Mon, 31 Oct 2011 03:06:42 -0700
7104561: UseRDPCForConstantTableBase doesn't work after shorten branches changes
Reviewed-by: never, kvn
1 /*
2 * Copyright (c) 2005, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "libadt/vectset.hpp"
28 #include "memory/allocation.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/compile.hpp"
33 #include "opto/escape.hpp"
34 #include "opto/phaseX.hpp"
35 #include "opto/rootnode.hpp"
37 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) {
38 uint v = (targIdx << EdgeShift) + ((uint) et);
39 if (_edges == NULL) {
40 Arena *a = Compile::current()->comp_arena();
41 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0);
42 }
43 _edges->append_if_missing(v);
44 }
46 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) {
47 uint v = (targIdx << EdgeShift) + ((uint) et);
49 _edges->remove(v);
50 }
52 #ifndef PRODUCT
53 static const char *node_type_names[] = {
54 "UnknownType",
55 "JavaObject",
56 "LocalVar",
57 "Field"
58 };
60 static const char *esc_names[] = {
61 "UnknownEscape",
62 "NoEscape",
63 "ArgEscape",
64 "GlobalEscape"
65 };
67 static const char *edge_type_suffix[] = {
68 "?", // UnknownEdge
69 "P", // PointsToEdge
70 "D", // DeferredEdge
71 "F" // FieldEdge
72 };
74 void PointsToNode::dump(bool print_state) const {
75 NodeType nt = node_type();
76 tty->print("%s ", node_type_names[(int) nt]);
77 if (print_state) {
78 EscapeState es = escape_state();
79 tty->print("%s %s ", esc_names[(int) es], _scalar_replaceable ? "":"NSR");
80 }
81 tty->print("[[");
82 for (uint i = 0; i < edge_count(); i++) {
83 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]);
84 }
85 tty->print("]] ");
86 if (_node == NULL)
87 tty->print_cr("<null>");
88 else
89 _node->dump();
90 }
91 #endif
93 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
94 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()),
95 _processed(C->comp_arena()),
96 pt_ptset(C->comp_arena()),
97 pt_visited(C->comp_arena()),
98 pt_worklist(C->comp_arena(), 4, 0, 0),
99 _collecting(true),
100 _progress(false),
101 _compile(C),
102 _igvn(igvn),
103 _node_map(C->comp_arena()) {
105 _phantom_object = C->top()->_idx,
106 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true);
108 // Add ConP(#NULL) and ConN(#NULL) nodes.
109 Node* oop_null = igvn->zerocon(T_OBJECT);
110 _oop_null = oop_null->_idx;
111 assert(_oop_null < nodes_size(), "should be created already");
112 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
114 if (UseCompressedOops) {
115 Node* noop_null = igvn->zerocon(T_NARROWOOP);
116 _noop_null = noop_null->_idx;
117 assert(_noop_null < nodes_size(), "should be created already");
118 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
119 } else {
120 _noop_null = _oop_null; // Should be initialized
121 }
122 }
124 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) {
125 PointsToNode *f = ptnode_adr(from_i);
126 PointsToNode *t = ptnode_adr(to_i);
128 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
129 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge");
130 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge");
131 add_edge(f, to_i, PointsToNode::PointsToEdge);
132 }
134 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) {
135 PointsToNode *f = ptnode_adr(from_i);
136 PointsToNode *t = ptnode_adr(to_i);
138 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
139 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
140 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
141 // don't add a self-referential edge, this can occur during removal of
142 // deferred edges
143 if (from_i != to_i)
144 add_edge(f, to_i, PointsToNode::DeferredEdge);
145 }
147 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
148 const Type *adr_type = phase->type(adr);
149 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
150 adr->in(AddPNode::Address)->is_Proj() &&
151 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
152 // We are computing a raw address for a store captured by an Initialize
153 // compute an appropriate address type. AddP cases #3 and #5 (see below).
154 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
155 assert(offs != Type::OffsetBot ||
156 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
157 "offset must be a constant or it is initialization of array");
158 return offs;
159 }
160 const TypePtr *t_ptr = adr_type->isa_ptr();
161 assert(t_ptr != NULL, "must be a pointer type");
162 return t_ptr->offset();
163 }
165 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) {
166 PointsToNode *f = ptnode_adr(from_i);
167 PointsToNode *t = ptnode_adr(to_i);
169 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
170 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge");
171 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge");
172 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets");
173 t->set_offset(offset);
175 add_edge(f, to_i, PointsToNode::FieldEdge);
176 }
178 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) {
179 // Don't change non-escaping state of NULL pointer.
180 if (ni == _noop_null || ni == _oop_null)
181 return;
182 PointsToNode *npt = ptnode_adr(ni);
183 PointsToNode::EscapeState old_es = npt->escape_state();
184 if (es > old_es)
185 npt->set_escape_state(es);
186 }
188 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt,
189 PointsToNode::EscapeState es, bool done) {
190 PointsToNode* ptadr = ptnode_adr(n->_idx);
191 ptadr->_node = n;
192 ptadr->set_node_type(nt);
194 // inline set_escape_state(idx, es);
195 PointsToNode::EscapeState old_es = ptadr->escape_state();
196 if (es > old_es)
197 ptadr->set_escape_state(es);
199 if (done)
200 _processed.set(n->_idx);
201 }
203 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) {
204 uint idx = n->_idx;
205 PointsToNode::EscapeState es;
207 // If we are still collecting or there were no non-escaping allocations
208 // we don't know the answer yet
209 if (_collecting)
210 return PointsToNode::UnknownEscape;
212 // if the node was created after the escape computation, return
213 // UnknownEscape
214 if (idx >= nodes_size())
215 return PointsToNode::UnknownEscape;
217 es = ptnode_adr(idx)->escape_state();
219 // if we have already computed a value, return it
220 if (es != PointsToNode::UnknownEscape &&
221 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject)
222 return es;
224 // PointsTo() calls n->uncast() which can return a new ideal node.
225 if (n->uncast()->_idx >= nodes_size())
226 return PointsToNode::UnknownEscape;
228 PointsToNode::EscapeState orig_es = es;
230 // compute max escape state of anything this node could point to
231 for(VectorSetI i(PointsTo(n)); i.test() && es != PointsToNode::GlobalEscape; ++i) {
232 uint pt = i.elem;
233 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
234 if (pes > es)
235 es = pes;
236 }
237 if (orig_es != es) {
238 // cache the computed escape state
239 assert(es > orig_es, "should have computed an escape state");
240 set_escape_state(idx, es);
241 } // orig_es could be PointsToNode::UnknownEscape
242 return es;
243 }
245 VectorSet* ConnectionGraph::PointsTo(Node * n) {
246 pt_ptset.Reset();
247 pt_visited.Reset();
248 pt_worklist.clear();
250 #ifdef ASSERT
251 Node *orig_n = n;
252 #endif
254 n = n->uncast();
255 PointsToNode* npt = ptnode_adr(n->_idx);
257 // If we have a JavaObject, return just that object
258 if (npt->node_type() == PointsToNode::JavaObject) {
259 pt_ptset.set(n->_idx);
260 return &pt_ptset;
261 }
262 #ifdef ASSERT
263 if (npt->_node == NULL) {
264 if (orig_n != n)
265 orig_n->dump();
266 n->dump();
267 assert(npt->_node != NULL, "unregistered node");
268 }
269 #endif
270 pt_worklist.push(n->_idx);
271 while(pt_worklist.length() > 0) {
272 int ni = pt_worklist.pop();
273 if (pt_visited.test_set(ni))
274 continue;
276 PointsToNode* pn = ptnode_adr(ni);
277 // ensure that all inputs of a Phi have been processed
278 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
280 int edges_processed = 0;
281 uint e_cnt = pn->edge_count();
282 for (uint e = 0; e < e_cnt; e++) {
283 uint etgt = pn->edge_target(e);
284 PointsToNode::EdgeType et = pn->edge_type(e);
285 if (et == PointsToNode::PointsToEdge) {
286 pt_ptset.set(etgt);
287 edges_processed++;
288 } else if (et == PointsToNode::DeferredEdge) {
289 pt_worklist.push(etgt);
290 edges_processed++;
291 } else {
292 assert(false,"neither PointsToEdge or DeferredEdge");
293 }
294 }
295 if (edges_processed == 0) {
296 // no deferred or pointsto edges found. Assume the value was set
297 // outside this method. Add the phantom object to the pointsto set.
298 pt_ptset.set(_phantom_object);
299 }
300 }
301 return &pt_ptset;
302 }
304 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
305 // This method is most expensive during ConnectionGraph construction.
306 // Reuse vectorSet and an additional growable array for deferred edges.
307 deferred_edges->clear();
308 visited->Reset();
310 visited->set(ni);
311 PointsToNode *ptn = ptnode_adr(ni);
313 // Mark current edges as visited and move deferred edges to separate array.
314 for (uint i = 0; i < ptn->edge_count(); ) {
315 uint t = ptn->edge_target(i);
316 #ifdef ASSERT
317 assert(!visited->test_set(t), "expecting no duplications");
318 #else
319 visited->set(t);
320 #endif
321 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
322 ptn->remove_edge(t, PointsToNode::DeferredEdge);
323 deferred_edges->append(t);
324 } else {
325 i++;
326 }
327 }
328 for (int next = 0; next < deferred_edges->length(); ++next) {
329 uint t = deferred_edges->at(next);
330 PointsToNode *ptt = ptnode_adr(t);
331 uint e_cnt = ptt->edge_count();
332 for (uint e = 0; e < e_cnt; e++) {
333 uint etgt = ptt->edge_target(e);
334 if (visited->test_set(etgt))
335 continue;
337 PointsToNode::EdgeType et = ptt->edge_type(e);
338 if (et == PointsToNode::PointsToEdge) {
339 add_pointsto_edge(ni, etgt);
340 if(etgt == _phantom_object) {
341 // Special case - field set outside (globally escaping).
342 set_escape_state(ni, PointsToNode::GlobalEscape);
343 }
344 } else if (et == PointsToNode::DeferredEdge) {
345 deferred_edges->append(etgt);
346 } else {
347 assert(false,"invalid connection graph");
348 }
349 }
350 }
351 }
354 // Add an edge to node given by "to_i" from any field of adr_i whose offset
355 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
356 // a pointsto edge is added if it is a JavaObject
358 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
359 PointsToNode* an = ptnode_adr(adr_i);
360 PointsToNode* to = ptnode_adr(to_i);
361 bool deferred = (to->node_type() == PointsToNode::LocalVar);
363 for (uint fe = 0; fe < an->edge_count(); fe++) {
364 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
365 int fi = an->edge_target(fe);
366 PointsToNode* pf = ptnode_adr(fi);
367 int po = pf->offset();
368 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
369 if (deferred)
370 add_deferred_edge(fi, to_i);
371 else
372 add_pointsto_edge(fi, to_i);
373 }
374 }
375 }
377 // Add a deferred edge from node given by "from_i" to any field of adr_i
378 // whose offset matches "offset".
379 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
380 PointsToNode* an = ptnode_adr(adr_i);
381 for (uint fe = 0; fe < an->edge_count(); fe++) {
382 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
383 int fi = an->edge_target(fe);
384 PointsToNode* pf = ptnode_adr(fi);
385 int po = pf->offset();
386 if (pf->edge_count() == 0) {
387 // we have not seen any stores to this field, assume it was set outside this method
388 add_pointsto_edge(fi, _phantom_object);
389 }
390 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
391 add_deferred_edge(from_i, fi);
392 }
393 }
394 }
396 // Helper functions
398 static Node* get_addp_base(Node *addp) {
399 assert(addp->is_AddP(), "must be AddP");
400 //
401 // AddP cases for Base and Address inputs:
402 // case #1. Direct object's field reference:
403 // Allocate
404 // |
405 // Proj #5 ( oop result )
406 // |
407 // CheckCastPP (cast to instance type)
408 // | |
409 // AddP ( base == address )
410 //
411 // case #2. Indirect object's field reference:
412 // Phi
413 // |
414 // CastPP (cast to instance type)
415 // | |
416 // AddP ( base == address )
417 //
418 // case #3. Raw object's field reference for Initialize node:
419 // Allocate
420 // |
421 // Proj #5 ( oop result )
422 // top |
423 // \ |
424 // AddP ( base == top )
425 //
426 // case #4. Array's element reference:
427 // {CheckCastPP | CastPP}
428 // | | |
429 // | AddP ( array's element offset )
430 // | |
431 // AddP ( array's offset )
432 //
433 // case #5. Raw object's field reference for arraycopy stub call:
434 // The inline_native_clone() case when the arraycopy stub is called
435 // after the allocation before Initialize and CheckCastPP nodes.
436 // Allocate
437 // |
438 // Proj #5 ( oop result )
439 // | |
440 // AddP ( base == address )
441 //
442 // case #6. Constant Pool, ThreadLocal, CastX2P or
443 // Raw object's field reference:
444 // {ConP, ThreadLocal, CastX2P, raw Load}
445 // top |
446 // \ |
447 // AddP ( base == top )
448 //
449 // case #7. Klass's field reference.
450 // LoadKlass
451 // | |
452 // AddP ( base == address )
453 //
454 // case #8. narrow Klass's field reference.
455 // LoadNKlass
456 // |
457 // DecodeN
458 // | |
459 // AddP ( base == address )
460 //
461 Node *base = addp->in(AddPNode::Base)->uncast();
462 if (base->is_top()) { // The AddP case #3 and #6.
463 base = addp->in(AddPNode::Address)->uncast();
464 while (base->is_AddP()) {
465 // Case #6 (unsafe access) may have several chained AddP nodes.
466 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only");
467 base = base->in(AddPNode::Address)->uncast();
468 }
469 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
470 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
471 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
472 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
473 }
474 return base;
475 }
477 static Node* find_second_addp(Node* addp, Node* n) {
478 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
480 Node* addp2 = addp->raw_out(0);
481 if (addp->outcnt() == 1 && addp2->is_AddP() &&
482 addp2->in(AddPNode::Base) == n &&
483 addp2->in(AddPNode::Address) == addp) {
485 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
486 //
487 // Find array's offset to push it on worklist first and
488 // as result process an array's element offset first (pushed second)
489 // to avoid CastPP for the array's offset.
490 // Otherwise the inserted CastPP (LocalVar) will point to what
491 // the AddP (Field) points to. Which would be wrong since
492 // the algorithm expects the CastPP has the same point as
493 // as AddP's base CheckCastPP (LocalVar).
494 //
495 // ArrayAllocation
496 // |
497 // CheckCastPP
498 // |
499 // memProj (from ArrayAllocation CheckCastPP)
500 // | ||
501 // | || Int (element index)
502 // | || | ConI (log(element size))
503 // | || | /
504 // | || LShift
505 // | || /
506 // | AddP (array's element offset)
507 // | |
508 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
509 // | / /
510 // AddP (array's offset)
511 // |
512 // Load/Store (memory operation on array's element)
513 //
514 return addp2;
515 }
516 return NULL;
517 }
519 //
520 // Adjust the type and inputs of an AddP which computes the
521 // address of a field of an instance
522 //
523 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
524 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
525 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
526 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
527 if (t == NULL) {
528 // We are computing a raw address for a store captured by an Initialize
529 // compute an appropriate address type (cases #3 and #5).
530 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
531 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
532 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
533 assert(offs != Type::OffsetBot, "offset must be a constant");
534 t = base_t->add_offset(offs)->is_oopptr();
535 }
536 int inst_id = base_t->instance_id();
537 assert(!t->is_known_instance() || t->instance_id() == inst_id,
538 "old type must be non-instance or match new type");
540 // The type 't' could be subclass of 'base_t'.
541 // As result t->offset() could be large then base_t's size and it will
542 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
543 // constructor verifies correctness of the offset.
544 //
545 // It could happened on subclass's branch (from the type profiling
546 // inlining) which was not eliminated during parsing since the exactness
547 // of the allocation type was not propagated to the subclass type check.
548 //
549 // Or the type 't' could be not related to 'base_t' at all.
550 // It could happened when CHA type is different from MDO type on a dead path
551 // (for example, from instanceof check) which is not collapsed during parsing.
552 //
553 // Do nothing for such AddP node and don't process its users since
554 // this code branch will go away.
555 //
556 if (!t->is_known_instance() &&
557 !base_t->klass()->is_subtype_of(t->klass())) {
558 return false; // bail out
559 }
561 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
562 // Do NOT remove the next line: ensure a new alias index is allocated
563 // for the instance type. Note: C++ will not remove it since the call
564 // has side effect.
565 int alias_idx = _compile->get_alias_index(tinst);
566 igvn->set_type(addp, tinst);
567 // record the allocation in the node map
568 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered");
569 set_map(addp->_idx, get_map(base->_idx));
571 // Set addp's Base and Address to 'base'.
572 Node *abase = addp->in(AddPNode::Base);
573 Node *adr = addp->in(AddPNode::Address);
574 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
575 adr->in(0)->_idx == (uint)inst_id) {
576 // Skip AddP cases #3 and #5.
577 } else {
578 assert(!abase->is_top(), "sanity"); // AddP case #3
579 if (abase != base) {
580 igvn->hash_delete(addp);
581 addp->set_req(AddPNode::Base, base);
582 if (abase == adr) {
583 addp->set_req(AddPNode::Address, base);
584 } else {
585 // AddP case #4 (adr is array's element offset AddP node)
586 #ifdef ASSERT
587 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
588 assert(adr->is_AddP() && atype != NULL &&
589 atype->instance_id() == inst_id, "array's element offset should be processed first");
590 #endif
591 }
592 igvn->hash_insert(addp);
593 }
594 }
595 // Put on IGVN worklist since at least addp's type was changed above.
596 record_for_optimizer(addp);
597 return true;
598 }
600 //
601 // Create a new version of orig_phi if necessary. Returns either the newly
602 // created phi or an existing phi. Sets create_new to indicate whether a new
603 // phi was created. Cache the last newly created phi in the node map.
604 //
605 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
606 Compile *C = _compile;
607 new_created = false;
608 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
609 // nothing to do if orig_phi is bottom memory or matches alias_idx
610 if (phi_alias_idx == alias_idx) {
611 return orig_phi;
612 }
613 // Have we recently created a Phi for this alias index?
614 PhiNode *result = get_map_phi(orig_phi->_idx);
615 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
616 return result;
617 }
618 // Previous check may fail when the same wide memory Phi was split into Phis
619 // for different memory slices. Search all Phis for this region.
620 if (result != NULL) {
621 Node* region = orig_phi->in(0);
622 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
623 Node* phi = region->fast_out(i);
624 if (phi->is_Phi() &&
625 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
626 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
627 return phi->as_Phi();
628 }
629 }
630 }
631 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
632 if (C->do_escape_analysis() == true && !C->failing()) {
633 // Retry compilation without escape analysis.
634 // If this is the first failure, the sentinel string will "stick"
635 // to the Compile object, and the C2Compiler will see it and retry.
636 C->record_failure(C2Compiler::retry_no_escape_analysis());
637 }
638 return NULL;
639 }
640 orig_phi_worklist.append_if_missing(orig_phi);
641 const TypePtr *atype = C->get_adr_type(alias_idx);
642 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
643 C->copy_node_notes_to(result, orig_phi);
644 igvn->set_type(result, result->bottom_type());
645 record_for_optimizer(result);
647 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;)
648 assert(pn == NULL || pn == orig_phi, "wrong node");
649 set_map(orig_phi->_idx, result);
650 ptnode_adr(orig_phi->_idx)->_node = orig_phi;
652 new_created = true;
653 return result;
654 }
656 //
657 // Return a new version of Memory Phi "orig_phi" with the inputs having the
658 // specified alias index.
659 //
660 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
662 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
663 Compile *C = _compile;
664 bool new_phi_created;
665 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
666 if (!new_phi_created) {
667 return result;
668 }
670 GrowableArray<PhiNode *> phi_list;
671 GrowableArray<uint> cur_input;
673 PhiNode *phi = orig_phi;
674 uint idx = 1;
675 bool finished = false;
676 while(!finished) {
677 while (idx < phi->req()) {
678 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
679 if (mem != NULL && mem->is_Phi()) {
680 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
681 if (new_phi_created) {
682 // found an phi for which we created a new split, push current one on worklist and begin
683 // processing new one
684 phi_list.push(phi);
685 cur_input.push(idx);
686 phi = mem->as_Phi();
687 result = newphi;
688 idx = 1;
689 continue;
690 } else {
691 mem = newphi;
692 }
693 }
694 if (C->failing()) {
695 return NULL;
696 }
697 result->set_req(idx++, mem);
698 }
699 #ifdef ASSERT
700 // verify that the new Phi has an input for each input of the original
701 assert( phi->req() == result->req(), "must have same number of inputs.");
702 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
703 #endif
704 // Check if all new phi's inputs have specified alias index.
705 // Otherwise use old phi.
706 for (uint i = 1; i < phi->req(); i++) {
707 Node* in = result->in(i);
708 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
709 }
710 // we have finished processing a Phi, see if there are any more to do
711 finished = (phi_list.length() == 0 );
712 if (!finished) {
713 phi = phi_list.pop();
714 idx = cur_input.pop();
715 PhiNode *prev_result = get_map_phi(phi->_idx);
716 prev_result->set_req(idx++, result);
717 result = prev_result;
718 }
719 }
720 return result;
721 }
724 //
725 // The next methods are derived from methods in MemNode.
726 //
727 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
728 Node *mem = mmem;
729 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
730 // means an array I have not precisely typed yet. Do not do any
731 // alias stuff with it any time soon.
732 if( toop->base() != Type::AnyPtr &&
733 !(toop->klass() != NULL &&
734 toop->klass()->is_java_lang_Object() &&
735 toop->offset() == Type::OffsetBot) ) {
736 mem = mmem->memory_at(alias_idx);
737 // Update input if it is progress over what we have now
738 }
739 return mem;
740 }
742 //
743 // Move memory users to their memory slices.
744 //
745 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) {
746 Compile* C = _compile;
748 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
749 assert(tp != NULL, "ptr type");
750 int alias_idx = C->get_alias_index(tp);
751 int general_idx = C->get_general_index(alias_idx);
753 // Move users first
754 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
755 Node* use = n->fast_out(i);
756 if (use->is_MergeMem()) {
757 MergeMemNode* mmem = use->as_MergeMem();
758 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
759 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
760 continue; // Nothing to do
761 }
762 // Replace previous general reference to mem node.
763 uint orig_uniq = C->unique();
764 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
765 assert(orig_uniq == C->unique(), "no new nodes");
766 mmem->set_memory_at(general_idx, m);
767 --imax;
768 --i;
769 } else if (use->is_MemBar()) {
770 assert(!use->is_Initialize(), "initializing stores should not be moved");
771 if (use->req() > MemBarNode::Precedent &&
772 use->in(MemBarNode::Precedent) == n) {
773 // Don't move related membars.
774 record_for_optimizer(use);
775 continue;
776 }
777 tp = use->as_MemBar()->adr_type()->isa_ptr();
778 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
779 alias_idx == general_idx) {
780 continue; // Nothing to do
781 }
782 // Move to general memory slice.
783 uint orig_uniq = C->unique();
784 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
785 assert(orig_uniq == C->unique(), "no new nodes");
786 igvn->hash_delete(use);
787 imax -= use->replace_edge(n, m);
788 igvn->hash_insert(use);
789 record_for_optimizer(use);
790 --i;
791 #ifdef ASSERT
792 } else if (use->is_Mem()) {
793 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
794 // Don't move related cardmark.
795 continue;
796 }
797 // Memory nodes should have new memory input.
798 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
799 assert(tp != NULL, "ptr type");
800 int idx = C->get_alias_index(tp);
801 assert(get_map(use->_idx) != NULL || idx == alias_idx,
802 "Following memory nodes should have new memory input or be on the same memory slice");
803 } else if (use->is_Phi()) {
804 // Phi nodes should be split and moved already.
805 tp = use->as_Phi()->adr_type()->isa_ptr();
806 assert(tp != NULL, "ptr type");
807 int idx = C->get_alias_index(tp);
808 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
809 } else {
810 use->dump();
811 assert(false, "should not be here");
812 #endif
813 }
814 }
815 }
817 //
818 // Search memory chain of "mem" to find a MemNode whose address
819 // is the specified alias index.
820 //
821 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
822 if (orig_mem == NULL)
823 return orig_mem;
824 Compile* C = phase->C;
825 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
826 bool is_instance = (toop != NULL) && toop->is_known_instance();
827 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
828 Node *prev = NULL;
829 Node *result = orig_mem;
830 while (prev != result) {
831 prev = result;
832 if (result == start_mem)
833 break; // hit one of our sentinels
834 if (result->is_Mem()) {
835 const Type *at = phase->type(result->in(MemNode::Address));
836 if (at == Type::TOP)
837 break; // Dead
838 assert (at->isa_ptr() != NULL, "pointer type required.");
839 int idx = C->get_alias_index(at->is_ptr());
840 if (idx == alias_idx)
841 break; // Found
842 if (!is_instance && (at->isa_oopptr() == NULL ||
843 !at->is_oopptr()->is_known_instance())) {
844 break; // Do not skip store to general memory slice.
845 }
846 result = result->in(MemNode::Memory);
847 }
848 if (!is_instance)
849 continue; // don't search further for non-instance types
850 // skip over a call which does not affect this memory slice
851 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
852 Node *proj_in = result->in(0);
853 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
854 break; // hit one of our sentinels
855 } else if (proj_in->is_Call()) {
856 CallNode *call = proj_in->as_Call();
857 if (!call->may_modify(toop, phase)) {
858 result = call->in(TypeFunc::Memory);
859 }
860 } else if (proj_in->is_Initialize()) {
861 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
862 // Stop if this is the initialization for the object instance which
863 // which contains this memory slice, otherwise skip over it.
864 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
865 result = proj_in->in(TypeFunc::Memory);
866 }
867 } else if (proj_in->is_MemBar()) {
868 result = proj_in->in(TypeFunc::Memory);
869 }
870 } else if (result->is_MergeMem()) {
871 MergeMemNode *mmem = result->as_MergeMem();
872 result = step_through_mergemem(mmem, alias_idx, toop);
873 if (result == mmem->base_memory()) {
874 // Didn't find instance memory, search through general slice recursively.
875 result = mmem->memory_at(C->get_general_index(alias_idx));
876 result = find_inst_mem(result, alias_idx, orig_phis, phase);
877 if (C->failing()) {
878 return NULL;
879 }
880 mmem->set_memory_at(alias_idx, result);
881 }
882 } else if (result->is_Phi() &&
883 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
884 Node *un = result->as_Phi()->unique_input(phase);
885 if (un != NULL) {
886 orig_phis.append_if_missing(result->as_Phi());
887 result = un;
888 } else {
889 break;
890 }
891 } else if (result->is_ClearArray()) {
892 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) {
893 // Can not bypass initialization of the instance
894 // we are looking for.
895 break;
896 }
897 // Otherwise skip it (the call updated 'result' value).
898 } else if (result->Opcode() == Op_SCMemProj) {
899 assert(result->in(0)->is_LoadStore(), "sanity");
900 const Type *at = phase->type(result->in(0)->in(MemNode::Address));
901 if (at != Type::TOP) {
902 assert (at->isa_ptr() != NULL, "pointer type required.");
903 int idx = C->get_alias_index(at->is_ptr());
904 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
905 break;
906 }
907 result = result->in(0)->in(MemNode::Memory);
908 }
909 }
910 if (result->is_Phi()) {
911 PhiNode *mphi = result->as_Phi();
912 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
913 const TypePtr *t = mphi->adr_type();
914 if (!is_instance) {
915 // Push all non-instance Phis on the orig_phis worklist to update inputs
916 // during Phase 4 if needed.
917 orig_phis.append_if_missing(mphi);
918 } else if (C->get_alias_index(t) != alias_idx) {
919 // Create a new Phi with the specified alias index type.
920 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
921 }
922 }
923 // the result is either MemNode, PhiNode, InitializeNode.
924 return result;
925 }
927 //
928 // Convert the types of unescaped object to instance types where possible,
929 // propagate the new type information through the graph, and update memory
930 // edges and MergeMem inputs to reflect the new type.
931 //
932 // We start with allocations (and calls which may be allocations) on alloc_worklist.
933 // The processing is done in 4 phases:
934 //
935 // Phase 1: Process possible allocations from alloc_worklist. Create instance
936 // types for the CheckCastPP for allocations where possible.
937 // Propagate the the new types through users as follows:
938 // casts and Phi: push users on alloc_worklist
939 // AddP: cast Base and Address inputs to the instance type
940 // push any AddP users on alloc_worklist and push any memnode
941 // users onto memnode_worklist.
942 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
943 // search the Memory chain for a store with the appropriate type
944 // address type. If a Phi is found, create a new version with
945 // the appropriate memory slices from each of the Phi inputs.
946 // For stores, process the users as follows:
947 // MemNode: push on memnode_worklist
948 // MergeMem: push on mergemem_worklist
949 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
950 // moving the first node encountered of each instance type to the
951 // the input corresponding to its alias index.
952 // appropriate memory slice.
953 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
954 //
955 // In the following example, the CheckCastPP nodes are the cast of allocation
956 // results and the allocation of node 29 is unescaped and eligible to be an
957 // instance type.
958 //
959 // We start with:
960 //
961 // 7 Parm #memory
962 // 10 ConI "12"
963 // 19 CheckCastPP "Foo"
964 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
965 // 29 CheckCastPP "Foo"
966 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
967 //
968 // 40 StoreP 25 7 20 ... alias_index=4
969 // 50 StoreP 35 40 30 ... alias_index=4
970 // 60 StoreP 45 50 20 ... alias_index=4
971 // 70 LoadP _ 60 30 ... alias_index=4
972 // 80 Phi 75 50 60 Memory alias_index=4
973 // 90 LoadP _ 80 30 ... alias_index=4
974 // 100 LoadP _ 80 20 ... alias_index=4
975 //
976 //
977 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
978 // and creating a new alias index for node 30. This gives:
979 //
980 // 7 Parm #memory
981 // 10 ConI "12"
982 // 19 CheckCastPP "Foo"
983 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
984 // 29 CheckCastPP "Foo" iid=24
985 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
986 //
987 // 40 StoreP 25 7 20 ... alias_index=4
988 // 50 StoreP 35 40 30 ... alias_index=6
989 // 60 StoreP 45 50 20 ... alias_index=4
990 // 70 LoadP _ 60 30 ... alias_index=6
991 // 80 Phi 75 50 60 Memory alias_index=4
992 // 90 LoadP _ 80 30 ... alias_index=6
993 // 100 LoadP _ 80 20 ... alias_index=4
994 //
995 // In phase 2, new memory inputs are computed for the loads and stores,
996 // And a new version of the phi is created. In phase 4, the inputs to
997 // node 80 are updated and then the memory nodes are updated with the
998 // values computed in phase 2. This results in:
999 //
1000 // 7 Parm #memory
1001 // 10 ConI "12"
1002 // 19 CheckCastPP "Foo"
1003 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
1004 // 29 CheckCastPP "Foo" iid=24
1005 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
1006 //
1007 // 40 StoreP 25 7 20 ... alias_index=4
1008 // 50 StoreP 35 7 30 ... alias_index=6
1009 // 60 StoreP 45 40 20 ... alias_index=4
1010 // 70 LoadP _ 50 30 ... alias_index=6
1011 // 80 Phi 75 40 60 Memory alias_index=4
1012 // 120 Phi 75 50 50 Memory alias_index=6
1013 // 90 LoadP _ 120 30 ... alias_index=6
1014 // 100 LoadP _ 80 20 ... alias_index=4
1015 //
1016 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
1017 GrowableArray<Node *> memnode_worklist;
1018 GrowableArray<PhiNode *> orig_phis;
1020 PhaseIterGVN *igvn = _igvn;
1021 uint new_index_start = (uint) _compile->num_alias_types();
1022 Arena* arena = Thread::current()->resource_area();
1023 VectorSet visited(arena);
1026 // Phase 1: Process possible allocations from alloc_worklist.
1027 // Create instance types for the CheckCastPP for allocations where possible.
1028 //
1029 // (Note: don't forget to change the order of the second AddP node on
1030 // the alloc_worklist if the order of the worklist processing is changed,
1031 // see the comment in find_second_addp().)
1032 //
1033 while (alloc_worklist.length() != 0) {
1034 Node *n = alloc_worklist.pop();
1035 uint ni = n->_idx;
1036 const TypeOopPtr* tinst = NULL;
1037 if (n->is_Call()) {
1038 CallNode *alloc = n->as_Call();
1039 // copy escape information to call node
1040 PointsToNode* ptn = ptnode_adr(alloc->_idx);
1041 PointsToNode::EscapeState es = escape_state(alloc);
1042 // We have an allocation or call which returns a Java object,
1043 // see if it is unescaped.
1044 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable)
1045 continue;
1047 // Find CheckCastPP for the allocate or for the return value of a call
1048 n = alloc->result_cast();
1049 if (n == NULL) { // No uses except Initialize node
1050 if (alloc->is_Allocate()) {
1051 // Set the scalar_replaceable flag for allocation
1052 // so it could be eliminated if it has no uses.
1053 alloc->as_Allocate()->_is_scalar_replaceable = true;
1054 }
1055 continue;
1056 }
1057 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
1058 assert(!alloc->is_Allocate(), "allocation should have unique type");
1059 continue;
1060 }
1062 // The inline code for Object.clone() casts the allocation result to
1063 // java.lang.Object and then to the actual type of the allocated
1064 // object. Detect this case and use the second cast.
1065 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
1066 // the allocation result is cast to java.lang.Object and then
1067 // to the actual Array type.
1068 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
1069 && (alloc->is_AllocateArray() ||
1070 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
1071 Node *cast2 = NULL;
1072 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1073 Node *use = n->fast_out(i);
1074 if (use->is_CheckCastPP()) {
1075 cast2 = use;
1076 break;
1077 }
1078 }
1079 if (cast2 != NULL) {
1080 n = cast2;
1081 } else {
1082 // Non-scalar replaceable if the allocation type is unknown statically
1083 // (reflection allocation), the object can't be restored during
1084 // deoptimization without precise type.
1085 continue;
1086 }
1087 }
1088 if (alloc->is_Allocate()) {
1089 // Set the scalar_replaceable flag for allocation
1090 // so it could be eliminated.
1091 alloc->as_Allocate()->_is_scalar_replaceable = true;
1092 }
1093 set_escape_state(n->_idx, es);
1094 // in order for an object to be scalar-replaceable, it must be:
1095 // - a direct allocation (not a call returning an object)
1096 // - non-escaping
1097 // - eligible to be a unique type
1098 // - not determined to be ineligible by escape analysis
1099 assert(ptnode_adr(alloc->_idx)->_node != NULL &&
1100 ptnode_adr(n->_idx)->_node != NULL, "should be registered");
1101 set_map(alloc->_idx, n);
1102 set_map(n->_idx, alloc);
1103 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
1104 if (t == NULL)
1105 continue; // not a TypeInstPtr
1106 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
1107 igvn->hash_delete(n);
1108 igvn->set_type(n, tinst);
1109 n->raise_bottom_type(tinst);
1110 igvn->hash_insert(n);
1111 record_for_optimizer(n);
1112 if (alloc->is_Allocate() && ptn->_scalar_replaceable &&
1113 (t->isa_instptr() || t->isa_aryptr())) {
1115 // First, put on the worklist all Field edges from Connection Graph
1116 // which is more accurate then putting immediate users from Ideal Graph.
1117 for (uint e = 0; e < ptn->edge_count(); e++) {
1118 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
1119 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
1120 "only AddP nodes are Field edges in CG");
1121 if (use->outcnt() > 0) { // Don't process dead nodes
1122 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
1123 if (addp2 != NULL) {
1124 assert(alloc->is_AllocateArray(),"array allocation was expected");
1125 alloc_worklist.append_if_missing(addp2);
1126 }
1127 alloc_worklist.append_if_missing(use);
1128 }
1129 }
1131 // An allocation may have an Initialize which has raw stores. Scan
1132 // the users of the raw allocation result and push AddP users
1133 // on alloc_worklist.
1134 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
1135 assert (raw_result != NULL, "must have an allocation result");
1136 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
1137 Node *use = raw_result->fast_out(i);
1138 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
1139 Node* addp2 = find_second_addp(use, raw_result);
1140 if (addp2 != NULL) {
1141 assert(alloc->is_AllocateArray(),"array allocation was expected");
1142 alloc_worklist.append_if_missing(addp2);
1143 }
1144 alloc_worklist.append_if_missing(use);
1145 } else if (use->is_MemBar()) {
1146 memnode_worklist.append_if_missing(use);
1147 }
1148 }
1149 }
1150 } else if (n->is_AddP()) {
1151 VectorSet* ptset = PointsTo(get_addp_base(n));
1152 assert(ptset->Size() == 1, "AddP address is unique");
1153 uint elem = ptset->getelem(); // Allocation node's index
1154 if (elem == _phantom_object) {
1155 assert(false, "escaped allocation");
1156 continue; // Assume the value was set outside this method.
1157 }
1158 Node *base = get_map(elem); // CheckCastPP node
1159 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path
1160 tinst = igvn->type(base)->isa_oopptr();
1161 } else if (n->is_Phi() ||
1162 n->is_CheckCastPP() ||
1163 n->is_EncodeP() ||
1164 n->is_DecodeN() ||
1165 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
1166 if (visited.test_set(n->_idx)) {
1167 assert(n->is_Phi(), "loops only through Phi's");
1168 continue; // already processed
1169 }
1170 VectorSet* ptset = PointsTo(n);
1171 if (ptset->Size() == 1) {
1172 uint elem = ptset->getelem(); // Allocation node's index
1173 if (elem == _phantom_object) {
1174 assert(false, "escaped allocation");
1175 continue; // Assume the value was set outside this method.
1176 }
1177 Node *val = get_map(elem); // CheckCastPP node
1178 TypeNode *tn = n->as_Type();
1179 tinst = igvn->type(val)->isa_oopptr();
1180 assert(tinst != NULL && tinst->is_known_instance() &&
1181 (uint)tinst->instance_id() == elem , "instance type expected.");
1183 const Type *tn_type = igvn->type(tn);
1184 const TypeOopPtr *tn_t;
1185 if (tn_type->isa_narrowoop()) {
1186 tn_t = tn_type->make_ptr()->isa_oopptr();
1187 } else {
1188 tn_t = tn_type->isa_oopptr();
1189 }
1191 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
1192 if (tn_type->isa_narrowoop()) {
1193 tn_type = tinst->make_narrowoop();
1194 } else {
1195 tn_type = tinst;
1196 }
1197 igvn->hash_delete(tn);
1198 igvn->set_type(tn, tn_type);
1199 tn->set_type(tn_type);
1200 igvn->hash_insert(tn);
1201 record_for_optimizer(n);
1202 } else {
1203 assert(tn_type == TypePtr::NULL_PTR ||
1204 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
1205 "unexpected type");
1206 continue; // Skip dead path with different type
1207 }
1208 }
1209 } else {
1210 debug_only(n->dump();)
1211 assert(false, "EA: unexpected node");
1212 continue;
1213 }
1214 // push allocation's users on appropriate worklist
1215 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1216 Node *use = n->fast_out(i);
1217 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1218 // Load/store to instance's field
1219 memnode_worklist.append_if_missing(use);
1220 } else if (use->is_MemBar()) {
1221 memnode_worklist.append_if_missing(use);
1222 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1223 Node* addp2 = find_second_addp(use, n);
1224 if (addp2 != NULL) {
1225 alloc_worklist.append_if_missing(addp2);
1226 }
1227 alloc_worklist.append_if_missing(use);
1228 } else if (use->is_Phi() ||
1229 use->is_CheckCastPP() ||
1230 use->is_EncodeP() ||
1231 use->is_DecodeN() ||
1232 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1233 alloc_worklist.append_if_missing(use);
1234 #ifdef ASSERT
1235 } else if (use->is_Mem()) {
1236 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
1237 } else if (use->is_MergeMem()) {
1238 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1239 } else if (use->is_SafePoint()) {
1240 // Look for MergeMem nodes for calls which reference unique allocation
1241 // (through CheckCastPP nodes) even for debug info.
1242 Node* m = use->in(TypeFunc::Memory);
1243 if (m->is_MergeMem()) {
1244 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1245 }
1246 } else {
1247 uint op = use->Opcode();
1248 if (!(op == Op_CmpP || op == Op_Conv2B ||
1249 op == Op_CastP2X || op == Op_StoreCM ||
1250 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
1251 op == Op_StrEquals || op == Op_StrIndexOf)) {
1252 n->dump();
1253 use->dump();
1254 assert(false, "EA: missing allocation reference path");
1255 }
1256 #endif
1257 }
1258 }
1260 }
1261 // New alias types were created in split_AddP().
1262 uint new_index_end = (uint) _compile->num_alias_types();
1264 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1265 // compute new values for Memory inputs (the Memory inputs are not
1266 // actually updated until phase 4.)
1267 if (memnode_worklist.length() == 0)
1268 return; // nothing to do
1270 while (memnode_worklist.length() != 0) {
1271 Node *n = memnode_worklist.pop();
1272 if (visited.test_set(n->_idx))
1273 continue;
1274 if (n->is_Phi() || n->is_ClearArray()) {
1275 // we don't need to do anything, but the users must be pushed
1276 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
1277 // we don't need to do anything, but the users must be pushed
1278 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
1279 if (n == NULL)
1280 continue;
1281 } else {
1282 assert(n->is_Mem(), "memory node required.");
1283 Node *addr = n->in(MemNode::Address);
1284 const Type *addr_t = igvn->type(addr);
1285 if (addr_t == Type::TOP)
1286 continue;
1287 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1288 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1289 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1290 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1291 if (_compile->failing()) {
1292 return;
1293 }
1294 if (mem != n->in(MemNode::Memory)) {
1295 // We delay the memory edge update since we need old one in
1296 // MergeMem code below when instances memory slices are separated.
1297 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;)
1298 assert(pn == NULL || pn == n, "wrong node");
1299 set_map(n->_idx, mem);
1300 ptnode_adr(n->_idx)->_node = n;
1301 }
1302 if (n->is_Load()) {
1303 continue; // don't push users
1304 } else if (n->is_LoadStore()) {
1305 // get the memory projection
1306 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1307 Node *use = n->fast_out(i);
1308 if (use->Opcode() == Op_SCMemProj) {
1309 n = use;
1310 break;
1311 }
1312 }
1313 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1314 }
1315 }
1316 // push user on appropriate worklist
1317 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1318 Node *use = n->fast_out(i);
1319 if (use->is_Phi() || use->is_ClearArray()) {
1320 memnode_worklist.append_if_missing(use);
1321 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1322 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
1323 continue;
1324 memnode_worklist.append_if_missing(use);
1325 } else if (use->is_MemBar()) {
1326 memnode_worklist.append_if_missing(use);
1327 #ifdef ASSERT
1328 } else if(use->is_Mem()) {
1329 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
1330 } else if (use->is_MergeMem()) {
1331 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1332 } else {
1333 uint op = use->Opcode();
1334 if (!(op == Op_StoreCM ||
1335 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
1336 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
1337 op == Op_AryEq || op == Op_StrComp ||
1338 op == Op_StrEquals || op == Op_StrIndexOf)) {
1339 n->dump();
1340 use->dump();
1341 assert(false, "EA: missing memory path");
1342 }
1343 #endif
1344 }
1345 }
1346 }
1348 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1349 // Walk each memory slice moving the first node encountered of each
1350 // instance type to the the input corresponding to its alias index.
1351 uint length = _mergemem_worklist.length();
1352 for( uint next = 0; next < length; ++next ) {
1353 MergeMemNode* nmm = _mergemem_worklist.at(next);
1354 assert(!visited.test_set(nmm->_idx), "should not be visited before");
1355 // Note: we don't want to use MergeMemStream here because we only want to
1356 // scan inputs which exist at the start, not ones we add during processing.
1357 // Note 2: MergeMem may already contains instance memory slices added
1358 // during find_inst_mem() call when memory nodes were processed above.
1359 igvn->hash_delete(nmm);
1360 uint nslices = nmm->req();
1361 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1362 Node* mem = nmm->in(i);
1363 Node* cur = NULL;
1364 if (mem == NULL || mem->is_top())
1365 continue;
1366 // First, update mergemem by moving memory nodes to corresponding slices
1367 // if their type became more precise since this mergemem was created.
1368 while (mem->is_Mem()) {
1369 const Type *at = igvn->type(mem->in(MemNode::Address));
1370 if (at != Type::TOP) {
1371 assert (at->isa_ptr() != NULL, "pointer type required.");
1372 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1373 if (idx == i) {
1374 if (cur == NULL)
1375 cur = mem;
1376 } else {
1377 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1378 nmm->set_memory_at(idx, mem);
1379 }
1380 }
1381 }
1382 mem = mem->in(MemNode::Memory);
1383 }
1384 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1385 // Find any instance of the current type if we haven't encountered
1386 // already a memory slice of the instance along the memory chain.
1387 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1388 if((uint)_compile->get_general_index(ni) == i) {
1389 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1390 if (nmm->is_empty_memory(m)) {
1391 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1392 if (_compile->failing()) {
1393 return;
1394 }
1395 nmm->set_memory_at(ni, result);
1396 }
1397 }
1398 }
1399 }
1400 // Find the rest of instances values
1401 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1402 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
1403 Node* result = step_through_mergemem(nmm, ni, tinst);
1404 if (result == nmm->base_memory()) {
1405 // Didn't find instance memory, search through general slice recursively.
1406 result = nmm->memory_at(_compile->get_general_index(ni));
1407 result = find_inst_mem(result, ni, orig_phis, igvn);
1408 if (_compile->failing()) {
1409 return;
1410 }
1411 nmm->set_memory_at(ni, result);
1412 }
1413 }
1414 igvn->hash_insert(nmm);
1415 record_for_optimizer(nmm);
1416 }
1418 // Phase 4: Update the inputs of non-instance memory Phis and
1419 // the Memory input of memnodes
1420 // First update the inputs of any non-instance Phi's from
1421 // which we split out an instance Phi. Note we don't have
1422 // to recursively process Phi's encounted on the input memory
1423 // chains as is done in split_memory_phi() since they will
1424 // also be processed here.
1425 for (int j = 0; j < orig_phis.length(); j++) {
1426 PhiNode *phi = orig_phis.at(j);
1427 int alias_idx = _compile->get_alias_index(phi->adr_type());
1428 igvn->hash_delete(phi);
1429 for (uint i = 1; i < phi->req(); i++) {
1430 Node *mem = phi->in(i);
1431 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1432 if (_compile->failing()) {
1433 return;
1434 }
1435 if (mem != new_mem) {
1436 phi->set_req(i, new_mem);
1437 }
1438 }
1439 igvn->hash_insert(phi);
1440 record_for_optimizer(phi);
1441 }
1443 // Update the memory inputs of MemNodes with the value we computed
1444 // in Phase 2 and move stores memory users to corresponding memory slices.
1446 // Disable memory split verification code until the fix for 6984348.
1447 // Currently it produces false negative results since it does not cover all cases.
1448 #if 0 // ifdef ASSERT
1449 visited.Reset();
1450 Node_Stack old_mems(arena, _compile->unique() >> 2);
1451 #endif
1452 for (uint i = 0; i < nodes_size(); i++) {
1453 Node *nmem = get_map(i);
1454 if (nmem != NULL) {
1455 Node *n = ptnode_adr(i)->_node;
1456 assert(n != NULL, "sanity");
1457 if (n->is_Mem()) {
1458 #if 0 // ifdef ASSERT
1459 Node* old_mem = n->in(MemNode::Memory);
1460 if (!visited.test_set(old_mem->_idx)) {
1461 old_mems.push(old_mem, old_mem->outcnt());
1462 }
1463 #endif
1464 assert(n->in(MemNode::Memory) != nmem, "sanity");
1465 if (!n->is_Load()) {
1466 // Move memory users of a store first.
1467 move_inst_mem(n, orig_phis, igvn);
1468 }
1469 // Now update memory input
1470 igvn->hash_delete(n);
1471 n->set_req(MemNode::Memory, nmem);
1472 igvn->hash_insert(n);
1473 record_for_optimizer(n);
1474 } else {
1475 assert(n->is_Allocate() || n->is_CheckCastPP() ||
1476 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
1477 }
1478 }
1479 }
1480 #if 0 // ifdef ASSERT
1481 // Verify that memory was split correctly
1482 while (old_mems.is_nonempty()) {
1483 Node* old_mem = old_mems.node();
1484 uint old_cnt = old_mems.index();
1485 old_mems.pop();
1486 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
1487 }
1488 #endif
1489 }
1491 bool ConnectionGraph::has_candidates(Compile *C) {
1492 // EA brings benefits only when the code has allocations and/or locks which
1493 // are represented by ideal Macro nodes.
1494 int cnt = C->macro_count();
1495 for( int i=0; i < cnt; i++ ) {
1496 Node *n = C->macro_node(i);
1497 if ( n->is_Allocate() )
1498 return true;
1499 if( n->is_Lock() ) {
1500 Node* obj = n->as_Lock()->obj_node()->uncast();
1501 if( !(obj->is_Parm() || obj->is_Con()) )
1502 return true;
1503 }
1504 }
1505 return false;
1506 }
1508 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
1509 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
1510 // to create space for them in ConnectionGraph::_nodes[].
1511 Node* oop_null = igvn->zerocon(T_OBJECT);
1512 Node* noop_null = igvn->zerocon(T_NARROWOOP);
1514 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
1515 // Perform escape analysis
1516 if (congraph->compute_escape()) {
1517 // There are non escaping objects.
1518 C->set_congraph(congraph);
1519 }
1521 // Cleanup.
1522 if (oop_null->outcnt() == 0)
1523 igvn->hash_delete(oop_null);
1524 if (noop_null->outcnt() == 0)
1525 igvn->hash_delete(noop_null);
1526 }
1528 bool ConnectionGraph::compute_escape() {
1529 Compile* C = _compile;
1531 // 1. Populate Connection Graph (CG) with Ideal nodes.
1533 Unique_Node_List worklist_init;
1534 worklist_init.map(C->unique(), NULL); // preallocate space
1536 // Initialize worklist
1537 if (C->root() != NULL) {
1538 worklist_init.push(C->root());
1539 }
1541 GrowableArray<int> cg_worklist;
1542 PhaseGVN* igvn = _igvn;
1543 bool has_allocations = false;
1545 // Push all useful nodes onto CG list and set their type.
1546 for( uint next = 0; next < worklist_init.size(); ++next ) {
1547 Node* n = worklist_init.at(next);
1548 record_for_escape_analysis(n, igvn);
1549 // Only allocations and java static calls results are checked
1550 // for an escape status. See process_call_result() below.
1551 if (n->is_Allocate() || n->is_CallStaticJava() &&
1552 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1553 has_allocations = true;
1554 }
1555 if(n->is_AddP()) {
1556 // Collect address nodes. Use them during stage 3 below
1557 // to build initial connection graph field edges.
1558 cg_worklist.append(n->_idx);
1559 } else if (n->is_MergeMem()) {
1560 // Collect all MergeMem nodes to add memory slices for
1561 // scalar replaceable objects in split_unique_types().
1562 _mergemem_worklist.append(n->as_MergeMem());
1563 }
1564 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1565 Node* m = n->fast_out(i); // Get user
1566 worklist_init.push(m);
1567 }
1568 }
1570 if (!has_allocations) {
1571 _collecting = false;
1572 return false; // Nothing to do.
1573 }
1575 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1576 uint delayed_size = _delayed_worklist.size();
1577 for( uint next = 0; next < delayed_size; ++next ) {
1578 Node* n = _delayed_worklist.at(next);
1579 build_connection_graph(n, igvn);
1580 }
1582 // 3. Pass to create initial fields edges (JavaObject -F-> AddP)
1583 // to reduce number of iterations during stage 4 below.
1584 uint cg_length = cg_worklist.length();
1585 for( uint next = 0; next < cg_length; ++next ) {
1586 int ni = cg_worklist.at(next);
1587 Node* n = ptnode_adr(ni)->_node;
1588 Node* base = get_addp_base(n);
1589 if (base->is_Proj())
1590 base = base->in(0);
1591 PointsToNode::NodeType nt = ptnode_adr(base->_idx)->node_type();
1592 if (nt == PointsToNode::JavaObject) {
1593 build_connection_graph(n, igvn);
1594 }
1595 }
1597 cg_worklist.clear();
1598 cg_worklist.append(_phantom_object);
1599 GrowableArray<uint> worklist;
1601 // 4. Build Connection Graph which need
1602 // to walk the connection graph.
1603 _progress = false;
1604 for (uint ni = 0; ni < nodes_size(); ni++) {
1605 PointsToNode* ptn = ptnode_adr(ni);
1606 Node *n = ptn->_node;
1607 if (n != NULL) { // Call, AddP, LoadP, StoreP
1608 build_connection_graph(n, igvn);
1609 if (ptn->node_type() != PointsToNode::UnknownType)
1610 cg_worklist.append(n->_idx); // Collect CG nodes
1611 if (!_processed.test(n->_idx))
1612 worklist.append(n->_idx); // Collect C/A/L/S nodes
1613 }
1614 }
1616 // After IGVN user nodes may have smaller _idx than
1617 // their inputs so they will be processed first in
1618 // previous loop. Because of that not all Graph
1619 // edges will be created. Walk over interesting
1620 // nodes again until no new edges are created.
1621 //
1622 // Normally only 1-3 passes needed to build
1623 // Connection Graph depending on graph complexity.
1624 // Observed 8 passes in jvm2008 compiler.compiler.
1625 // Set limit to 20 to catch situation when something
1626 // did go wrong and recompile the method without EA.
1628 #define CG_BUILD_ITER_LIMIT 20
1630 uint length = worklist.length();
1631 int iterations = 0;
1632 while(_progress && (iterations++ < CG_BUILD_ITER_LIMIT)) {
1633 _progress = false;
1634 for( uint next = 0; next < length; ++next ) {
1635 int ni = worklist.at(next);
1636 PointsToNode* ptn = ptnode_adr(ni);
1637 Node* n = ptn->_node;
1638 assert(n != NULL, "should be known node");
1639 build_connection_graph(n, igvn);
1640 }
1641 }
1642 if (iterations >= CG_BUILD_ITER_LIMIT) {
1643 assert(iterations < CG_BUILD_ITER_LIMIT,
1644 err_msg("infinite EA connection graph build with %d nodes and worklist size %d",
1645 nodes_size(), length));
1646 // Possible infinite build_connection_graph loop,
1647 // retry compilation without escape analysis.
1648 C->record_failure(C2Compiler::retry_no_escape_analysis());
1649 _collecting = false;
1650 return false;
1651 }
1652 #undef CG_BUILD_ITER_LIMIT
1654 Arena* arena = Thread::current()->resource_area();
1655 VectorSet visited(arena);
1656 worklist.clear();
1658 // 5. Remove deferred edges from the graph and adjust
1659 // escape state of nonescaping objects.
1660 cg_length = cg_worklist.length();
1661 for( uint next = 0; next < cg_length; ++next ) {
1662 int ni = cg_worklist.at(next);
1663 PointsToNode* ptn = ptnode_adr(ni);
1664 PointsToNode::NodeType nt = ptn->node_type();
1665 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1666 remove_deferred(ni, &worklist, &visited);
1667 Node *n = ptn->_node;
1668 if (n->is_AddP()) {
1669 // Search for objects which are not scalar replaceable
1670 // and adjust their escape state.
1671 adjust_escape_state(ni, igvn);
1672 }
1673 }
1674 }
1676 // 6. Propagate escape states.
1677 worklist.clear();
1678 bool has_non_escaping_obj = false;
1680 // push all GlobalEscape nodes on the worklist
1681 for( uint next = 0; next < cg_length; ++next ) {
1682 int nk = cg_worklist.at(next);
1683 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape)
1684 worklist.push(nk);
1685 }
1686 // mark all nodes reachable from GlobalEscape nodes
1687 while(worklist.length() > 0) {
1688 PointsToNode* ptn = ptnode_adr(worklist.pop());
1689 uint e_cnt = ptn->edge_count();
1690 for (uint ei = 0; ei < e_cnt; ei++) {
1691 uint npi = ptn->edge_target(ei);
1692 PointsToNode *np = ptnode_adr(npi);
1693 if (np->escape_state() < PointsToNode::GlobalEscape) {
1694 set_escape_state(npi, PointsToNode::GlobalEscape);
1695 worklist.push(npi);
1696 }
1697 }
1698 }
1700 // push all ArgEscape nodes on the worklist
1701 for( uint next = 0; next < cg_length; ++next ) {
1702 int nk = cg_worklist.at(next);
1703 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape)
1704 worklist.push(nk);
1705 }
1706 // mark all nodes reachable from ArgEscape nodes
1707 while(worklist.length() > 0) {
1708 PointsToNode* ptn = ptnode_adr(worklist.pop());
1709 if (ptn->node_type() == PointsToNode::JavaObject)
1710 has_non_escaping_obj = true; // Non GlobalEscape
1711 uint e_cnt = ptn->edge_count();
1712 for (uint ei = 0; ei < e_cnt; ei++) {
1713 uint npi = ptn->edge_target(ei);
1714 PointsToNode *np = ptnode_adr(npi);
1715 if (np->escape_state() < PointsToNode::ArgEscape) {
1716 set_escape_state(npi, PointsToNode::ArgEscape);
1717 worklist.push(npi);
1718 }
1719 }
1720 }
1722 GrowableArray<Node*> alloc_worklist;
1724 // push all NoEscape nodes on the worklist
1725 for( uint next = 0; next < cg_length; ++next ) {
1726 int nk = cg_worklist.at(next);
1727 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape)
1728 worklist.push(nk);
1729 }
1730 // mark all nodes reachable from NoEscape nodes
1731 while(worklist.length() > 0) {
1732 uint nk = worklist.pop();
1733 PointsToNode* ptn = ptnode_adr(nk);
1734 if (ptn->node_type() == PointsToNode::JavaObject &&
1735 !(nk == _noop_null || nk == _oop_null))
1736 has_non_escaping_obj = true; // Non Escape
1737 Node* n = ptn->_node;
1738 if (n->is_Allocate() && ptn->_scalar_replaceable ) {
1739 // Push scalar replaceable allocations on alloc_worklist
1740 // for processing in split_unique_types().
1741 alloc_worklist.append(n);
1742 }
1743 uint e_cnt = ptn->edge_count();
1744 for (uint ei = 0; ei < e_cnt; ei++) {
1745 uint npi = ptn->edge_target(ei);
1746 PointsToNode *np = ptnode_adr(npi);
1747 if (np->escape_state() < PointsToNode::NoEscape) {
1748 set_escape_state(npi, PointsToNode::NoEscape);
1749 worklist.push(npi);
1750 }
1751 }
1752 }
1754 _collecting = false;
1755 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1757 assert(ptnode_adr(_oop_null)->escape_state() == PointsToNode::NoEscape, "sanity");
1758 if (UseCompressedOops) {
1759 assert(ptnode_adr(_noop_null)->escape_state() == PointsToNode::NoEscape, "sanity");
1760 }
1762 if (EliminateLocks) {
1763 // Mark locks before changing ideal graph.
1764 int cnt = C->macro_count();
1765 for( int i=0; i < cnt; i++ ) {
1766 Node *n = C->macro_node(i);
1767 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1768 AbstractLockNode* alock = n->as_AbstractLock();
1769 if (!alock->is_eliminated()) {
1770 PointsToNode::EscapeState es = escape_state(alock->obj_node());
1771 assert(es != PointsToNode::UnknownEscape, "should know");
1772 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1773 // Mark it eliminated
1774 alock->set_eliminated();
1775 }
1776 }
1777 }
1778 }
1779 }
1781 #ifndef PRODUCT
1782 if (PrintEscapeAnalysis) {
1783 dump(); // Dump ConnectionGraph
1784 }
1785 #endif
1787 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0;
1788 if ( has_scalar_replaceable_candidates &&
1789 C->AliasLevel() >= 3 && EliminateAllocations ) {
1791 // Now use the escape information to create unique types for
1792 // scalar replaceable objects.
1793 split_unique_types(alloc_worklist);
1795 if (C->failing()) return false;
1797 C->print_method("After Escape Analysis", 2);
1799 #ifdef ASSERT
1800 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1801 tty->print("=== No allocations eliminated for ");
1802 C->method()->print_short_name();
1803 if(!EliminateAllocations) {
1804 tty->print(" since EliminateAllocations is off ===");
1805 } else if(!has_scalar_replaceable_candidates) {
1806 tty->print(" since there are no scalar replaceable candidates ===");
1807 } else if(C->AliasLevel() < 3) {
1808 tty->print(" since AliasLevel < 3 ===");
1809 }
1810 tty->cr();
1811 #endif
1812 }
1813 return has_non_escaping_obj;
1814 }
1816 // Adjust escape state after Connection Graph is built.
1817 void ConnectionGraph::adjust_escape_state(int nidx, PhaseTransform* phase) {
1818 PointsToNode* ptn = ptnode_adr(nidx);
1819 Node* n = ptn->_node;
1820 assert(n->is_AddP(), "Should be called for AddP nodes only");
1821 // Search for objects which are not scalar replaceable.
1822 // Mark their escape state as ArgEscape to propagate the state
1823 // to referenced objects.
1824 // Note: currently there are no difference in compiler optimizations
1825 // for ArgEscape objects and NoEscape objects which are not
1826 // scalar replaceable.
1828 Compile* C = _compile;
1830 int offset = ptn->offset();
1831 Node* base = get_addp_base(n);
1832 VectorSet* ptset = PointsTo(base);
1833 int ptset_size = ptset->Size();
1835 // Check if a oop field's initializing value is recorded and add
1836 // a corresponding NULL field's value if it is not recorded.
1837 // Connection Graph does not record a default initialization by NULL
1838 // captured by Initialize node.
1839 //
1840 // Note: it will disable scalar replacement in some cases:
1841 //
1842 // Point p[] = new Point[1];
1843 // p[0] = new Point(); // Will be not scalar replaced
1844 //
1845 // but it will save us from incorrect optimizations in next cases:
1846 //
1847 // Point p[] = new Point[1];
1848 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1849 //
1850 // Do a simple control flow analysis to distinguish above cases.
1851 //
1852 if (offset != Type::OffsetBot && ptset_size == 1) {
1853 uint elem = ptset->getelem(); // Allocation node's index
1854 // It does not matter if it is not Allocation node since
1855 // only non-escaping allocations are scalar replaced.
1856 if (ptnode_adr(elem)->_node->is_Allocate() &&
1857 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) {
1858 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate();
1859 InitializeNode* ini = alloc->initialization();
1861 // Check only oop fields.
1862 const Type* adr_type = n->as_AddP()->bottom_type();
1863 BasicType basic_field_type = T_INT;
1864 if (adr_type->isa_instptr()) {
1865 ciField* field = C->alias_type(adr_type->isa_instptr())->field();
1866 if (field != NULL) {
1867 basic_field_type = field->layout_type();
1868 } else {
1869 // Ignore non field load (for example, klass load)
1870 }
1871 } else if (adr_type->isa_aryptr()) {
1872 const Type* elemtype = adr_type->isa_aryptr()->elem();
1873 basic_field_type = elemtype->array_element_basic_type();
1874 } else {
1875 // Raw pointers are used for initializing stores so skip it.
1876 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1877 (base->in(0) == alloc),"unexpected pointer type");
1878 }
1879 if (basic_field_type == T_OBJECT ||
1880 basic_field_type == T_NARROWOOP ||
1881 basic_field_type == T_ARRAY) {
1882 Node* value = NULL;
1883 if (ini != NULL) {
1884 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
1885 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase);
1886 if (store != NULL && store->is_Store()) {
1887 value = store->in(MemNode::ValueIn);
1888 } else if (ptn->edge_count() > 0) { // Are there oop stores?
1889 // Check for a store which follows allocation without branches.
1890 // For example, a volatile field store is not collected
1891 // by Initialize node. TODO: it would be nice to use idom() here.
1892 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1893 store = n->fast_out(i);
1894 if (store->is_Store() && store->in(0) != NULL) {
1895 Node* ctrl = store->in(0);
1896 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL ||
1897 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() ||
1898 ctrl->is_IfTrue() || ctrl->is_IfFalse())) {
1899 ctrl = ctrl->in(0);
1900 }
1901 if (ctrl == ini || ctrl == alloc) {
1902 value = store->in(MemNode::ValueIn);
1903 break;
1904 }
1905 }
1906 }
1907 }
1908 }
1909 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
1910 // A field's initializing value was not recorded. Add NULL.
1911 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1912 add_pointsto_edge(nidx, null_idx);
1913 }
1914 }
1915 }
1916 }
1918 // An object is not scalar replaceable if the field which may point
1919 // to it has unknown offset (unknown element of an array of objects).
1920 //
1921 if (offset == Type::OffsetBot) {
1922 uint e_cnt = ptn->edge_count();
1923 for (uint ei = 0; ei < e_cnt; ei++) {
1924 uint npi = ptn->edge_target(ei);
1925 set_escape_state(npi, PointsToNode::ArgEscape);
1926 ptnode_adr(npi)->_scalar_replaceable = false;
1927 }
1928 }
1930 // Currently an object is not scalar replaceable if a LoadStore node
1931 // access its field since the field value is unknown after it.
1932 //
1933 bool has_LoadStore = false;
1934 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1935 Node *use = n->fast_out(i);
1936 if (use->is_LoadStore()) {
1937 has_LoadStore = true;
1938 break;
1939 }
1940 }
1941 // An object is not scalar replaceable if the address points
1942 // to unknown field (unknown element for arrays, offset is OffsetBot).
1943 //
1944 // Or the address may point to more then one object. This may produce
1945 // the false positive result (set scalar_replaceable to false)
1946 // since the flow-insensitive escape analysis can't separate
1947 // the case when stores overwrite the field's value from the case
1948 // when stores happened on different control branches.
1949 //
1950 if (ptset_size > 1 || ptset_size != 0 &&
1951 (has_LoadStore || offset == Type::OffsetBot)) {
1952 for( VectorSetI j(ptset); j.test(); ++j ) {
1953 set_escape_state(j.elem, PointsToNode::ArgEscape);
1954 ptnode_adr(j.elem)->_scalar_replaceable = false;
1955 }
1956 }
1957 }
1959 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
1961 switch (call->Opcode()) {
1962 #ifdef ASSERT
1963 case Op_Allocate:
1964 case Op_AllocateArray:
1965 case Op_Lock:
1966 case Op_Unlock:
1967 assert(false, "should be done already");
1968 break;
1969 #endif
1970 case Op_CallLeaf:
1971 case Op_CallLeafNoFP:
1972 {
1973 // Stub calls, objects do not escape but they are not scale replaceable.
1974 // Adjust escape state for outgoing arguments.
1975 const TypeTuple * d = call->tf()->domain();
1976 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1977 const Type* at = d->field_at(i);
1978 Node *arg = call->in(i)->uncast();
1979 const Type *aat = phase->type(arg);
1980 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() &&
1981 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) {
1983 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
1984 aat->isa_ptr() != NULL, "expecting an Ptr");
1985 #ifdef ASSERT
1986 if (!(call->Opcode() == Op_CallLeafNoFP &&
1987 call->as_CallLeaf()->_name != NULL &&
1988 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) ||
1989 call->as_CallLeaf()->_name != NULL &&
1990 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
1991 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ))
1992 ) {
1993 call->dump();
1994 assert(false, "EA: unexpected CallLeaf");
1995 }
1996 #endif
1997 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1998 if (arg->is_AddP()) {
1999 //
2000 // The inline_native_clone() case when the arraycopy stub is called
2001 // after the allocation before Initialize and CheckCastPP nodes.
2002 //
2003 // Set AddP's base (Allocate) as not scalar replaceable since
2004 // pointer to the base (with offset) is passed as argument.
2005 //
2006 arg = get_addp_base(arg);
2007 }
2008 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2009 uint pt = j.elem;
2010 set_escape_state(pt, PointsToNode::ArgEscape);
2011 }
2012 }
2013 }
2014 break;
2015 }
2017 case Op_CallStaticJava:
2018 // For a static call, we know exactly what method is being called.
2019 // Use bytecode estimator to record the call's escape affects
2020 {
2021 ciMethod *meth = call->as_CallJava()->method();
2022 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
2023 // fall-through if not a Java method or no analyzer information
2024 if (call_analyzer != NULL) {
2025 const TypeTuple * d = call->tf()->domain();
2026 bool copy_dependencies = false;
2027 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2028 const Type* at = d->field_at(i);
2029 int k = i - TypeFunc::Parms;
2030 Node *arg = call->in(i)->uncast();
2032 if (at->isa_oopptr() != NULL &&
2033 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) {
2035 bool global_escapes = false;
2036 bool fields_escapes = false;
2037 if (!call_analyzer->is_arg_stack(k)) {
2038 // The argument global escapes, mark everything it could point to
2039 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2040 global_escapes = true;
2041 } else {
2042 if (!call_analyzer->is_arg_local(k)) {
2043 // The argument itself doesn't escape, but any fields might
2044 fields_escapes = true;
2045 }
2046 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
2047 copy_dependencies = true;
2048 }
2050 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2051 uint pt = j.elem;
2052 if (global_escapes) {
2053 //The argument global escapes, mark everything it could point to
2054 set_escape_state(pt, PointsToNode::GlobalEscape);
2055 } else {
2056 if (fields_escapes) {
2057 // The argument itself doesn't escape, but any fields might
2058 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
2059 }
2060 set_escape_state(pt, PointsToNode::ArgEscape);
2061 }
2062 }
2063 }
2064 }
2065 if (copy_dependencies)
2066 call_analyzer->copy_dependencies(_compile->dependencies());
2067 break;
2068 }
2069 }
2071 default:
2072 // Fall-through here if not a Java method or no analyzer information
2073 // or some other type of call, assume the worst case: all arguments
2074 // globally escape.
2075 {
2076 // adjust escape state for outgoing arguments
2077 const TypeTuple * d = call->tf()->domain();
2078 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2079 const Type* at = d->field_at(i);
2080 if (at->isa_oopptr() != NULL) {
2081 Node *arg = call->in(i)->uncast();
2082 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2083 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2084 uint pt = j.elem;
2085 set_escape_state(pt, PointsToNode::GlobalEscape);
2086 }
2087 }
2088 }
2089 }
2090 }
2091 }
2092 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
2093 CallNode *call = resproj->in(0)->as_Call();
2094 uint call_idx = call->_idx;
2095 uint resproj_idx = resproj->_idx;
2097 switch (call->Opcode()) {
2098 case Op_Allocate:
2099 {
2100 Node *k = call->in(AllocateNode::KlassNode);
2101 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2102 assert(kt != NULL, "TypeKlassPtr required.");
2103 ciKlass* cik = kt->klass();
2105 PointsToNode::EscapeState es;
2106 uint edge_to;
2107 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
2108 !cik->is_instance_klass() || // StressReflectiveCode
2109 cik->as_instance_klass()->has_finalizer()) {
2110 es = PointsToNode::GlobalEscape;
2111 edge_to = _phantom_object; // Could not be worse
2112 } else {
2113 es = PointsToNode::NoEscape;
2114 edge_to = call_idx;
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_AllocateArray:
2123 {
2125 Node *k = call->in(AllocateNode::KlassNode);
2126 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2127 assert(kt != NULL, "TypeKlassPtr required.");
2128 ciKlass* cik = kt->klass();
2130 PointsToNode::EscapeState es;
2131 uint edge_to;
2132 if (!cik->is_array_klass()) { // StressReflectiveCode
2133 es = PointsToNode::GlobalEscape;
2134 edge_to = _phantom_object;
2135 } else {
2136 es = PointsToNode::NoEscape;
2137 edge_to = call_idx;
2138 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2139 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
2140 // Not scalar replaceable if the length is not constant or too big.
2141 ptnode_adr(call_idx)->_scalar_replaceable = false;
2142 }
2143 }
2144 set_escape_state(call_idx, es);
2145 add_pointsto_edge(resproj_idx, edge_to);
2146 _processed.set(resproj_idx);
2147 break;
2148 }
2150 case Op_CallStaticJava:
2151 // For a static call, we know exactly what method is being called.
2152 // Use bytecode estimator to record whether the call's return value escapes
2153 {
2154 bool done = true;
2155 const TypeTuple *r = call->tf()->range();
2156 const Type* ret_type = NULL;
2158 if (r->cnt() > TypeFunc::Parms)
2159 ret_type = r->field_at(TypeFunc::Parms);
2161 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2162 // _multianewarray functions return a TypeRawPtr.
2163 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
2164 _processed.set(resproj_idx);
2165 break; // doesn't return a pointer type
2166 }
2167 ciMethod *meth = call->as_CallJava()->method();
2168 const TypeTuple * d = call->tf()->domain();
2169 if (meth == NULL) {
2170 // not a Java method, assume global escape
2171 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2172 add_pointsto_edge(resproj_idx, _phantom_object);
2173 } else {
2174 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
2175 bool copy_dependencies = false;
2177 if (call_analyzer->is_return_allocated()) {
2178 // Returns a newly allocated unescaped object, simply
2179 // update dependency information.
2180 // Mark it as NoEscape so that objects referenced by
2181 // it's fields will be marked as NoEscape at least.
2182 set_escape_state(call_idx, PointsToNode::NoEscape);
2183 add_pointsto_edge(resproj_idx, call_idx);
2184 copy_dependencies = true;
2185 } else if (call_analyzer->is_return_local()) {
2186 // determine whether any arguments are returned
2187 set_escape_state(call_idx, PointsToNode::NoEscape);
2188 bool ret_arg = false;
2189 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2190 const Type* at = d->field_at(i);
2192 if (at->isa_oopptr() != NULL) {
2193 Node *arg = call->in(i)->uncast();
2195 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2196 ret_arg = true;
2197 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2198 if (arg_esp->node_type() == PointsToNode::UnknownType)
2199 done = false;
2200 else if (arg_esp->node_type() == PointsToNode::JavaObject)
2201 add_pointsto_edge(resproj_idx, arg->_idx);
2202 else
2203 add_deferred_edge(resproj_idx, arg->_idx);
2204 arg_esp->_hidden_alias = true;
2205 }
2206 }
2207 }
2208 if (done && !ret_arg) {
2209 // Returns unknown object.
2210 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2211 add_pointsto_edge(resproj_idx, _phantom_object);
2212 }
2213 copy_dependencies = true;
2214 } else {
2215 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2216 add_pointsto_edge(resproj_idx, _phantom_object);
2217 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2218 const Type* at = d->field_at(i);
2219 if (at->isa_oopptr() != NULL) {
2220 Node *arg = call->in(i)->uncast();
2221 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2222 arg_esp->_hidden_alias = true;
2223 }
2224 }
2225 }
2226 if (copy_dependencies)
2227 call_analyzer->copy_dependencies(_compile->dependencies());
2228 }
2229 if (done)
2230 _processed.set(resproj_idx);
2231 break;
2232 }
2234 default:
2235 // Some other type of call, assume the worst case that the
2236 // returned value, if any, globally escapes.
2237 {
2238 const TypeTuple *r = call->tf()->range();
2239 if (r->cnt() > TypeFunc::Parms) {
2240 const Type* ret_type = r->field_at(TypeFunc::Parms);
2242 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2243 // _multianewarray functions return a TypeRawPtr.
2244 if (ret_type->isa_ptr() != NULL) {
2245 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2246 add_pointsto_edge(resproj_idx, _phantom_object);
2247 }
2248 }
2249 _processed.set(resproj_idx);
2250 }
2251 }
2252 }
2254 // Populate Connection Graph with Ideal nodes and create simple
2255 // connection graph edges (do not need to check the node_type of inputs
2256 // or to call PointsTo() to walk the connection graph).
2257 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
2258 if (_processed.test(n->_idx))
2259 return; // No need to redefine node's state.
2261 if (n->is_Call()) {
2262 // Arguments to allocation and locking don't escape.
2263 if (n->is_Allocate()) {
2264 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
2265 record_for_optimizer(n);
2266 } else if (n->is_Lock() || n->is_Unlock()) {
2267 // Put Lock and Unlock nodes on IGVN worklist to process them during
2268 // the first IGVN optimization when escape information is still available.
2269 record_for_optimizer(n);
2270 _processed.set(n->_idx);
2271 } else {
2272 // Don't mark as processed since call's arguments have to be processed.
2273 PointsToNode::NodeType nt = PointsToNode::UnknownType;
2274 PointsToNode::EscapeState es = PointsToNode::UnknownEscape;
2276 // Check if a call returns an object.
2277 const TypeTuple *r = n->as_Call()->tf()->range();
2278 if (r->cnt() > TypeFunc::Parms &&
2279 r->field_at(TypeFunc::Parms)->isa_ptr() &&
2280 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
2281 nt = PointsToNode::JavaObject;
2282 if (!n->is_CallStaticJava()) {
2283 // Since the called mathod is statically unknown assume
2284 // the worst case that the returned value globally escapes.
2285 es = PointsToNode::GlobalEscape;
2286 }
2287 }
2288 add_node(n, nt, es, false);
2289 }
2290 return;
2291 }
2293 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
2294 // ThreadLocal has RawPrt type.
2295 switch (n->Opcode()) {
2296 case Op_AddP:
2297 {
2298 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
2299 break;
2300 }
2301 case Op_CastX2P:
2302 { // "Unsafe" memory access.
2303 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2304 break;
2305 }
2306 case Op_CastPP:
2307 case Op_CheckCastPP:
2308 case Op_EncodeP:
2309 case Op_DecodeN:
2310 {
2311 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2312 int ti = n->in(1)->_idx;
2313 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2314 if (nt == PointsToNode::UnknownType) {
2315 _delayed_worklist.push(n); // Process it later.
2316 break;
2317 } else if (nt == PointsToNode::JavaObject) {
2318 add_pointsto_edge(n->_idx, ti);
2319 } else {
2320 add_deferred_edge(n->_idx, ti);
2321 }
2322 _processed.set(n->_idx);
2323 break;
2324 }
2325 case Op_ConP:
2326 {
2327 // assume all pointer constants globally escape except for null
2328 PointsToNode::EscapeState es;
2329 if (phase->type(n) == TypePtr::NULL_PTR)
2330 es = PointsToNode::NoEscape;
2331 else
2332 es = PointsToNode::GlobalEscape;
2334 add_node(n, PointsToNode::JavaObject, es, true);
2335 break;
2336 }
2337 case Op_ConN:
2338 {
2339 // assume all narrow oop constants globally escape except for null
2340 PointsToNode::EscapeState es;
2341 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
2342 es = PointsToNode::NoEscape;
2343 else
2344 es = PointsToNode::GlobalEscape;
2346 add_node(n, PointsToNode::JavaObject, es, true);
2347 break;
2348 }
2349 case Op_CreateEx:
2350 {
2351 // assume that all exception objects globally escape
2352 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2353 break;
2354 }
2355 case Op_LoadKlass:
2356 case Op_LoadNKlass:
2357 {
2358 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2359 break;
2360 }
2361 case Op_LoadP:
2362 case Op_LoadN:
2363 {
2364 const Type *t = phase->type(n);
2365 if (t->make_ptr() == NULL) {
2366 _processed.set(n->_idx);
2367 return;
2368 }
2369 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2370 break;
2371 }
2372 case Op_Parm:
2373 {
2374 _processed.set(n->_idx); // No need to redefine it state.
2375 uint con = n->as_Proj()->_con;
2376 if (con < TypeFunc::Parms)
2377 return;
2378 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2379 if (t->isa_ptr() == NULL)
2380 return;
2381 // We have to assume all input parameters globally escape
2382 // (Note: passing 'false' since _processed is already set).
2383 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2384 break;
2385 }
2386 case Op_Phi:
2387 {
2388 const Type *t = n->as_Phi()->type();
2389 if (t->make_ptr() == NULL) {
2390 // nothing to do if not an oop or narrow oop
2391 _processed.set(n->_idx);
2392 return;
2393 }
2394 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2395 uint i;
2396 for (i = 1; i < n->req() ; i++) {
2397 Node* in = n->in(i);
2398 if (in == NULL)
2399 continue; // ignore NULL
2400 in = in->uncast();
2401 if (in->is_top() || in == n)
2402 continue; // ignore top or inputs which go back this node
2403 int ti = in->_idx;
2404 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2405 if (nt == PointsToNode::UnknownType) {
2406 break;
2407 } else if (nt == PointsToNode::JavaObject) {
2408 add_pointsto_edge(n->_idx, ti);
2409 } else {
2410 add_deferred_edge(n->_idx, ti);
2411 }
2412 }
2413 if (i >= n->req())
2414 _processed.set(n->_idx);
2415 else
2416 _delayed_worklist.push(n);
2417 break;
2418 }
2419 case Op_Proj:
2420 {
2421 // we are only interested in the oop result projection from a call
2422 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2423 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2424 assert(r->cnt() > TypeFunc::Parms, "sanity");
2425 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2426 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2427 int ti = n->in(0)->_idx;
2428 // The call may not be registered yet (since not all its inputs are registered)
2429 // if this is the projection from backbranch edge of Phi.
2430 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) {
2431 process_call_result(n->as_Proj(), phase);
2432 }
2433 if (!_processed.test(n->_idx)) {
2434 // The call's result may need to be processed later if the call
2435 // returns it's argument and the argument is not processed yet.
2436 _delayed_worklist.push(n);
2437 }
2438 break;
2439 }
2440 }
2441 _processed.set(n->_idx);
2442 break;
2443 }
2444 case Op_Return:
2445 {
2446 if( n->req() > TypeFunc::Parms &&
2447 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2448 // Treat Return value as LocalVar with GlobalEscape escape state.
2449 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2450 int ti = n->in(TypeFunc::Parms)->_idx;
2451 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2452 if (nt == PointsToNode::UnknownType) {
2453 _delayed_worklist.push(n); // Process it later.
2454 break;
2455 } else if (nt == PointsToNode::JavaObject) {
2456 add_pointsto_edge(n->_idx, ti);
2457 } else {
2458 add_deferred_edge(n->_idx, ti);
2459 }
2460 }
2461 _processed.set(n->_idx);
2462 break;
2463 }
2464 case Op_StoreP:
2465 case Op_StoreN:
2466 {
2467 const Type *adr_type = phase->type(n->in(MemNode::Address));
2468 adr_type = adr_type->make_ptr();
2469 if (adr_type->isa_oopptr()) {
2470 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2471 } else {
2472 Node* adr = n->in(MemNode::Address);
2473 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2474 adr->in(AddPNode::Address)->is_Proj() &&
2475 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2476 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2477 // We are computing a raw address for a store captured
2478 // by an Initialize compute an appropriate address type.
2479 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2480 assert(offs != Type::OffsetBot, "offset must be a constant");
2481 } else {
2482 _processed.set(n->_idx);
2483 return;
2484 }
2485 }
2486 break;
2487 }
2488 case Op_StorePConditional:
2489 case Op_CompareAndSwapP:
2490 case Op_CompareAndSwapN:
2491 {
2492 const Type *adr_type = phase->type(n->in(MemNode::Address));
2493 adr_type = adr_type->make_ptr();
2494 if (adr_type->isa_oopptr()) {
2495 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2496 } else {
2497 _processed.set(n->_idx);
2498 return;
2499 }
2500 break;
2501 }
2502 case Op_AryEq:
2503 case Op_StrComp:
2504 case Op_StrEquals:
2505 case Op_StrIndexOf:
2506 {
2507 // char[] arrays passed to string intrinsics are not scalar replaceable.
2508 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2509 break;
2510 }
2511 case Op_ThreadLocal:
2512 {
2513 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2514 break;
2515 }
2516 default:
2517 ;
2518 // nothing to do
2519 }
2520 return;
2521 }
2523 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2524 uint n_idx = n->_idx;
2525 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered");
2527 // Don't set processed bit for AddP, LoadP, StoreP since
2528 // they may need more then one pass to process.
2529 // Also don't mark as processed Call nodes since their
2530 // arguments may need more then one pass to process.
2531 if (_processed.test(n_idx))
2532 return; // No need to redefine node's state.
2534 if (n->is_Call()) {
2535 CallNode *call = n->as_Call();
2536 process_call_arguments(call, phase);
2537 return;
2538 }
2540 switch (n->Opcode()) {
2541 case Op_AddP:
2542 {
2543 Node *base = get_addp_base(n);
2544 // Create a field edge to this node from everything base could point to.
2545 for( VectorSetI i(PointsTo(base)); i.test(); ++i ) {
2546 uint pt = i.elem;
2547 add_field_edge(pt, n_idx, address_offset(n, phase));
2548 }
2549 break;
2550 }
2551 case Op_CastX2P:
2552 {
2553 assert(false, "Op_CastX2P");
2554 break;
2555 }
2556 case Op_CastPP:
2557 case Op_CheckCastPP:
2558 case Op_EncodeP:
2559 case Op_DecodeN:
2560 {
2561 int ti = n->in(1)->_idx;
2562 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered");
2563 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2564 add_pointsto_edge(n_idx, ti);
2565 } else {
2566 add_deferred_edge(n_idx, ti);
2567 }
2568 _processed.set(n_idx);
2569 break;
2570 }
2571 case Op_ConP:
2572 {
2573 assert(false, "Op_ConP");
2574 break;
2575 }
2576 case Op_ConN:
2577 {
2578 assert(false, "Op_ConN");
2579 break;
2580 }
2581 case Op_CreateEx:
2582 {
2583 assert(false, "Op_CreateEx");
2584 break;
2585 }
2586 case Op_LoadKlass:
2587 case Op_LoadNKlass:
2588 {
2589 assert(false, "Op_LoadKlass");
2590 break;
2591 }
2592 case Op_LoadP:
2593 case Op_LoadN:
2594 {
2595 const Type *t = phase->type(n);
2596 #ifdef ASSERT
2597 if (t->make_ptr() == NULL)
2598 assert(false, "Op_LoadP");
2599 #endif
2601 Node* adr = n->in(MemNode::Address)->uncast();
2602 Node* adr_base;
2603 if (adr->is_AddP()) {
2604 adr_base = get_addp_base(adr);
2605 } else {
2606 adr_base = adr;
2607 }
2609 // For everything "adr_base" could point to, create a deferred edge from
2610 // this node to each field with the same offset.
2611 int offset = address_offset(adr, phase);
2612 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2613 uint pt = i.elem;
2614 add_deferred_edge_to_fields(n_idx, pt, offset);
2615 }
2616 break;
2617 }
2618 case Op_Parm:
2619 {
2620 assert(false, "Op_Parm");
2621 break;
2622 }
2623 case Op_Phi:
2624 {
2625 #ifdef ASSERT
2626 const Type *t = n->as_Phi()->type();
2627 if (t->make_ptr() == NULL)
2628 assert(false, "Op_Phi");
2629 #endif
2630 for (uint i = 1; i < n->req() ; i++) {
2631 Node* in = n->in(i);
2632 if (in == NULL)
2633 continue; // ignore NULL
2634 in = in->uncast();
2635 if (in->is_top() || in == n)
2636 continue; // ignore top or inputs which go back this node
2637 int ti = in->_idx;
2638 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2639 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
2640 if (nt == PointsToNode::JavaObject) {
2641 add_pointsto_edge(n_idx, ti);
2642 } else {
2643 add_deferred_edge(n_idx, ti);
2644 }
2645 }
2646 _processed.set(n_idx);
2647 break;
2648 }
2649 case Op_Proj:
2650 {
2651 // we are only interested in the oop result projection from a call
2652 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2653 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType,
2654 "all nodes should be registered");
2655 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2656 assert(r->cnt() > TypeFunc::Parms, "sanity");
2657 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2658 process_call_result(n->as_Proj(), phase);
2659 assert(_processed.test(n_idx), "all call results should be processed");
2660 break;
2661 }
2662 }
2663 assert(false, "Op_Proj");
2664 break;
2665 }
2666 case Op_Return:
2667 {
2668 #ifdef ASSERT
2669 if( n->req() <= TypeFunc::Parms ||
2670 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2671 assert(false, "Op_Return");
2672 }
2673 #endif
2674 int ti = n->in(TypeFunc::Parms)->_idx;
2675 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered");
2676 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2677 add_pointsto_edge(n_idx, ti);
2678 } else {
2679 add_deferred_edge(n_idx, ti);
2680 }
2681 _processed.set(n_idx);
2682 break;
2683 }
2684 case Op_StoreP:
2685 case Op_StoreN:
2686 case Op_StorePConditional:
2687 case Op_CompareAndSwapP:
2688 case Op_CompareAndSwapN:
2689 {
2690 Node *adr = n->in(MemNode::Address);
2691 const Type *adr_type = phase->type(adr)->make_ptr();
2692 #ifdef ASSERT
2693 if (!adr_type->isa_oopptr())
2694 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
2695 #endif
2697 assert(adr->is_AddP(), "expecting an AddP");
2698 Node *adr_base = get_addp_base(adr);
2699 Node *val = n->in(MemNode::ValueIn)->uncast();
2700 // For everything "adr_base" could point to, create a deferred edge
2701 // to "val" from each field with the same offset.
2702 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2703 uint pt = i.elem;
2704 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase));
2705 }
2706 break;
2707 }
2708 case Op_AryEq:
2709 case Op_StrComp:
2710 case Op_StrEquals:
2711 case Op_StrIndexOf:
2712 {
2713 // char[] arrays passed to string intrinsic do not escape but
2714 // they are not scalar replaceable. Adjust escape state for them.
2715 // Start from in(2) edge since in(1) is memory edge.
2716 for (uint i = 2; i < n->req(); i++) {
2717 Node* adr = n->in(i)->uncast();
2718 const Type *at = phase->type(adr);
2719 if (!adr->is_top() && at->isa_ptr()) {
2720 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
2721 at->isa_ptr() != NULL, "expecting an Ptr");
2722 if (adr->is_AddP()) {
2723 adr = get_addp_base(adr);
2724 }
2725 // Mark as ArgEscape everything "adr" could point to.
2726 set_escape_state(adr->_idx, PointsToNode::ArgEscape);
2727 }
2728 }
2729 _processed.set(n_idx);
2730 break;
2731 }
2732 case Op_ThreadLocal:
2733 {
2734 assert(false, "Op_ThreadLocal");
2735 break;
2736 }
2737 default:
2738 // This method should be called only for EA specific nodes.
2739 ShouldNotReachHere();
2740 }
2741 }
2743 #ifndef PRODUCT
2744 void ConnectionGraph::dump() {
2745 bool first = true;
2747 uint size = nodes_size();
2748 for (uint ni = 0; ni < size; ni++) {
2749 PointsToNode *ptn = ptnode_adr(ni);
2750 PointsToNode::NodeType ptn_type = ptn->node_type();
2752 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
2753 continue;
2754 PointsToNode::EscapeState es = escape_state(ptn->_node);
2755 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
2756 if (first) {
2757 tty->cr();
2758 tty->print("======== Connection graph for ");
2759 _compile->method()->print_short_name();
2760 tty->cr();
2761 first = false;
2762 }
2763 tty->print("%6d ", ni);
2764 ptn->dump();
2765 // Print all locals which reference this allocation
2766 for (uint li = ni; li < size; li++) {
2767 PointsToNode *ptn_loc = ptnode_adr(li);
2768 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
2769 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
2770 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
2771 ptnode_adr(li)->dump(false);
2772 }
2773 }
2774 if (Verbose) {
2775 // Print all fields which reference this allocation
2776 for (uint i = 0; i < ptn->edge_count(); i++) {
2777 uint ei = ptn->edge_target(i);
2778 ptnode_adr(ei)->dump(false);
2779 }
2780 }
2781 tty->cr();
2782 }
2783 }
2784 }
2785 #endif