 }
1430 igvn->hash_insert(phi);
1431 record_for_optimizer(phi);
1432 }
1434 // Update the memory inputs of MemNodes with the value we computed
1435 // in Phase 2 and move stores memory users to corresponding memory slices.
1436 #ifdef ASSERT
1437 visited.Clear();
1438 Node_Stack old_mems(arena, _compile->unique() >> 2);
1439 #endif
1440 for (uint i = 0; i < nodes_size(); i++) {
1441 Node *nmem = get_map(i);
1442 if (nmem != NULL) {
1443 Node *n = ptnode_adr(i)->_node;
1444 assert(n != NULL, "sanity");
1445 if (n->is_Mem()) {
1446 #ifdef ASSERT
1447 Node* old_mem = n->in(MemNode::Memory);
1448 if (!visited.test_set(old_mem->_idx)) {
1449 old_mems.push(old_mem, old_mem->outcnt());
1450 }
1451 #endif
1452 assert(n->in(MemNode::Memory) != nmem, "sanity");
1453 if (!n->is_Load()) {
1454 // Move memory users of a store first.
1455 move_inst_mem(n, orig_phis, igvn);
1456 }
1457 // Now update memory input
1458 igvn->hash_delete(n);
1459 n->set_req(MemNode::Memory, nmem);
1460 igvn->hash_insert(n);
1461 record_for_optimizer(n);
1462 } else {
1463 assert(n->is_Allocate() || n->is_CheckCastPP() ||
1464 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
1465 }
1466 }
1467 }
1468 #ifdef ASSERT
1469 // Verify that memory was split correctly
1470 while (old_mems.is_nonempty()) {
1471 Node* old_mem = old_mems.node();
1472 uint old_cnt = old_mems.index();
1473 old_mems.pop();
1474 assert(old_cnt = old_mem->outcnt(), "old mem could be lost");
1475 }
1476 #endif
1477 }
1479 bool ConnectionGraph::has_candidates(Compile *C) {
1480 // EA brings benefits only when the code has allocations and/or locks which
1481 // are represented by ideal Macro nodes.
1482 int cnt = C->macro_count();
1483 for( int i=0; i < cnt; i++ ) {
1484 Node *n = C->macro_node(i);
1485 if ( n->is_Allocate() )
1486 return true;
1487 if( n->is_Lock() ) {
1488 Node* obj = n->as_Lock()->obj_node()->uncast();
1489 if( !(obj->is_Parm() || obj->is_Con()) )
1490 return true;
1491 }
1492 }
1493 return false;
1494 }
1496 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
1497 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
1498 // to create space for them in ConnectionGraph::_nodes[].
1499 Node* oop_null = igvn->zerocon(T_OBJECT);
1500 Node* noop_null = igvn->zerocon(T_NARROWOOP);
1502 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
1503 // Perform escape analysis
1504 if (congraph->compute_escape()) {
1505 // There are non escaping objects.
1506 C->set_congraph(congraph);
1507 }
1509 // Cleanup.
1510 if (oop_null->outcnt() == 0)
1511 igvn->hash_delete(oop_null);
1512 if (noop_null->outcnt() == 0)
1513 igvn->hash_delete(noop_null);
1514 }
1516 bool ConnectionGraph::compute_escape() {
1517 Compile* C = _compile;
1519 // 1. Populate Connection Graph (CG) with Ideal nodes.
1521 Unique_Node_List worklist_init;
1522 worklist_init.map(C->unique(), NULL); // preallocate space
1524 // Initialize worklist
1525 if (C->root() != NULL) {
1526 worklist_init.push(C->root());
1527 }
1529 GrowableArray<int> cg_worklist;
1530 PhaseGVN* igvn = _igvn;
1531 bool has_allocations = false;
1533 // Push all useful nodes onto CG list and set their type.
1534 for( uint next = 0; next < worklist_init.size(); ++next ) {
1535 Node* n = worklist_init.at(next);
1536 record_for_escape_analysis(n, igvn);
1537 // Only allocations and java static calls results are checked
1538 // for an escape status. See process_call_result() below.
1539 if (n->is_Allocate() || n->is_CallStaticJava() &&
1540 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1541 has_allocations = true;
1542 }
1543 if(n->is_AddP()) {
1544 // Collect address nodes. Use them during stage 3 below
1545 // to build initial connection graph field edges.
1546 cg_worklist.append(n->_idx);
1547 } else if (n->is_MergeMem()) {
1548 // Collect all MergeMem nodes to add memory slices for
1549 // scalar replaceable objects in split_unique_types().
1550 _mergemem_worklist.append(n->as_MergeMem());
1551 }
1552 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1553 Node* m = n->fast_out(i); // Get user
1554 worklist_init.push(m);
1555 }
1556 }
1558 if (!has_allocations) {
1559 _collecting = false;
1560 return false; // Nothing to do.
1561 }
1563 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1564 uint delayed_size = _delayed_worklist.size();
1565 for( uint next = 0; next < delayed_size; ++next ) {
1566 Node* n = _delayed_worklist.at(next);
1567 build_connection_graph(n, igvn);
1568 }
1570 // 3. Pass to create initial fields edges (JavaObject -F-> AddP)
1571 // to reduce number of iterations during stage 4 below.
1572 uint cg_length = cg_worklist.length();
1573 for( uint next = 0; next < cg_length; ++next ) {
1574 int ni = cg_worklist.at(next);
1575 Node* n = ptnode_adr(ni)->_node;
1576 Node* base = get_addp_base(n);
1577 if (base->is_Proj())
1578 base = base->in(0);
1579 PointsToNode::NodeType nt = ptnode_adr(base->_idx)->node_type();
1580 if (nt == PointsToNode::JavaObject) {
1581 build_connection_graph(n, igvn);
1582 }
1583 }
1585 cg_worklist.clear();
1586 cg_worklist.append(_phantom_object);
1587 GrowableArray<uint> worklist;
1589 // 4. Build Connection Graph which need
1590 // to walk the connection graph.
1591 _progress = false;
1592 for (uint ni = 0; ni < nodes_size(); ni++) {
1593 PointsToNode* ptn = ptnode_adr(ni);
1594 Node *n = ptn->_node;
1595 if (n != NULL) { // Call, AddP, LoadP, StoreP
1596 build_connection_graph(n, igvn);
1597 if (ptn->node_type() != PointsToNode::UnknownType)
1598 cg_worklist.append(n->_idx); // Collect CG nodes
1599 if (!_processed.test(n->_idx))
1600 worklist.append(n->_idx); // Collect C/A/L/S nodes
1601 }
1602 }
1604 // After IGVN user nodes may have smaller _idx than
1605 // their inputs so they will be processed first in
1606 // previous loop. Because of that not all Graph
1607 // edges will be created. Walk over interesting
1608 // nodes again until no new edges are created.
1609 //
1610 // Normally only 1-3 passes needed to build
1611 // Connection Graph depending on graph complexity.
1612 // Set limit to 10 to catch situation when something
1613 // did go wrong and recompile the method without EA.
1615 #define CG_BUILD_ITER_LIMIT 10
1617 uint length = worklist.length();
1618 int iterations = 0;
1619 while(_progress && (iterations++ < CG_BUILD_ITER_LIMIT)) {
1620 _progress = false;
1621 for( uint next = 0; next < length; ++next ) {
1622 int ni = worklist.at(next);
1623 PointsToNode* ptn = ptnode_adr(ni);
1624 Node* n = ptn->_node;
1625 assert(n != NULL, "should be known node");
1626 build_connection_graph(n, igvn);
1627 }
1628 }
1629 if (iterations >= CG_BUILD_ITER_LIMIT) {
1630 assert(iterations < CG_BUILD_ITER_LIMIT,
1631 err_msg("infinite EA connection graph build with %d nodes and worklist size %d",
1632 nodes_size(), length));
1633 // Possible infinite build_connection_graph loop,
1634 // retry compilation without escape analysis.
1635 C->record_failure(C2Compiler::retry_no_escape_analysis());
1636 _collecting = false;
1637 return false;
1638 }
1639 #undef CG_BUILD_ITER_LIMIT
1641 Arena* arena = Thread::current()->resource_area();
1642 VectorSet ptset(arena);
1643 VectorSet visited(arena);
1644 worklist.clear();
1646 // 5. Remove deferred edges from the graph and adjust
1647 // escape state of nonescaping objects.
1648 cg_length = cg_worklist.length();
1649 for( uint next = 0; next < cg_length; ++next ) {
1650 int ni = cg_worklist.at(next);
1651 PointsToNode* ptn = ptnode_adr(ni);
1652 PointsToNode::NodeType nt = ptn->node_type();
1653 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1654 remove_deferred(ni, &worklist, &visited);
1655 Node *n = ptn->_node;
1656 if (n->is_AddP()) {
1657 // Search for objects which are not scalar replaceable
1658 // and adjust their escape state.
1659 verify_escape_state(ni, ptset, igvn);
1660 }
1661 }
1662 }
1664 // 6. Propagate escape states.
1665 worklist.clear();
1666 bool has_non_escaping_obj = false;
1668 // push all GlobalEscape nodes on the worklist
1669 for( uint next = 0; next < cg_length; ++next ) {
1670 int nk = cg_worklist.at(next);
1671 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape)
1672 worklist.push(nk);
1673 }
1674 // mark all nodes reachable from GlobalEscape nodes
1675 while(worklist.length() > 0) {
1676 PointsToNode* ptn = ptnode_adr(worklist.pop());
1677 uint e_cnt = ptn->edge_count();
1678 for (uint ei = 0; ei < e_cnt; ei++) {
1679 uint npi = ptn->edge_target(ei);
1680 PointsToNode *np = ptnode_adr(npi);
1681 if (np->escape_state() < PointsToNode::GlobalEscape) {
1682 np->set_escape_state(PointsToNode::GlobalEscape);
1683 worklist.push(npi);
1684 }
1685 }
1686 }
1688 // push all ArgEscape nodes on the worklist
1689 for( uint next = 0; next < cg_length; ++next ) {
1690 int nk = cg_worklist.at(next);
1691 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape)
1692 worklist.push(nk);
1693 }
1694 // mark all nodes reachable from ArgEscape nodes
1695 while(worklist.length() > 0) {
1696 PointsToNode* ptn = ptnode_adr(worklist.pop());
1697 if (ptn->node_type() == PointsToNode::JavaObject)
1698 has_non_escaping_obj = true; // Non GlobalEscape
1699 uint e_cnt = ptn->edge_count();
1700 for (uint ei = 0; ei < e_cnt; ei++) {
1701 uint npi = ptn->edge_target(ei);
1702 PointsToNode *np = ptnode_adr(npi);
1703 if (np->escape_state() < PointsToNode::ArgEscape) {
1704 np->set_escape_state(PointsToNode::ArgEscape);
1705 worklist.push(npi);
1706 }
1707 }
1708 }
1710 GrowableArray<Node*> alloc_worklist;
1712 // push all NoEscape nodes on the worklist
1713 for( uint next = 0; next < cg_length; ++next ) {
1714 int nk = cg_worklist.at(next);
1715 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape)
1716 worklist.push(nk);
1717 }
1718 // mark all nodes reachable from NoEscape nodes
1719 while(worklist.length() > 0) {
1720 PointsToNode* ptn = ptnode_adr(worklist.pop());
1721 if (ptn->node_type() == PointsToNode::JavaObject)
1722 has_non_escaping_obj = true; // Non GlobalEscape
1723 Node* n = ptn->_node;
1724 if (n->is_Allocate() && ptn->_scalar_replaceable ) {
1725 // Push scalar replaceable allocations on alloc_worklist
1726 // for processing in split_unique_types().
1727 alloc_worklist.append(n);
1728 }
1729 uint e_cnt = ptn->edge_count();
1730 for (uint ei = 0; ei < e_cnt; ei++) {
1731 uint npi = ptn->edge_target(ei);
1732 PointsToNode *np = ptnode_adr(npi);
1733 if (np->escape_state() < PointsToNode::NoEscape) {
1734 np->set_escape_state(PointsToNode::NoEscape);
1735 worklist.push(npi);
1736 }
1737 }
1738 }
1740 _collecting = false;
1741 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1743 #ifndef PRODUCT
1744 if (PrintEscapeAnalysis) {
1745 dump(); // Dump ConnectionGraph
1746 }
1747 #endif
1749 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0;
1750 if ( has_scalar_replaceable_candidates &&
1751 C->AliasLevel() >= 3 && EliminateAllocations ) {
1753 // Now use the escape information to create unique types for
1754 // scalar replaceable objects.
1755 split_unique_types(alloc_worklist);
1757 if (C->failing()) return false;
1759 C->print_method("After Escape Analysis", 2);
1761 #ifdef ASSERT
1762 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1763 tty->print("=== No allocations eliminated for ");
1764 C->method()->print_short_name();
1765 if(!EliminateAllocations) {
1766 tty->print(" since EliminateAllocations is off ===");
1767 } else if(!has_scalar_replaceable_candidates) {
1768 tty->print(" since there are no scalar replaceable candidates ===");
1769 } else if(C->AliasLevel() < 3) {
1770 tty->print(" since AliasLevel < 3 ===");
1771 }
1772 tty->cr();
1773 #endif
1774 }
1775 return has_non_escaping_obj;
1776 }
1778 // Search for objects which are not scalar replaceable.
1779 void ConnectionGraph::verify_escape_state(int nidx, VectorSet& ptset, PhaseTransform* phase) {
1780 PointsToNode* ptn = ptnode_adr(nidx);
1781 Node* n = ptn->_node;
1782 assert(n->is_AddP(), "Should be called for AddP nodes only");
1783 // Search for objects which are not scalar replaceable.
1784 // Mark their escape state as ArgEscape to propagate the state
1785 // to referenced objects.
1786 // Note: currently there are no difference in compiler optimizations
1787 // for ArgEscape objects and NoEscape objects which are not
1788 // scalar replaceable.
1790 Compile* C = _compile;
1792 int offset = ptn->offset();
1793 Node* base = get_addp_base(n);
1794 ptset.Clear();
1795 PointsTo(ptset, base);
1796 int ptset_size = ptset.Size();
1798 // Check if a oop field's initializing value is recorded and add
1799 // a corresponding NULL field's value if it is not recorded.
1800 // Connection Graph does not record a default initialization by NULL
1801 // captured by Initialize node.
1802 //
1803 // Note: it will disable scalar replacement in some cases:
1804 //
1805 // Point p[] = new Point[1];
1806 // p[0] = new Point(); // Will be not scalar replaced
1807 //
1808 // but it will save us from incorrect optimizations in next cases:
1809 //
1810 // Point p[] = new Point[1];
1811 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1812 //
1813 // Do a simple control flow analysis to distinguish above cases.
1814 //
1815 if (offset != Type::OffsetBot && ptset_size == 1) {
1816 uint elem = ptset.getelem(); // Allocation node's index
1817 // It does not matter if it is not Allocation node since
1818 // only non-escaping allocations are scalar replaced.
1819 if (ptnode_adr(elem)->_node->is_Allocate() &&
1820 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) {
1821 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate();
1822 InitializeNode* ini = alloc->initialization();
1824 // Check only oop fields.
1825 const Type* adr_type = n->as_AddP()->bottom_type();
1826 BasicType basic_field_type = T_INT;
1827 if (adr_type->isa_instptr()) {
1828 ciField* field = C->alias_type(adr_type->isa_instptr())->field();
1829 if (field != NULL) {
1830 basic_field_type = field->layout_type();
1831 } else {
1832 // Ignore non field load (for example, klass load)
1833 }
1834 } else if (adr_type->isa_aryptr()) {
1835 const Type* elemtype = adr_type->isa_aryptr()->elem();
1836 basic_field_type = elemtype->array_element_basic_type();
1837 } else {
1838 // Raw pointers are used for initializing stores so skip it.
1839 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1840 (base->in(0) == alloc),"unexpected pointer type");
1841 }
1842 if (basic_field_type == T_OBJECT ||
1843 basic_field_type == T_NARROWOOP ||
1844 basic_field_type == T_ARRAY) {
1845 Node* value = NULL;
1846 if (ini != NULL) {
1847 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
1848 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase);
1849 if (store != NULL && store->is_Store()) {
1850 value = store->in(MemNode::ValueIn);
1851 } else if (ptn->edge_count() > 0) { // Are there oop stores?
1852 // Check for a store which follows allocation without branches.
1853 // For example, a volatile field store is not collected
1854 // by Initialize node. TODO: it would be nice to use idom() here.
1855 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1856 store = n->fast_out(i);
1857 if (store->is_Store() && store->in(0) != NULL) {
1858 Node* ctrl = store->in(0);
1859 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL ||
1860 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() ||
1861 ctrl->is_IfTrue() || ctrl->is_IfFalse())) {
1862 ctrl = ctrl->in(0);
1863 }
1864 if (ctrl == ini || ctrl == alloc) {
1865 value = store->in(MemNode::ValueIn);
1866 break;
1867 }
1868 }
1869 }
1870 }
1871 }
1872 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
1873 // A field's initializing value was not recorded. Add NULL.
1874 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1875 add_pointsto_edge(nidx, null_idx);
1876 }
1877 }
1878 }
1879 }
1881 // An object is not scalar replaceable if the field which may point
1882 // to it has unknown offset (unknown element of an array of objects).
1883 //
1884 if (offset == Type::OffsetBot) {
1885 uint e_cnt = ptn->edge_count();
1886 for (uint ei = 0; ei < e_cnt; ei++) {
1887 uint npi = ptn->edge_target(ei);
1888 set_escape_state(npi, PointsToNode::ArgEscape);
1889 ptnode_adr(npi)->_scalar_replaceable = false;
1890 }
1891 }
1893 // Currently an object is not scalar replaceable if a LoadStore node
1894 // access its field since the field value is unknown after it.
1895 //
1896 bool has_LoadStore = false;
1897 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1898 Node *use = n->fast_out(i);
1899 if (use->is_LoadStore()) {
1900 has_LoadStore = true;
1901 break;
1902 }
1903 }
1904 // An object is not scalar replaceable if the address points
1905 // to unknown field (unknown element for arrays, offset is OffsetBot).
1906 //
1907 // Or the address may point to more then one object. This may produce
1908 // the false positive result (set scalar_replaceable to false)
1909 // since the flow-insensitive escape analysis can't separate
1910 // the case when stores overwrite the field's value from the case
1911 // when stores happened on different control branches.
1912 //
1913 if (ptset_size > 1 || ptset_size != 0 &&
1914 (has_LoadStore || offset == Type::OffsetBot)) {
1915 for( VectorSetI j(&ptset); j.test(); ++j ) {
1916 set_escape_state(j.elem, PointsToNode::ArgEscape);
1917 ptnode_adr(j.elem)->_scalar_replaceable = false;
1918 }
1919 }
1920 }
1922 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
1924 switch (call->Opcode()) {
1925 #ifdef ASSERT
1926 case Op_Allocate:
1927 case Op_AllocateArray:
1928 case Op_Lock:
1929 case Op_Unlock:
1930 assert(false, "should be done already");
1931 break;
1932 #endif
1933 case Op_CallLeaf:
1934 case Op_CallLeafNoFP:
1935 {
1936 // Stub calls, objects do not escape but they are not scale replaceable.
1937 // Adjust escape state for outgoing arguments.
1938 const TypeTuple * d = call->tf()->domain();
1939 VectorSet ptset(Thread::current()->resource_area());
1940 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1941 const Type* at = d->field_at(i);
1942 Node *arg = call->in(i)->uncast();
1943 const Type *aat = phase->type(arg);
1944 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() &&
1945 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) {
1947 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
1948 aat->isa_ptr() != NULL, "expecting an Ptr");
1949 #ifdef ASSERT
1950 if (!(call->Opcode() == Op_CallLeafNoFP &&
1951 call->as_CallLeaf()->_name != NULL &&
1952 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) ||
1953 call->as_CallLeaf()->_name != NULL &&
1954 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
1955 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ))
1956 ) {
1957 call->dump();
1958 assert(false, "EA: unexpected CallLeaf");
1959 }
1960 #endif
1961 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1962 if (arg->is_AddP()) {
1963 //
1964 // The inline_native_clone() case when the arraycopy stub is called
1965 // after the allocation before Initialize and CheckCastPP nodes.
1966 //
1967 // Set AddP's base (Allocate) as not scalar replaceable since
1968 // pointer to the base (with offset) is passed as argument.
1969 //
1970 arg = get_addp_base(arg);
1971 }
1972 ptset.Clear();
1973 PointsTo(ptset, arg);
1974 for( VectorSetI j(&ptset); j.test(); ++j ) {
1975 uint pt = j.elem;
1976 set_escape_state(pt, PointsToNode::ArgEscape);
1977 }
1978 }
1979 }
1980 break;
1981 }
1983 case Op_CallStaticJava:
1984 // For a static call, we know exactly what method is being called.
1985 // Use bytecode estimator to record the call's escape affects
1986 {
1987 ciMethod *meth = call->as_CallJava()->method();
1988 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1989 // fall-through if not a Java method or no analyzer information
1990 if (call_analyzer != NULL) {
1991 const TypeTuple * d = call->tf()->domain();
1992 VectorSet ptset(Thread::current()->resource_area());
1993 bool copy_dependencies = false;
1994 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1995 const Type* at = d->field_at(i);
1996 int k = i - TypeFunc::Parms;
1997 Node *arg = call->in(i)->uncast();
1999 if (at->isa_oopptr() != NULL &&
2000 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) {
2002 bool global_escapes = false;
2003 bool fields_escapes = false;
2004 if (!call_analyzer->is_arg_stack(k)) {
2005 // The argument global escapes, mark everything it could point to
2006 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2007 global_escapes = true;
2008 } else {
2009 if (!call_analyzer->is_arg_local(k)) {
2010 // The argument itself doesn't escape, but any fields might
2011 fields_escapes = true;
2012 }
2013 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
2014 copy_dependencies = true;
2015 }
2017 ptset.Clear();
2018 PointsTo(ptset, arg);
2019 for( VectorSetI j(&ptset); j.test(); ++j ) {
2020 uint pt = j.elem;
2021 if (global_escapes) {
2022 //The argument global escapes, mark everything it could point to
2023 set_escape_state(pt, PointsToNode::GlobalEscape);
2024 } else {
2025 if (fields_escapes) {
2026 // The argument itself doesn't escape, but any fields might
2027 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
2028 }
2029 set_escape_state(pt, PointsToNode::ArgEscape);
2030 }
2031 }
2032 }
2033 }
2034 if (copy_dependencies)
2035 call_analyzer->copy_dependencies(_compile->dependencies());
2036 break;
2037 }
2038 }
2040 default:
2041 // Fall-through here if not a Java method or no analyzer information
2042 // or some other type of call, assume the worst case: all arguments
2043 // globally escape.
2044 {
2045 // adjust escape state for outgoing arguments
2046 const TypeTuple * d = call->tf()->domain();
2047 VectorSet ptset(Thread::current()->resource_area());
2048 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2049 const Type* at = d->field_at(i);
2050 if (at->isa_oopptr() != NULL) {
2051 Node *arg = call->in(i)->uncast();
2052 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2053 ptset.Clear();
2054 PointsTo(ptset, arg);
2055 for( VectorSetI j(&ptset); j.test(); ++j ) {
2056 uint pt = j.elem;
2057 set_escape_state(pt, PointsToNode::GlobalEscape);
2058 }
2059 }
2060 }
2061 }
2062 }
2063 }
2064 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
2065 CallNode *call = resproj->in(0)->as_Call();
2066 uint call_idx = call->_idx;
2067 uint resproj_idx = resproj->_idx;
2069 switch (call->Opcode()) {
2070 case Op_Allocate:
2071 {
2072 Node *k = call->in(AllocateNode::KlassNode);
2073 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2074 assert(kt != NULL, "TypeKlassPtr required.");
2075 ciKlass* cik = kt->klass();
2077 PointsToNode::EscapeState es;
2078 uint edge_to;
2079 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
2080 !cik->is_instance_klass() || // StressReflectiveCode
2081 cik->as_instance_klass()->has_finalizer()) {
2082 es = PointsToNode::GlobalEscape;
2083 edge_to = _phantom_object; // Could not be worse
2084 } else {
2085 es = PointsToNode::NoEscape;
2086 edge_to = call_idx;
2087 }
2088 set_escape_state(call_idx, es);
2089 add_pointsto_edge(resproj_idx, edge_to);
2090 _processed.set(resproj_idx);
2091 break;
2092 }
2094 case Op_AllocateArray:
2095 {
2097 Node *k = call->in(AllocateNode::KlassNode);
2098 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2099 assert(kt != NULL, "TypeKlassPtr required.");
2100 ciKlass* cik = kt->klass();
2102 PointsToNode::EscapeState es;
2103 uint edge_to;
2104 if (!cik->is_array_klass()) { // StressReflectiveCode
2105 es = PointsToNode::GlobalEscape;
2106 edge_to = _phantom_object;
2107 } else {
2108 es = PointsToNode::NoEscape;
2109 edge_to = call_idx;
2110 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2111 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
2112 // Not scalar replaceable if the length is not constant or too big.
2113 ptnode_adr(call_idx)->_scalar_replaceable = false;
2114 }
2115 }
2116 set_escape_state(call_idx, es);
2117 add_pointsto_edge(resproj_idx, edge_to);
2118 _processed.set(resproj_idx);
2119 break;
2120 }
2122 case Op_CallStaticJava:
2123 // For a static call, we know exactly what method is being called.
2124 // Use bytecode estimator to record whether the call's return value escapes
2125 {
2126 bool done = true;
2127 const TypeTuple *r = call->tf()->range();
2128 const Type* ret_type = NULL;
2130 if (r->cnt() > TypeFunc::Parms)
2131 ret_type = r->field_at(TypeFunc::Parms);
2133 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2134 // _multianewarray functions return a TypeRawPtr.
2135 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
2136 _processed.set(resproj_idx);
2137 break; // doesn't return a pointer type
2138 }
2139 ciMethod *meth = call->as_CallJava()->method();
2140 const TypeTuple * d = call->tf()->domain();
2141 if (meth == NULL) {
2142 // not a Java method, assume global escape
2143 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2144 add_pointsto_edge(resproj_idx, _phantom_object);
2145 } else {
2146 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
2147 bool copy_dependencies = false;
2149 if (call_analyzer->is_return_allocated()) {
2150 // Returns a newly allocated unescaped object, simply
2151 // update dependency information.
2152 // Mark it as NoEscape so that objects referenced by
2153 // it's fields will be marked as NoEscape at least.
2154 set_escape_state(call_idx, PointsToNode::NoEscape);
2155 add_pointsto_edge(resproj_idx, call_idx);
2156 copy_dependencies = true;
2157 } else if (call_analyzer->is_return_local()) {
2158 // determine whether any arguments are returned
2159 set_escape_state(call_idx, PointsToNode::NoEscape);
2160 bool ret_arg = false;
2161 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2162 const Type* at = d->field_at(i);
2164 if (at->isa_oopptr() != NULL) {
2165 Node *arg = call->in(i)->uncast();
2167 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2168 ret_arg = true;
2169 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2170 if (arg_esp->node_type() == PointsToNode::UnknownType)
2171 done = false;
2172 else if (arg_esp->node_type() == PointsToNode::JavaObject)
2173 add_pointsto_edge(resproj_idx, arg->_idx);
2174 else
2175 add_deferred_edge(resproj_idx, arg->_idx);
2176 arg_esp->_hidden_alias = true;
2177 }
2178 }
2179 }
2180 if (done && !ret_arg) {
2181 // Returns unknown object.
2182 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2183 add_pointsto_edge(resproj_idx, _phantom_object);
2184 }
2185 copy_dependencies = true;
2186 } else {
2187 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2188 add_pointsto_edge(resproj_idx, _phantom_object);
2189 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2190 const Type* at = d->field_at(i);
2191 if (at->isa_oopptr() != NULL) {
2192 Node *arg = call->in(i)->uncast();
2193 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2194 arg_esp->_hidden_alias = true;
2195 }
2196 }
2197 }
2198 if (copy_dependencies)
2199 call_analyzer->copy_dependencies(_compile->dependencies());
2200 }
2201 if (done)
2202 _processed.set(resproj_idx);
2203 break;
2204 }
2206 default:
2207 // Some other type of call, assume the worst case that the
2208 // returned value, if any, globally escapes.
2209 {
2210 const TypeTuple *r = call->tf()->range();
2211 if (r->cnt() > TypeFunc::Parms) {
2212 const Type* ret_type = r->field_at(TypeFunc::Parms);
2214 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2215 // _multianewarray functions return a TypeRawPtr.
2216 if (ret_type->isa_ptr() != NULL) {
2217 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2218 add_pointsto_edge(resproj_idx, _phantom_object);
2219 }
2220 }
2221 _processed.set(resproj_idx);
2222 }
2223 }
2224 }
2226 // Populate Connection Graph with Ideal nodes and create simple
2227 // connection graph edges (do not need to check the node_type of inputs
2228 // or to call PointsTo() to walk the connection graph).
2229 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
2230 if (_processed.test(n->_idx))
2231 return; // No need to redefine node's state.
2233 if (n->is_Call()) {
2234 // Arguments to allocation and locking don't escape.
2235 if (n->is_Allocate()) {
2236 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
2237 record_for_optimizer(n);
2238 } else if (n->is_Lock() || n->is_Unlock()) {
2239 // Put Lock and Unlock nodes on IGVN worklist to process them during
2240 // the first IGVN optimization when escape information is still available.
2241 record_for_optimizer(n);
2242 _processed.set(n->_idx);
2243 } else {
2244 // Don't mark as processed since call's arguments have to be processed.
2245 PointsToNode::NodeType nt = PointsToNode::UnknownType;
2246 PointsToNode::EscapeState es = PointsToNode::UnknownEscape;
2248 // Check if a call returns an object.
2249 const TypeTuple *r = n->as_Call()->tf()->range();
2250 if (r->cnt() > TypeFunc::Parms &&
2251 r->field_at(TypeFunc::Parms)->isa_ptr() &&
2252 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
2253 nt = PointsToNode::JavaObject;
2254 if (!n->is_CallStaticJava()) {
2255 // Since the called mathod is statically unknown assume
2256 // the worst case that the returned value globally escapes.
2257 es = PointsToNode::GlobalEscape;
2258 }
2259 }
2260 add_node(n, nt, es, false);
2261 }
2262 return;
2263 }
2265 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
2266 // ThreadLocal has RawPrt type.
2267 switch (n->Opcode()) {
2268 case Op_AddP:
2269 {
2270 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
2271 break;
2272 }
2273 case Op_CastX2P:
2274 { // "Unsafe" memory access.
2275 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2276 break;
2277 }
2278 case Op_CastPP:
2279 case Op_CheckCastPP:
2280 case Op_EncodeP:
2281 case Op_DecodeN:
2282 {
2283 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2284 int ti = n->in(1)->_idx;
2285 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2286 if (nt == PointsToNode::UnknownType) {
2287 _delayed_worklist.push(n); // Process it later.
2288 break;
2289 } else if (nt == PointsToNode::JavaObject) {
2290 add_pointsto_edge(n->_idx, ti);
2291 } else {
2292 add_deferred_edge(n->_idx, ti);
2293 }
2294 _processed.set(n->_idx);
2295 break;
2296 }
2297 case Op_ConP:
2298 {
2299 // assume all pointer constants globally escape except for null
2300 PointsToNode::EscapeState es;
2301 if (phase->type(n) == TypePtr::NULL_PTR)
2302 es = PointsToNode::NoEscape;
2303 else
2304 es = PointsToNode::GlobalEscape;
2306 add_node(n, PointsToNode::JavaObject, es, true);
2307 break;
2308 }
2309 case Op_ConN:
2310 {
2311 // assume all narrow oop constants globally escape except for null
2312 PointsToNode::EscapeState es;
2313 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
2314 es = PointsToNode::NoEscape;
2315 else
2316 es = PointsToNode::GlobalEscape;
2318 add_node(n, PointsToNode::JavaObject, es, true);
2319 break;
2320 }
2321 case Op_CreateEx:
2322 {
2323 // assume that all exception objects globally escape
2324 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2325 break;
2326 }
2327 case Op_LoadKlass:
2328 case Op_LoadNKlass:
2329 {
2330 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2331 break;
2332 }
2333 case Op_LoadP:
2334 case Op_LoadN:
2335 {
2336 const Type *t = phase->type(n);
2337 if (t->make_ptr() == NULL) {
2338 _processed.set(n->_idx);
2339 return;
2340 }
2341 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2342 break;
2343 }
2344 case Op_Parm:
2345 {
2346 _processed.set(n->_idx); // No need to redefine it state.
2347 uint con = n->as_Proj()->_con;
2348 if (con < TypeFunc::Parms)
2349 return;
2350 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2351 if (t->isa_ptr() == NULL)
2352 return;
2353 // We have to assume all input parameters globally escape
2354 // (Note: passing 'false' since _processed is already set).
2355 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2356 break;
2357 }
2358 case Op_Phi:
2359 {
2360 const Type *t = n->as_Phi()->type();
2361 if (t->make_ptr() == NULL) {
2362 // nothing to do if not an oop or narrow oop
2363 _processed.set(n->_idx);
2364 return;
2365 }
2366 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2367 uint i;
2368 for (i = 1; i < n->req() ; i++) {
2369 Node* in = n->in(i);
2370 if (in == NULL)
2371 continue; // ignore NULL
2372 in = in->uncast();
2373 if (in->is_top() || in == n)
2374 continue; // ignore top or inputs which go back this node
2375 int ti = in->_idx;
2376 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2377 if (nt == PointsToNode::UnknownType) {
2378 break;
2379 } else if (nt == PointsToNode::JavaObject) {
2380 add_pointsto_edge(n->_idx, ti);
2381 } else {
2382 add_deferred_edge(n->_idx, ti);
2383 }
2384 }
2385 if (i >= n->req())
2386 _processed.set(n->_idx);
2387 else
2388 _delayed_worklist.push(n);
2389 break;
2390 }
2391 case Op_Proj:
2392 {
2393 // we are only interested in the oop result projection from a call
2394 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2395 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2396 assert(r->cnt() > TypeFunc::Parms, "sanity");
2397 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2398 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2399 int ti = n->in(0)->_idx;
2400 // The call may not be registered yet (since not all its inputs are registered)
2401 // if this is the projection from backbranch edge of Phi.
2402 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) {
2403 process_call_result(n->as_Proj(), phase);
2404 }
2405 if (!_processed.test(n->_idx)) {
2406 // The call's result may need to be processed later if the call
2407 // returns it's argument and the argument is not processed yet.
2408 _delayed_worklist.push(n);
2409 }
2410 break;
2411 }
2412 }
2413 _processed.set(n->_idx);
2414 break;
2415 }
2416 case Op_Return:
2417 {
2418 if( n->req() > TypeFunc::Parms &&
2419 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2420 // Treat Return value as LocalVar with GlobalEscape escape state.
2421 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2422 int ti = n->in(TypeFunc::Parms)->_idx;
2423 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2424 if (nt == PointsToNode::UnknownType) {
2425 _delayed_worklist.push(n); // Process it later.
2426 break;
2427 } else if (nt == PointsToNode::JavaObject) {
2428 add_pointsto_edge(n->_idx, ti);
2429 } else {
2430 add_deferred_edge(n->_idx, ti);
2431 }
2432 }
2433 _processed.set(n->_idx);
2434 break;
2435 }
2436 case Op_StoreP:
2437 case Op_StoreN:
2438 {
2439 const Type *adr_type = phase->type(n->in(MemNode::Address));
2440 adr_type = adr_type->make_ptr();
2441 if (adr_type->isa_oopptr()) {
2442 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2443 } else {
2444 Node* adr = n->in(MemNode::Address);
2445 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2446 adr->in(AddPNode::Address)->is_Proj() &&
2447 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2448 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2449 // We are computing a raw address for a store captured
2450 // by an Initialize compute an appropriate address type.
2451 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2452 assert(offs != Type::OffsetBot, "offset must be a constant");
2453 } else {
2454 _processed.set(n->_idx);
2455 return;
2456 }
2457 }
2458 break;
2459 }
2460 case Op_StorePConditional:
2461 case Op_CompareAndSwapP:
2462 case Op_CompareAndSwapN:
2463 {
2464 const Type *adr_type = phase->type(n->in(MemNode::Address));
2465 adr_type = adr_type->make_ptr();
2466 if (adr_type->isa_oopptr()) {
2467 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2468 } else {
2469 _processed.set(n->_idx);
2470 return;
2471 }
2472 break;
2473 }
2474 case Op_AryEq:
2475 case Op_StrComp:
2476 case Op_StrEquals:
2477 case Op_StrIndexOf:
2478 {
2479 // char[] arrays passed to string intrinsics are not scalar replaceable.
2480 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2481 break;
2482 }
2483 case Op_ThreadLocal:
2484 {
2485 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2486 break;
2487 }
2488 default:
2489 ;
2490 // nothing to do
2491 }
2492 return;
2493 }
2495 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2496 uint n_idx = n->_idx;
2497 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered");
2499 // Don't set processed bit for AddP, LoadP, StoreP since
2500 // they may need more then one pass to process.
2501 // Also don't mark as processed Call nodes since their
2502 // arguments may need more then one pass to process.
2503 if (_processed.test(n_idx))
2504 return; // No need to redefine node's state.
2506 if (n->is_Call()) {
2507 CallNode *call = n->as_Call();
2508 process_call_arguments(call, phase);
2509 return;
2510 }
2512 switch (n->Opcode()) {
2513 case Op_AddP:
2514 {
2515 Node *base = get_addp_base(n);
2516 // Create a field edge to this node from everything base could point to.
2517 VectorSet ptset(Thread::current()->resource_area());
2518 PointsTo(ptset, base);
2519 for( VectorSetI i(&ptset); i.test(); ++i ) {
2520 uint pt = i.elem;
2521 add_field_edge(pt, n_idx, address_offset(n, phase));
2522 }
2523 break;
2524 }
2525 case Op_CastX2P:
2526 {
2527 assert(false, "Op_CastX2P");
2528 break;
2529 }
2530 case Op_CastPP:
2531 case Op_CheckCastPP:
2532 case Op_EncodeP:
2533 case Op_DecodeN:
2534 {
2535 int ti = n->in(1)->_idx;
2536 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered");
2537 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2538 add_pointsto_edge(n_idx, ti);
2539 } else {
2540 add_deferred_edge(n_idx, ti);
2541 }
2542 _processed.set(n_idx);
2543 break;
2544 }
2545 case Op_ConP:
2546 {
2547 assert(false, "Op_ConP");
2548 break;
2549 }
2550 case Op_ConN:
2551 {
2552 assert(false, "Op_ConN");
2553 break;
2554 }
2555 case Op_CreateEx:
2556 {
2557 assert(false, "Op_CreateEx");
2558 break;
2559 }
2560 case Op_LoadKlass:
2561 case Op_LoadNKlass:
2562 {
2563 assert(false, "Op_LoadKlass");
2564 break;
2565 }
2566 case Op_LoadP:
2567 case Op_LoadN:
2568 {
2569 const Type *t = phase->type(n);
2570 #ifdef ASSERT
2571 if (t->make_ptr() == NULL)
2572 assert(false, "Op_LoadP");
2573 #endif
2575 Node* adr = n->in(MemNode::Address)->uncast();
2576 Node* adr_base;
2577 if (adr->is_AddP()) {
2578 adr_base = get_addp_base(adr);
2579 } else {
2580 adr_base = adr;
2581 }
2583 // For everything "adr_base" could point to, create a deferred edge from
2584 // this node to each field with the same offset.
2585 VectorSet ptset(Thread::current()->resource_area());
2586 PointsTo(ptset, adr_base);
2587 int offset = address_offset(adr, phase);
2588 for( VectorSetI i(&ptset); i.test(); ++i ) {
2589 uint pt = i.elem;
2590 add_deferred_edge_to_fields(n_idx, pt, offset);
2591 }
2592 break;
2593 }
2594 case Op_Parm:
2595 {
2596 assert(false, "Op_Parm");
2597 break;
2598 }
2599 case Op_Phi:
2600 {
2601 #ifdef ASSERT
2602 const Type *t = n->as_Phi()->type();
2603 if (t->make_ptr() == NULL)
2604 assert(false, "Op_Phi");
2605 #endif
2606 for (uint i = 1; i < n->req() ; i++) {
2607 Node* in = n->in(i);
2608 if (in == NULL)
2609 continue; // ignore NULL
2610 in = in->uncast();
2611 if (in->is_top() || in == n)
2612 continue; // ignore top or inputs which go back this node
2613 int ti = in->_idx;
2614 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2615 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
2616 if (nt == PointsToNode::JavaObject) {
2617 add_pointsto_edge(n_idx, ti);
2618 } else {
2619 add_deferred_edge(n_idx, ti);
2620 }
2621 }
2622 _processed.set(n_idx);
2623 break;
2624 }
2625 case Op_Proj:
2626 {
2627 // we are only interested in the oop result projection from a call
2628 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2629 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType,
2630 "all nodes should be registered");
2631 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2632 assert(r->cnt() > TypeFunc::Parms, "sanity");
2633 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2634 process_call_result(n->as_Proj(), phase);
2635 assert(_processed.test(n_idx), "all call results should be processed");
2636 break;
2637 }
2638 }
2639 assert(false, "Op_Proj");
2640 break;
2641 }
2642 case Op_Return:
2643 {
2644 #ifdef ASSERT
2645 if( n->req() <= TypeFunc::Parms ||
2646 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2647 assert(false, "Op_Return");
2648 }
2649 #endif
2650 int ti = n->in(TypeFunc::Parms)->_idx;
2651 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered");
2652 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2653 add_pointsto_edge(n_idx, ti);
2654 } else {
2655 add_deferred_edge(n_idx, ti);
2656 }
2657 _processed.set(n_idx);
2658 break;
2659 }
2660 case Op_StoreP:
2661 case Op_StoreN:
2662 case Op_StorePConditional:
2663 case Op_CompareAndSwapP:
2664 case Op_CompareAndSwapN:
2665 {
2666 Node *adr = n->in(MemNode::Address);
2667 const Type *adr_type = phase->type(adr)->make_ptr();
2668 #ifdef ASSERT
2669 if (!adr_type->isa_oopptr())
2670 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
2671 #endif
2673 assert(adr->is_AddP(), "expecting an AddP");
2674 Node *adr_base = get_addp_base(adr);
2675 Node *val = n->in(MemNode::ValueIn)->uncast();
2676 // For everything "adr_base" could point to, create a deferred edge
2677 // to "val" from each field with the same offset.
2678 VectorSet ptset(Thread::current()->resource_area());
2679 PointsTo(ptset, adr_base);
2680 for( VectorSetI i(&ptset); i.test(); ++i ) {
2681 uint pt = i.elem;
2682 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase));
2683 }
2684 break;
2685 }
2686 case Op_AryEq:
2687 case Op_StrComp:
2688 case Op_StrEquals:
2689 case Op_StrIndexOf:
2690 {
2691 // char[] arrays passed to string intrinsic do not escape but
2692 // they are not scalar replaceable. Adjust escape state for them.
2693 // Start from in(2) edge since in(1) is memory edge.
2694 for (uint i = 2; i < n->req(); i++) {
2695 Node* adr = n->in(i)->uncast();
2696 const Type *at = phase->type(adr);
2697 if (!adr->is_top() && at->isa_ptr()) {
2698 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
2699 at->isa_ptr() != NULL, "expecting an Ptr");
2700 if (adr->is_AddP()) {
2701 adr = get_addp_base(adr);
2702 }
2703 // Mark as ArgEscape everything "adr" could point to.
2704 set_escape_state(adr->_idx, PointsToNode::ArgEscape);
2705 }
2706 }
2707 _processed.set(n_idx);
2708 break;
2709 }
2710 case Op_ThreadLocal:
2711 {
2712 assert(false, "Op_ThreadLocal");
2713 break;
2714 }
2715 default:
2716 // This method should be called only for EA specific nodes.
2717 ShouldNotReachHere();
2718 }
2719 }
2721 #ifndef PRODUCT
2722 void ConnectionGraph::dump() {
2723 bool first = true;
2725 uint size = nodes_size();
2726 for (uint ni = 0; ni < size; ni++) {
2727 PointsToNode *ptn = ptnode_adr(ni);
2728 PointsToNode::NodeType ptn_type = ptn->node_type();
2730 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
2731 continue;
2732 PointsToNode::EscapeState es = escape_state(ptn->_node);
2733 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
2734 if (first) {
2735 tty->cr();
2736 tty->print("======== Connection graph for ");
2737 _compile->method()->print_short_name();
2738 tty->cr();
2739 first = false;
2740 }
2741 tty->print("%6d ", ni);
2742 ptn->dump();
2743 // Print all locals which reference this allocation
2744 for (uint li = ni; li < size; li++) {
2745 PointsToNode *ptn_loc = ptnode_adr(li);
2746 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
2747 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
2748 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
2749 ptnode_adr(li)->dump(false);
2750 }
2751 }
2752 if (Verbose) {
2753 // Print all fields which reference this allocation
2754 for (uint i = 0; i < ptn->edge_count(); i++) {
2755 uint ei = ptn->edge_target(i);
2756 ptnode_adr(ei)->dump(false);
2757 }
2758 }
2759 tty->cr();
2760 }
2761 }
2762 }
2763 #endif