Mercurial > hg > truffle
annotate src/share/vm/opto/escape.cpp @ 244:524eca34ea76
6684714: Optimize EA Connection Graph build performance
Summary: switch on EA by default, optimize Connection Graph construction
Reviewed-by: rasbold, never
| author | kvn |
|---|---|
| date | Thu, 03 Jul 2008 18:02:47 -0700 |
| parents | 1dd146f17531 |
| children | 4a4c365f777d |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 2 * Copyright 2005-2006 Sun Microsystems, Inc. All Rights Reserved. | |
| 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
| 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
| 20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
| 21 * have any questions. | |
| 22 * | |
| 23 */ | |
| 24 | |
| 25 #include "incls/_precompiled.incl" | |
| 26 #include "incls/_escape.cpp.incl" | |
| 27 | |
| 28 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) { | |
| 29 uint v = (targIdx << EdgeShift) + ((uint) et); | |
| 30 if (_edges == NULL) { | |
| 31 Arena *a = Compile::current()->comp_arena(); | |
| 32 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0); | |
| 33 } | |
| 34 _edges->append_if_missing(v); | |
| 35 } | |
| 36 | |
| 37 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) { | |
| 38 uint v = (targIdx << EdgeShift) + ((uint) et); | |
| 39 | |
| 40 _edges->remove(v); | |
| 41 } | |
| 42 | |
| 43 #ifndef PRODUCT | |
| 77 | 44 static const char *node_type_names[] = { |
| 0 | 45 "UnknownType", |
| 46 "JavaObject", | |
| 47 "LocalVar", | |
| 48 "Field" | |
| 49 }; | |
| 50 | |
| 77 | 51 static const char *esc_names[] = { |
| 0 | 52 "UnknownEscape", |
| 65 | 53 "NoEscape", |
| 54 "ArgEscape", | |
| 55 "GlobalEscape" | |
| 0 | 56 }; |
| 57 | |
| 77 | 58 static const char *edge_type_suffix[] = { |
| 0 | 59 "?", // UnknownEdge |
| 60 "P", // PointsToEdge | |
| 61 "D", // DeferredEdge | |
| 62 "F" // FieldEdge | |
| 63 }; | |
| 64 | |
| 65 void PointsToNode::dump() const { | |
| 66 NodeType nt = node_type(); | |
| 67 EscapeState es = escape_state(); | |
| 65 | 68 tty->print("%s %s %s [[", node_type_names[(int) nt], esc_names[(int) es], _scalar_replaceable ? "" : "NSR"); |
| 0 | 69 for (uint i = 0; i < edge_count(); i++) { |
| 70 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]); | |
| 71 } | |
| 72 tty->print("]] "); | |
| 73 if (_node == NULL) | |
| 74 tty->print_cr("<null>"); | |
| 75 else | |
| 76 _node->dump(); | |
| 77 } | |
| 78 #endif | |
| 79 | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
80 ConnectionGraph::ConnectionGraph(Compile * C) : |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
81 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()), |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
82 _processed(C->comp_arena()), |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
83 _collecting(true), |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
84 _compile(C), |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
85 _node_map(C->comp_arena()) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
86 |
| 0 | 87 _phantom_object = C->top()->_idx; |
| 88 PointsToNode *phn = ptnode_adr(_phantom_object); | |
| 65 | 89 phn->_node = C->top(); |
| 0 | 90 phn->set_node_type(PointsToNode::JavaObject); |
| 91 phn->set_escape_state(PointsToNode::GlobalEscape); | |
| 92 } | |
| 93 | |
| 94 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) { | |
| 95 PointsToNode *f = ptnode_adr(from_i); | |
| 96 PointsToNode *t = ptnode_adr(to_i); | |
| 97 | |
| 98 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); | |
| 99 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge"); | |
| 100 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge"); | |
| 101 f->add_edge(to_i, PointsToNode::PointsToEdge); | |
| 102 } | |
| 103 | |
| 104 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) { | |
| 105 PointsToNode *f = ptnode_adr(from_i); | |
| 106 PointsToNode *t = ptnode_adr(to_i); | |
| 107 | |
| 108 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); | |
| 109 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge"); | |
| 110 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge"); | |
| 111 // don't add a self-referential edge, this can occur during removal of | |
| 112 // deferred edges | |
| 113 if (from_i != to_i) | |
| 114 f->add_edge(to_i, PointsToNode::DeferredEdge); | |
| 115 } | |
| 116 | |
| 65 | 117 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { |
| 118 const Type *adr_type = phase->type(adr); | |
| 119 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && | |
| 120 adr->in(AddPNode::Address)->is_Proj() && | |
| 121 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { | |
| 122 // We are computing a raw address for a store captured by an Initialize | |
| 123 // compute an appropriate address type. AddP cases #3 and #5 (see below). | |
| 124 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); | |
| 125 assert(offs != Type::OffsetBot || | |
| 126 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), | |
| 127 "offset must be a constant or it is initialization of array"); | |
| 128 return offs; | |
| 129 } | |
| 130 const TypePtr *t_ptr = adr_type->isa_ptr(); | |
| 0 | 131 assert(t_ptr != NULL, "must be a pointer type"); |
| 132 return t_ptr->offset(); | |
| 133 } | |
| 134 | |
| 135 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) { | |
| 136 PointsToNode *f = ptnode_adr(from_i); | |
| 137 PointsToNode *t = ptnode_adr(to_i); | |
| 138 | |
| 139 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); | |
| 140 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge"); | |
| 141 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge"); | |
| 142 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets"); | |
| 143 t->set_offset(offset); | |
| 144 | |
| 145 f->add_edge(to_i, PointsToNode::FieldEdge); | |
| 146 } | |
| 147 | |
| 148 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) { | |
| 149 PointsToNode *npt = ptnode_adr(ni); | |
| 150 PointsToNode::EscapeState old_es = npt->escape_state(); | |
| 151 if (es > old_es) | |
| 152 npt->set_escape_state(es); | |
| 153 } | |
| 154 | |
| 65 | 155 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt, |
| 156 PointsToNode::EscapeState es, bool done) { | |
| 157 PointsToNode* ptadr = ptnode_adr(n->_idx); | |
| 158 ptadr->_node = n; | |
| 159 ptadr->set_node_type(nt); | |
| 160 | |
| 161 // inline set_escape_state(idx, es); | |
| 162 PointsToNode::EscapeState old_es = ptadr->escape_state(); | |
| 163 if (es > old_es) | |
| 164 ptadr->set_escape_state(es); | |
| 165 | |
| 166 if (done) | |
| 167 _processed.set(n->_idx); | |
| 168 } | |
| 169 | |
| 0 | 170 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n, PhaseTransform *phase) { |
| 171 uint idx = n->_idx; | |
| 172 PointsToNode::EscapeState es; | |
| 173 | |
| 65 | 174 // If we are still collecting or there were no non-escaping allocations |
| 175 // we don't know the answer yet | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
176 if (_collecting) |
| 0 | 177 return PointsToNode::UnknownEscape; |
| 178 | |
| 179 // if the node was created after the escape computation, return | |
| 180 // UnknownEscape | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
181 if (idx >= nodes_size()) |
| 0 | 182 return PointsToNode::UnknownEscape; |
| 183 | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
184 es = ptnode_adr(idx)->escape_state(); |
| 0 | 185 |
| 186 // if we have already computed a value, return it | |
| 187 if (es != PointsToNode::UnknownEscape) | |
| 188 return es; | |
| 189 | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
190 // PointsTo() calls n->uncast() which can return a new ideal node. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
191 if (n->uncast()->_idx >= nodes_size()) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
192 return PointsToNode::UnknownEscape; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
193 |
| 0 | 194 // compute max escape state of anything this node could point to |
| 195 VectorSet ptset(Thread::current()->resource_area()); | |
| 196 PointsTo(ptset, n, phase); | |
| 65 | 197 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) { |
| 0 | 198 uint pt = i.elem; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
199 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state(); |
| 0 | 200 if (pes > es) |
| 201 es = pes; | |
| 202 } | |
| 203 // cache the computed escape state | |
| 204 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state"); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
205 ptnode_adr(idx)->set_escape_state(es); |
| 0 | 206 return es; |
| 207 } | |
| 208 | |
| 209 void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n, PhaseTransform *phase) { | |
| 210 VectorSet visited(Thread::current()->resource_area()); | |
| 211 GrowableArray<uint> worklist; | |
| 212 | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
213 #ifdef ASSERT |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
214 Node *orig_n = n; |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
215 #endif |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
216 |
| 65 | 217 n = n->uncast(); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
218 PointsToNode* npt = ptnode_adr(n->_idx); |
| 0 | 219 |
| 220 // If we have a JavaObject, return just that object | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
221 if (npt->node_type() == PointsToNode::JavaObject) { |
| 0 | 222 ptset.set(n->_idx); |
| 223 return; | |
| 224 } | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
225 #ifdef ASSERT |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
226 if (npt->_node == NULL) { |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
227 if (orig_n != n) |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
228 orig_n->dump(); |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
229 n->dump(); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
230 assert(npt->_node != NULL, "unregistered node"); |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
231 } |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
232 #endif |
| 0 | 233 worklist.push(n->_idx); |
| 234 while(worklist.length() > 0) { | |
| 235 int ni = worklist.pop(); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
236 if (visited.test_set(ni)) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
237 continue; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
238 |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
239 PointsToNode* pn = ptnode_adr(ni); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
240 // ensure that all inputs of a Phi have been processed |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
241 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),""); |
| 0 | 242 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
243 int edges_processed = 0; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
244 uint e_cnt = pn->edge_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
245 for (uint e = 0; e < e_cnt; e++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
246 uint etgt = pn->edge_target(e); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
247 PointsToNode::EdgeType et = pn->edge_type(e); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
248 if (et == PointsToNode::PointsToEdge) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
249 ptset.set(etgt); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
250 edges_processed++; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
251 } else if (et == PointsToNode::DeferredEdge) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
252 worklist.push(etgt); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
253 edges_processed++; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
254 } else { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
255 assert(false,"neither PointsToEdge or DeferredEdge"); |
| 0 | 256 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
257 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
258 if (edges_processed == 0) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
259 // no deferred or pointsto edges found. Assume the value was set |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
260 // outside this method. Add the phantom object to the pointsto set. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
261 ptset.set(_phantom_object); |
| 0 | 262 } |
| 263 } | |
| 264 } | |
| 265 | |
|
101
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
266 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) { |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
267 // This method is most expensive during ConnectionGraph construction. |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
268 // Reuse vectorSet and an additional growable array for deferred edges. |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
269 deferred_edges->clear(); |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
270 visited->Clear(); |
| 0 | 271 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
272 visited->set(ni); |
| 0 | 273 PointsToNode *ptn = ptnode_adr(ni); |
| 274 | |
|
101
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
275 // Mark current edges as visited and move deferred edges to separate array. |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
276 for (uint i = 0; i < ptn->edge_count(); ) { |
| 65 | 277 uint t = ptn->edge_target(i); |
|
101
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
278 #ifdef ASSERT |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
279 assert(!visited->test_set(t), "expecting no duplications"); |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
280 #else |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
281 visited->set(t); |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
282 #endif |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
283 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) { |
| 0 | 284 ptn->remove_edge(t, PointsToNode::DeferredEdge); |
|
101
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
285 deferred_edges->append(t); |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
286 } else { |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
287 i++; |
|
101
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
288 } |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
289 } |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
290 for (int next = 0; next < deferred_edges->length(); ++next) { |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
291 uint t = deferred_edges->at(next); |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
292 PointsToNode *ptt = ptnode_adr(t); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
293 uint e_cnt = ptt->edge_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
294 for (uint e = 0; e < e_cnt; e++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
295 uint etgt = ptt->edge_target(e); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
296 if (visited->test_set(etgt)) |
|
101
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
297 continue; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
298 |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
299 PointsToNode::EdgeType et = ptt->edge_type(e); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
300 if (et == PointsToNode::PointsToEdge) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
301 add_pointsto_edge(ni, etgt); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
302 if(etgt == _phantom_object) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
303 // Special case - field set outside (globally escaping). |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
304 ptn->set_escape_state(PointsToNode::GlobalEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
305 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
306 } else if (et == PointsToNode::DeferredEdge) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
307 deferred_edges->append(etgt); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
308 } else { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
309 assert(false,"invalid connection graph"); |
| 0 | 310 } |
| 311 } | |
| 312 } | |
| 313 } | |
| 314 | |
| 315 | |
| 316 // Add an edge to node given by "to_i" from any field of adr_i whose offset | |
| 317 // matches "offset" A deferred edge is added if to_i is a LocalVar, and | |
| 318 // a pointsto edge is added if it is a JavaObject | |
| 319 | |
| 320 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
321 PointsToNode* an = ptnode_adr(adr_i); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
322 PointsToNode* to = ptnode_adr(to_i); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
323 bool deferred = (to->node_type() == PointsToNode::LocalVar); |
| 0 | 324 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
325 for (uint fe = 0; fe < an->edge_count(); fe++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
326 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
327 int fi = an->edge_target(fe); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
328 PointsToNode* pf = ptnode_adr(fi); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
329 int po = pf->offset(); |
| 0 | 330 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { |
| 331 if (deferred) | |
| 332 add_deferred_edge(fi, to_i); | |
| 333 else | |
| 334 add_pointsto_edge(fi, to_i); | |
| 335 } | |
| 336 } | |
| 337 } | |
| 338 | |
| 65 | 339 // Add a deferred edge from node given by "from_i" to any field of adr_i |
| 340 // whose offset matches "offset". | |
| 0 | 341 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
342 PointsToNode* an = ptnode_adr(adr_i); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
343 for (uint fe = 0; fe < an->edge_count(); fe++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
344 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
345 int fi = an->edge_target(fe); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
346 PointsToNode* pf = ptnode_adr(fi); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
347 int po = pf->offset(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
348 if (pf->edge_count() == 0) { |
| 0 | 349 // we have not seen any stores to this field, assume it was set outside this method |
| 350 add_pointsto_edge(fi, _phantom_object); | |
| 351 } | |
| 352 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { | |
| 353 add_deferred_edge(from_i, fi); | |
| 354 } | |
| 355 } | |
| 356 } | |
| 357 | |
| 65 | 358 // Helper functions |
| 359 | |
| 360 static Node* get_addp_base(Node *addp) { | |
| 361 assert(addp->is_AddP(), "must be AddP"); | |
| 362 // | |
| 363 // AddP cases for Base and Address inputs: | |
| 364 // case #1. Direct object's field reference: | |
| 365 // Allocate | |
| 366 // | | |
| 367 // Proj #5 ( oop result ) | |
| 368 // | | |
| 369 // CheckCastPP (cast to instance type) | |
| 370 // | | | |
| 371 // AddP ( base == address ) | |
| 372 // | |
| 373 // case #2. Indirect object's field reference: | |
| 374 // Phi | |
| 375 // | | |
| 376 // CastPP (cast to instance type) | |
| 377 // | | | |
| 378 // AddP ( base == address ) | |
| 379 // | |
| 380 // case #3. Raw object's field reference for Initialize node: | |
| 381 // Allocate | |
| 382 // | | |
| 383 // Proj #5 ( oop result ) | |
| 384 // top | | |
| 385 // \ | | |
| 386 // AddP ( base == top ) | |
| 387 // | |
| 388 // case #4. Array's element reference: | |
| 389 // {CheckCastPP | CastPP} | |
| 390 // | | | | |
| 391 // | AddP ( array's element offset ) | |
| 392 // | | | |
| 393 // AddP ( array's offset ) | |
| 394 // | |
| 395 // case #5. Raw object's field reference for arraycopy stub call: | |
| 396 // The inline_native_clone() case when the arraycopy stub is called | |
| 397 // after the allocation before Initialize and CheckCastPP nodes. | |
| 398 // Allocate | |
| 399 // | | |
| 400 // Proj #5 ( oop result ) | |
| 401 // | | | |
| 402 // AddP ( base == address ) | |
| 403 // | |
| 77 | 404 // case #6. Constant Pool, ThreadLocal, CastX2P or |
| 405 // Raw object's field reference: | |
| 406 // {ConP, ThreadLocal, CastX2P, raw Load} | |
| 65 | 407 // top | |
| 408 // \ | | |
| 409 // AddP ( base == top ) | |
| 410 // | |
| 77 | 411 // case #7. Klass's field reference. |
| 412 // LoadKlass | |
| 413 // | | | |
| 414 // AddP ( base == address ) | |
| 415 // | |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
416 // case #8. narrow Klass's field reference. |
|
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
417 // LoadNKlass |
|
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
418 // | |
|
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
419 // DecodeN |
|
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
420 // | | |
|
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
421 // AddP ( base == address ) |
|
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
422 // |
| 65 | 423 Node *base = addp->in(AddPNode::Base)->uncast(); |
| 424 if (base->is_top()) { // The AddP case #3 and #6. | |
| 425 base = addp->in(AddPNode::Address)->uncast(); | |
| 426 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal || | |
|
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
427 base->Opcode() == Op_CastX2P || base->is_DecodeN() || |
| 77 | 428 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) || |
| 429 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity"); | |
| 0 | 430 } |
| 65 | 431 return base; |
| 432 } | |
| 433 | |
| 434 static Node* find_second_addp(Node* addp, Node* n) { | |
| 435 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); | |
| 436 | |
| 437 Node* addp2 = addp->raw_out(0); | |
| 438 if (addp->outcnt() == 1 && addp2->is_AddP() && | |
| 439 addp2->in(AddPNode::Base) == n && | |
| 440 addp2->in(AddPNode::Address) == addp) { | |
| 441 | |
| 442 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); | |
| 443 // | |
| 444 // Find array's offset to push it on worklist first and | |
| 445 // as result process an array's element offset first (pushed second) | |
| 446 // to avoid CastPP for the array's offset. | |
| 447 // Otherwise the inserted CastPP (LocalVar) will point to what | |
| 448 // the AddP (Field) points to. Which would be wrong since | |
| 449 // the algorithm expects the CastPP has the same point as | |
| 450 // as AddP's base CheckCastPP (LocalVar). | |
| 451 // | |
| 452 // ArrayAllocation | |
| 453 // | | |
| 454 // CheckCastPP | |
| 455 // | | |
| 456 // memProj (from ArrayAllocation CheckCastPP) | |
| 457 // | || | |
| 458 // | || Int (element index) | |
| 459 // | || | ConI (log(element size)) | |
| 460 // | || | / | |
| 461 // | || LShift | |
| 462 // | || / | |
| 463 // | AddP (array's element offset) | |
| 464 // | | | |
| 465 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) | |
| 466 // | / / | |
| 467 // AddP (array's offset) | |
| 468 // | | |
| 469 // Load/Store (memory operation on array's element) | |
| 470 // | |
| 471 return addp2; | |
| 472 } | |
| 473 return NULL; | |
| 0 | 474 } |
| 475 | |
| 476 // | |
| 477 // Adjust the type and inputs of an AddP which computes the | |
| 478 // address of a field of an instance | |
| 479 // | |
| 480 void ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) { | |
| 65 | 481 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); |
| 223 | 482 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); |
| 0 | 483 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); |
| 65 | 484 if (t == NULL) { |
| 485 // We are computing a raw address for a store captured by an Initialize | |
| 486 // compute an appropriate address type. | |
| 487 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); | |
| 488 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); | |
| 489 int offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); | |
| 490 assert(offs != Type::OffsetBot, "offset must be a constant"); | |
| 491 t = base_t->add_offset(offs)->is_oopptr(); | |
| 492 } | |
| 223 | 493 int inst_id = base_t->instance_id(); |
| 494 assert(!t->is_known_instance() || t->instance_id() == inst_id, | |
| 0 | 495 "old type must be non-instance or match new type"); |
| 496 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); | |
| 65 | 497 // Do NOT remove the next call: ensure an new alias index is allocated |
| 498 // for the instance type | |
| 0 | 499 int alias_idx = _compile->get_alias_index(tinst); |
| 500 igvn->set_type(addp, tinst); | |
| 501 // record the allocation in the node map | |
| 502 set_map(addp->_idx, get_map(base->_idx)); | |
| 65 | 503 // if the Address input is not the appropriate instance type |
| 504 // (due to intervening casts,) insert a cast | |
| 0 | 505 Node *adr = addp->in(AddPNode::Address); |
| 506 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); | |
| 65 | 507 if (atype != NULL && atype->instance_id() != inst_id) { |
| 223 | 508 assert(!atype->is_known_instance(), "no conflicting instances"); |
| 0 | 509 const TypeOopPtr *new_atype = base_t->add_offset(atype->offset())->isa_oopptr(); |
| 510 Node *acast = new (_compile, 2) CastPPNode(adr, new_atype); | |
| 511 acast->set_req(0, adr->in(0)); | |
| 512 igvn->set_type(acast, new_atype); | |
| 513 record_for_optimizer(acast); | |
| 514 Node *bcast = acast; | |
| 515 Node *abase = addp->in(AddPNode::Base); | |
| 516 if (abase != adr) { | |
| 517 bcast = new (_compile, 2) CastPPNode(abase, base_t); | |
| 518 bcast->set_req(0, abase->in(0)); | |
| 519 igvn->set_type(bcast, base_t); | |
| 520 record_for_optimizer(bcast); | |
| 521 } | |
| 522 igvn->hash_delete(addp); | |
| 523 addp->set_req(AddPNode::Base, bcast); | |
| 524 addp->set_req(AddPNode::Address, acast); | |
| 525 igvn->hash_insert(addp); | |
| 526 } | |
| 65 | 527 // Put on IGVN worklist since at least addp's type was changed above. |
| 528 record_for_optimizer(addp); | |
| 0 | 529 } |
| 530 | |
| 531 // | |
| 532 // Create a new version of orig_phi if necessary. Returns either the newly | |
| 533 // created phi or an existing phi. Sets create_new to indicate wheter a new | |
| 534 // phi was created. Cache the last newly created phi in the node map. | |
| 535 // | |
| 536 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) { | |
| 537 Compile *C = _compile; | |
| 538 new_created = false; | |
| 539 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); | |
| 540 // nothing to do if orig_phi is bottom memory or matches alias_idx | |
| 65 | 541 if (phi_alias_idx == alias_idx) { |
| 0 | 542 return orig_phi; |
| 543 } | |
| 544 // have we already created a Phi for this alias index? | |
| 545 PhiNode *result = get_map_phi(orig_phi->_idx); | |
| 546 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { | |
| 547 return result; | |
| 548 } | |
|
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
549 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) { |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
550 if (C->do_escape_analysis() == true && !C->failing()) { |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
551 // Retry compilation without escape analysis. |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
552 // If this is the first failure, the sentinel string will "stick" |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
553 // to the Compile object, and the C2Compiler will see it and retry. |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
554 C->record_failure(C2Compiler::retry_no_escape_analysis()); |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
555 } |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
556 return NULL; |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
557 } |
| 0 | 558 orig_phi_worklist.append_if_missing(orig_phi); |
| 65 | 559 const TypePtr *atype = C->get_adr_type(alias_idx); |
| 0 | 560 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); |
| 561 set_map_phi(orig_phi->_idx, result); | |
| 562 igvn->set_type(result, result->bottom_type()); | |
| 563 record_for_optimizer(result); | |
| 564 new_created = true; | |
| 565 return result; | |
| 566 } | |
| 567 | |
| 568 // | |
| 569 // Return a new version of Memory Phi "orig_phi" with the inputs having the | |
| 570 // specified alias index. | |
| 571 // | |
| 572 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) { | |
| 573 | |
| 574 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); | |
| 575 Compile *C = _compile; | |
| 576 bool new_phi_created; | |
| 65 | 577 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created); |
| 0 | 578 if (!new_phi_created) { |
| 579 return result; | |
| 580 } | |
| 581 | |
| 582 GrowableArray<PhiNode *> phi_list; | |
| 583 GrowableArray<uint> cur_input; | |
| 584 | |
| 585 PhiNode *phi = orig_phi; | |
| 586 uint idx = 1; | |
| 587 bool finished = false; | |
| 588 while(!finished) { | |
| 589 while (idx < phi->req()) { | |
| 65 | 590 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn); |
| 0 | 591 if (mem != NULL && mem->is_Phi()) { |
| 65 | 592 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created); |
| 0 | 593 if (new_phi_created) { |
| 594 // found an phi for which we created a new split, push current one on worklist and begin | |
| 595 // processing new one | |
| 596 phi_list.push(phi); | |
| 597 cur_input.push(idx); | |
| 598 phi = mem->as_Phi(); | |
| 65 | 599 result = newphi; |
| 0 | 600 idx = 1; |
| 601 continue; | |
| 602 } else { | |
| 65 | 603 mem = newphi; |
| 0 | 604 } |
| 605 } | |
|
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
606 if (C->failing()) { |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
607 return NULL; |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
608 } |
| 0 | 609 result->set_req(idx++, mem); |
| 610 } | |
| 611 #ifdef ASSERT | |
| 612 // verify that the new Phi has an input for each input of the original | |
| 613 assert( phi->req() == result->req(), "must have same number of inputs."); | |
| 614 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); | |
| 65 | 615 #endif |
| 616 // Check if all new phi's inputs have specified alias index. | |
| 617 // Otherwise use old phi. | |
| 0 | 618 for (uint i = 1; i < phi->req(); i++) { |
| 65 | 619 Node* in = result->in(i); |
| 620 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); | |
| 0 | 621 } |
| 622 // we have finished processing a Phi, see if there are any more to do | |
| 623 finished = (phi_list.length() == 0 ); | |
| 624 if (!finished) { | |
| 625 phi = phi_list.pop(); | |
| 626 idx = cur_input.pop(); | |
| 65 | 627 PhiNode *prev_result = get_map_phi(phi->_idx); |
| 628 prev_result->set_req(idx++, result); | |
| 629 result = prev_result; | |
| 0 | 630 } |
| 631 } | |
| 632 return result; | |
| 633 } | |
| 634 | |
| 65 | 635 |
| 636 // | |
| 637 // The next methods are derived from methods in MemNode. | |
| 638 // | |
| 639 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *tinst) { | |
| 640 Node *mem = mmem; | |
| 641 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally | |
| 642 // means an array I have not precisely typed yet. Do not do any | |
| 643 // alias stuff with it any time soon. | |
| 644 if( tinst->base() != Type::AnyPtr && | |
| 645 !(tinst->klass()->is_java_lang_Object() && | |
| 646 tinst->offset() == Type::OffsetBot) ) { | |
| 647 mem = mmem->memory_at(alias_idx); | |
| 648 // Update input if it is progress over what we have now | |
| 649 } | |
| 650 return mem; | |
| 651 } | |
| 652 | |
| 653 // | |
| 654 // Search memory chain of "mem" to find a MemNode whose address | |
| 655 // is the specified alias index. | |
| 656 // | |
| 657 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) { | |
| 658 if (orig_mem == NULL) | |
| 659 return orig_mem; | |
| 660 Compile* C = phase->C; | |
| 661 const TypeOopPtr *tinst = C->get_adr_type(alias_idx)->isa_oopptr(); | |
| 223 | 662 bool is_instance = (tinst != NULL) && tinst->is_known_instance(); |
| 65 | 663 Node *prev = NULL; |
| 664 Node *result = orig_mem; | |
| 665 while (prev != result) { | |
| 666 prev = result; | |
| 667 if (result->is_Mem()) { | |
| 668 MemNode *mem = result->as_Mem(); | |
| 669 const Type *at = phase->type(mem->in(MemNode::Address)); | |
| 670 if (at != Type::TOP) { | |
| 671 assert (at->isa_ptr() != NULL, "pointer type required."); | |
| 672 int idx = C->get_alias_index(at->is_ptr()); | |
| 673 if (idx == alias_idx) | |
| 674 break; | |
| 675 } | |
| 676 result = mem->in(MemNode::Memory); | |
| 677 } | |
| 678 if (!is_instance) | |
| 679 continue; // don't search further for non-instance types | |
| 680 // skip over a call which does not affect this memory slice | |
| 681 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { | |
| 682 Node *proj_in = result->in(0); | |
| 683 if (proj_in->is_Call()) { | |
| 684 CallNode *call = proj_in->as_Call(); | |
| 685 if (!call->may_modify(tinst, phase)) { | |
| 686 result = call->in(TypeFunc::Memory); | |
| 687 } | |
| 688 } else if (proj_in->is_Initialize()) { | |
| 689 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); | |
| 690 // Stop if this is the initialization for the object instance which | |
| 691 // which contains this memory slice, otherwise skip over it. | |
| 223 | 692 if (alloc == NULL || alloc->_idx != (uint)tinst->instance_id()) { |
| 65 | 693 result = proj_in->in(TypeFunc::Memory); |
| 694 } | |
| 695 } else if (proj_in->is_MemBar()) { | |
| 696 result = proj_in->in(TypeFunc::Memory); | |
| 697 } | |
| 698 } else if (result->is_MergeMem()) { | |
| 699 MergeMemNode *mmem = result->as_MergeMem(); | |
| 700 result = step_through_mergemem(mmem, alias_idx, tinst); | |
| 701 if (result == mmem->base_memory()) { | |
| 702 // Didn't find instance memory, search through general slice recursively. | |
| 703 result = mmem->memory_at(C->get_general_index(alias_idx)); | |
| 704 result = find_inst_mem(result, alias_idx, orig_phis, phase); | |
| 705 if (C->failing()) { | |
| 706 return NULL; | |
| 707 } | |
| 708 mmem->set_memory_at(alias_idx, result); | |
| 709 } | |
| 710 } else if (result->is_Phi() && | |
| 711 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { | |
| 712 Node *un = result->as_Phi()->unique_input(phase); | |
| 713 if (un != NULL) { | |
| 714 result = un; | |
| 715 } else { | |
| 716 break; | |
| 717 } | |
| 718 } | |
| 719 } | |
| 720 if (is_instance && result->is_Phi()) { | |
| 721 PhiNode *mphi = result->as_Phi(); | |
| 722 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); | |
| 723 const TypePtr *t = mphi->adr_type(); | |
| 724 if (C->get_alias_index(t) != alias_idx) { | |
| 725 result = split_memory_phi(mphi, alias_idx, orig_phis, phase); | |
| 726 } | |
| 727 } | |
| 728 // the result is either MemNode, PhiNode, InitializeNode. | |
| 729 return result; | |
| 730 } | |
| 731 | |
| 732 | |
| 0 | 733 // |
| 734 // Convert the types of unescaped object to instance types where possible, | |
| 735 // propagate the new type information through the graph, and update memory | |
| 736 // edges and MergeMem inputs to reflect the new type. | |
| 737 // | |
| 738 // We start with allocations (and calls which may be allocations) on alloc_worklist. | |
| 739 // The processing is done in 4 phases: | |
| 740 // | |
| 741 // Phase 1: Process possible allocations from alloc_worklist. Create instance | |
| 742 // types for the CheckCastPP for allocations where possible. | |
| 743 // Propagate the the new types through users as follows: | |
| 744 // casts and Phi: push users on alloc_worklist | |
| 745 // AddP: cast Base and Address inputs to the instance type | |
| 746 // push any AddP users on alloc_worklist and push any memnode | |
| 747 // users onto memnode_worklist. | |
| 748 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and | |
| 749 // search the Memory chain for a store with the appropriate type | |
| 750 // address type. If a Phi is found, create a new version with | |
| 751 // the approriate memory slices from each of the Phi inputs. | |
| 752 // For stores, process the users as follows: | |
| 753 // MemNode: push on memnode_worklist | |
| 754 // MergeMem: push on mergemem_worklist | |
| 755 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice | |
| 756 // moving the first node encountered of each instance type to the | |
| 757 // the input corresponding to its alias index. | |
| 758 // appropriate memory slice. | |
| 759 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. | |
| 760 // | |
| 761 // In the following example, the CheckCastPP nodes are the cast of allocation | |
| 762 // results and the allocation of node 29 is unescaped and eligible to be an | |
| 763 // instance type. | |
| 764 // | |
| 765 // We start with: | |
| 766 // | |
| 767 // 7 Parm #memory | |
| 768 // 10 ConI "12" | |
| 769 // 19 CheckCastPP "Foo" | |
| 770 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 | |
| 771 // 29 CheckCastPP "Foo" | |
| 772 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 | |
| 773 // | |
| 774 // 40 StoreP 25 7 20 ... alias_index=4 | |
| 775 // 50 StoreP 35 40 30 ... alias_index=4 | |
| 776 // 60 StoreP 45 50 20 ... alias_index=4 | |
| 777 // 70 LoadP _ 60 30 ... alias_index=4 | |
| 778 // 80 Phi 75 50 60 Memory alias_index=4 | |
| 779 // 90 LoadP _ 80 30 ... alias_index=4 | |
| 780 // 100 LoadP _ 80 20 ... alias_index=4 | |
| 781 // | |
| 782 // | |
| 783 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 | |
| 784 // and creating a new alias index for node 30. This gives: | |
| 785 // | |
| 786 // 7 Parm #memory | |
| 787 // 10 ConI "12" | |
| 788 // 19 CheckCastPP "Foo" | |
| 789 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 | |
| 790 // 29 CheckCastPP "Foo" iid=24 | |
| 791 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 | |
| 792 // | |
| 793 // 40 StoreP 25 7 20 ... alias_index=4 | |
| 794 // 50 StoreP 35 40 30 ... alias_index=6 | |
| 795 // 60 StoreP 45 50 20 ... alias_index=4 | |
| 796 // 70 LoadP _ 60 30 ... alias_index=6 | |
| 797 // 80 Phi 75 50 60 Memory alias_index=4 | |
| 798 // 90 LoadP _ 80 30 ... alias_index=6 | |
| 799 // 100 LoadP _ 80 20 ... alias_index=4 | |
| 800 // | |
| 801 // In phase 2, new memory inputs are computed for the loads and stores, | |
| 802 // And a new version of the phi is created. In phase 4, the inputs to | |
| 803 // node 80 are updated and then the memory nodes are updated with the | |
| 804 // values computed in phase 2. This results in: | |
| 805 // | |
| 806 // 7 Parm #memory | |
| 807 // 10 ConI "12" | |
| 808 // 19 CheckCastPP "Foo" | |
| 809 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 | |
| 810 // 29 CheckCastPP "Foo" iid=24 | |
| 811 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 | |
| 812 // | |
| 813 // 40 StoreP 25 7 20 ... alias_index=4 | |
| 814 // 50 StoreP 35 7 30 ... alias_index=6 | |
| 815 // 60 StoreP 45 40 20 ... alias_index=4 | |
| 816 // 70 LoadP _ 50 30 ... alias_index=6 | |
| 817 // 80 Phi 75 40 60 Memory alias_index=4 | |
| 818 // 120 Phi 75 50 50 Memory alias_index=6 | |
| 819 // 90 LoadP _ 120 30 ... alias_index=6 | |
| 820 // 100 LoadP _ 80 20 ... alias_index=4 | |
| 821 // | |
| 822 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { | |
| 823 GrowableArray<Node *> memnode_worklist; | |
| 824 GrowableArray<Node *> mergemem_worklist; | |
| 825 GrowableArray<PhiNode *> orig_phis; | |
| 826 PhaseGVN *igvn = _compile->initial_gvn(); | |
| 827 uint new_index_start = (uint) _compile->num_alias_types(); | |
| 828 VectorSet visited(Thread::current()->resource_area()); | |
| 829 VectorSet ptset(Thread::current()->resource_area()); | |
| 830 | |
| 65 | 831 |
| 832 // Phase 1: Process possible allocations from alloc_worklist. | |
| 833 // Create instance types for the CheckCastPP for allocations where possible. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
834 // |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
835 // (Note: don't forget to change the order of the second AddP node on |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
836 // the alloc_worklist if the order of the worklist processing is changed, |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
837 // see the comment in find_second_addp().) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
838 // |
| 0 | 839 while (alloc_worklist.length() != 0) { |
| 840 Node *n = alloc_worklist.pop(); | |
| 841 uint ni = n->_idx; | |
| 65 | 842 const TypeOopPtr* tinst = NULL; |
| 0 | 843 if (n->is_Call()) { |
| 844 CallNode *alloc = n->as_Call(); | |
| 845 // copy escape information to call node | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
846 PointsToNode* ptn = ptnode_adr(alloc->_idx); |
| 0 | 847 PointsToNode::EscapeState es = escape_state(alloc, igvn); |
| 65 | 848 // We have an allocation or call which returns a Java object, |
| 849 // see if it is unescaped. | |
| 850 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable) | |
| 0 | 851 continue; |
|
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
38
diff
changeset
|
852 if (alloc->is_Allocate()) { |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
38
diff
changeset
|
853 // Set the scalar_replaceable flag before the next check. |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
38
diff
changeset
|
854 alloc->as_Allocate()->_is_scalar_replaceable = true; |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
38
diff
changeset
|
855 } |
| 65 | 856 // find CheckCastPP of call return value |
| 857 n = alloc->result_cast(); | |
| 858 if (n == NULL || // No uses accept Initialize or | |
| 859 !n->is_CheckCastPP()) // not unique CheckCastPP. | |
| 860 continue; | |
| 861 // The inline code for Object.clone() casts the allocation result to | |
| 862 // java.lang.Object and then to the the actual type of the allocated | |
| 863 // object. Detect this case and use the second cast. | |
| 864 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL | |
| 865 && igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT) { | |
| 866 Node *cast2 = NULL; | |
| 867 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 868 Node *use = n->fast_out(i); | |
| 869 if (use->is_CheckCastPP()) { | |
| 870 cast2 = use; | |
| 871 break; | |
| 872 } | |
| 873 } | |
| 874 if (cast2 != NULL) { | |
| 875 n = cast2; | |
| 876 } else { | |
| 877 continue; | |
| 878 } | |
| 879 } | |
| 880 set_escape_state(n->_idx, es); | |
| 881 // in order for an object to be stackallocatable, it must be: | |
| 882 // - a direct allocation (not a call returning an object) | |
| 883 // - non-escaping | |
| 884 // - eligible to be a unique type | |
| 885 // - not determined to be ineligible by escape analysis | |
| 0 | 886 set_map(alloc->_idx, n); |
| 887 set_map(n->_idx, alloc); | |
| 65 | 888 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); |
| 889 if (t == NULL) | |
| 0 | 890 continue; // not a TypeInstPtr |
| 223 | 891 tinst = t->cast_to_instance_id(ni); |
| 0 | 892 igvn->hash_delete(n); |
| 893 igvn->set_type(n, tinst); | |
| 894 n->raise_bottom_type(tinst); | |
| 895 igvn->hash_insert(n); | |
| 65 | 896 record_for_optimizer(n); |
| 897 if (alloc->is_Allocate() && ptn->_scalar_replaceable && | |
| 898 (t->isa_instptr() || t->isa_aryptr())) { | |
| 163 | 899 |
| 900 // First, put on the worklist all Field edges from Connection Graph | |
| 901 // which is more accurate then putting immediate users from Ideal Graph. | |
| 902 for (uint e = 0; e < ptn->edge_count(); e++) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
903 Node *use = ptnode_adr(ptn->edge_target(e))->_node; |
| 163 | 904 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(), |
| 905 "only AddP nodes are Field edges in CG"); | |
| 906 if (use->outcnt() > 0) { // Don't process dead nodes | |
| 907 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); | |
| 908 if (addp2 != NULL) { | |
| 909 assert(alloc->is_AllocateArray(),"array allocation was expected"); | |
| 910 alloc_worklist.append_if_missing(addp2); | |
| 911 } | |
| 912 alloc_worklist.append_if_missing(use); | |
| 913 } | |
| 914 } | |
| 915 | |
| 65 | 916 // An allocation may have an Initialize which has raw stores. Scan |
| 917 // the users of the raw allocation result and push AddP users | |
| 918 // on alloc_worklist. | |
| 919 Node *raw_result = alloc->proj_out(TypeFunc::Parms); | |
| 920 assert (raw_result != NULL, "must have an allocation result"); | |
| 921 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { | |
| 922 Node *use = raw_result->fast_out(i); | |
| 923 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes | |
| 924 Node* addp2 = find_second_addp(use, raw_result); | |
| 925 if (addp2 != NULL) { | |
| 926 assert(alloc->is_AllocateArray(),"array allocation was expected"); | |
| 927 alloc_worklist.append_if_missing(addp2); | |
| 928 } | |
| 929 alloc_worklist.append_if_missing(use); | |
| 930 } else if (use->is_Initialize()) { | |
| 931 memnode_worklist.append_if_missing(use); | |
| 932 } | |
| 933 } | |
| 934 } | |
| 0 | 935 } else if (n->is_AddP()) { |
| 936 ptset.Clear(); | |
| 65 | 937 PointsTo(ptset, get_addp_base(n), igvn); |
| 0 | 938 assert(ptset.Size() == 1, "AddP address is unique"); |
| 65 | 939 uint elem = ptset.getelem(); // Allocation node's index |
| 940 if (elem == _phantom_object) | |
| 941 continue; // Assume the value was set outside this method. | |
| 942 Node *base = get_map(elem); // CheckCastPP node | |
| 0 | 943 split_AddP(n, base, igvn); |
| 65 | 944 tinst = igvn->type(base)->isa_oopptr(); |
| 945 } else if (n->is_Phi() || | |
| 946 n->is_CheckCastPP() || | |
|
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
947 n->is_EncodeP() || |
|
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
948 n->is_DecodeN() || |
| 65 | 949 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { |
| 0 | 950 if (visited.test_set(n->_idx)) { |
| 951 assert(n->is_Phi(), "loops only through Phi's"); | |
| 952 continue; // already processed | |
| 953 } | |
| 954 ptset.Clear(); | |
| 955 PointsTo(ptset, n, igvn); | |
| 956 if (ptset.Size() == 1) { | |
| 65 | 957 uint elem = ptset.getelem(); // Allocation node's index |
| 958 if (elem == _phantom_object) | |
| 959 continue; // Assume the value was set outside this method. | |
| 960 Node *val = get_map(elem); // CheckCastPP node | |
| 0 | 961 TypeNode *tn = n->as_Type(); |
| 65 | 962 tinst = igvn->type(val)->isa_oopptr(); |
| 223 | 963 assert(tinst != NULL && tinst->is_known_instance() && |
| 964 (uint)tinst->instance_id() == elem , "instance type expected."); | |
| 163 | 965 |
| 966 const Type *tn_type = igvn->type(tn); | |
| 223 | 967 const TypeOopPtr *tn_t; |
| 968 if (tn_type->isa_narrowoop()) { | |
| 969 tn_t = tn_type->make_ptr()->isa_oopptr(); | |
| 970 } else { | |
| 971 tn_t = tn_type->isa_oopptr(); | |
| 972 } | |
| 0 | 973 |
| 65 | 974 if (tn_t != NULL && |
| 223 | 975 tinst->cast_to_instance_id(TypeOopPtr::InstanceBot)->higher_equal(tn_t)) { |
| 163 | 976 if (tn_type->isa_narrowoop()) { |
| 977 tn_type = tinst->make_narrowoop(); | |
| 978 } else { | |
| 979 tn_type = tinst; | |
| 980 } | |
| 0 | 981 igvn->hash_delete(tn); |
| 163 | 982 igvn->set_type(tn, tn_type); |
| 983 tn->set_type(tn_type); | |
| 0 | 984 igvn->hash_insert(tn); |
| 65 | 985 record_for_optimizer(n); |
| 0 | 986 } |
| 987 } | |
| 988 } else { | |
| 989 continue; | |
| 990 } | |
| 991 // push users on appropriate worklist | |
| 992 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 993 Node *use = n->fast_out(i); | |
| 994 if(use->is_Mem() && use->in(MemNode::Address) == n) { | |
| 65 | 995 memnode_worklist.append_if_missing(use); |
| 996 } else if (use->is_Initialize()) { | |
| 997 memnode_worklist.append_if_missing(use); | |
| 998 } else if (use->is_MergeMem()) { | |
| 999 mergemem_worklist.append_if_missing(use); | |
| 1000 } else if (use->is_Call() && tinst != NULL) { | |
| 1001 // Look for MergeMem nodes for calls which reference unique allocation | |
| 1002 // (through CheckCastPP nodes) even for debug info. | |
| 1003 Node* m = use->in(TypeFunc::Memory); | |
| 1004 uint iid = tinst->instance_id(); | |
| 1005 while (m->is_Proj() && m->in(0)->is_Call() && | |
| 1006 m->in(0) != use && !m->in(0)->_idx != iid) { | |
| 1007 m = m->in(0)->in(TypeFunc::Memory); | |
| 1008 } | |
| 1009 if (m->is_MergeMem()) { | |
| 1010 mergemem_worklist.append_if_missing(m); | |
| 1011 } | |
| 1012 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes | |
| 1013 Node* addp2 = find_second_addp(use, n); | |
| 1014 if (addp2 != NULL) { | |
| 1015 alloc_worklist.append_if_missing(addp2); | |
| 1016 } | |
| 1017 alloc_worklist.append_if_missing(use); | |
| 1018 } else if (use->is_Phi() || | |
| 1019 use->is_CheckCastPP() || | |
|
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1020 use->is_EncodeP() || |
|
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1021 use->is_DecodeN() || |
| 65 | 1022 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { |
| 1023 alloc_worklist.append_if_missing(use); | |
| 0 | 1024 } |
| 1025 } | |
| 1026 | |
| 1027 } | |
| 65 | 1028 // New alias types were created in split_AddP(). |
| 0 | 1029 uint new_index_end = (uint) _compile->num_alias_types(); |
| 1030 | |
| 1031 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and | |
| 1032 // compute new values for Memory inputs (the Memory inputs are not | |
| 1033 // actually updated until phase 4.) | |
| 1034 if (memnode_worklist.length() == 0) | |
| 1035 return; // nothing to do | |
| 1036 | |
| 1037 while (memnode_worklist.length() != 0) { | |
| 1038 Node *n = memnode_worklist.pop(); | |
| 65 | 1039 if (visited.test_set(n->_idx)) |
| 1040 continue; | |
| 0 | 1041 if (n->is_Phi()) { |
| 1042 assert(n->as_Phi()->adr_type() != TypePtr::BOTTOM, "narrow memory slice required"); | |
| 1043 // we don't need to do anything, but the users must be pushed if we haven't processed | |
| 1044 // this Phi before | |
| 65 | 1045 } else if (n->is_Initialize()) { |
| 1046 // we don't need to do anything, but the users of the memory projection must be pushed | |
| 1047 n = n->as_Initialize()->proj_out(TypeFunc::Memory); | |
| 1048 if (n == NULL) | |
| 0 | 1049 continue; |
| 1050 } else { | |
| 1051 assert(n->is_Mem(), "memory node required."); | |
| 1052 Node *addr = n->in(MemNode::Address); | |
| 65 | 1053 assert(addr->is_AddP(), "AddP required"); |
| 0 | 1054 const Type *addr_t = igvn->type(addr); |
| 1055 if (addr_t == Type::TOP) | |
| 1056 continue; | |
| 1057 assert (addr_t->isa_ptr() != NULL, "pointer type required."); | |
| 1058 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); | |
| 65 | 1059 assert ((uint)alias_idx < new_index_end, "wrong alias index"); |
| 1060 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn); | |
|
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
1061 if (_compile->failing()) { |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
1062 return; |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
1063 } |
| 65 | 1064 if (mem != n->in(MemNode::Memory)) { |
| 0 | 1065 set_map(n->_idx, mem); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1066 ptnode_adr(n->_idx)->_node = n; |
| 65 | 1067 } |
| 0 | 1068 if (n->is_Load()) { |
| 1069 continue; // don't push users | |
| 1070 } else if (n->is_LoadStore()) { | |
| 1071 // get the memory projection | |
| 1072 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1073 Node *use = n->fast_out(i); | |
| 1074 if (use->Opcode() == Op_SCMemProj) { | |
| 1075 n = use; | |
| 1076 break; | |
| 1077 } | |
| 1078 } | |
| 1079 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); | |
| 1080 } | |
| 1081 } | |
| 1082 // push user on appropriate worklist | |
| 1083 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1084 Node *use = n->fast_out(i); | |
| 1085 if (use->is_Phi()) { | |
| 65 | 1086 memnode_worklist.append_if_missing(use); |
| 0 | 1087 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) { |
| 65 | 1088 memnode_worklist.append_if_missing(use); |
| 1089 } else if (use->is_Initialize()) { | |
| 1090 memnode_worklist.append_if_missing(use); | |
| 0 | 1091 } else if (use->is_MergeMem()) { |
| 65 | 1092 mergemem_worklist.append_if_missing(use); |
| 0 | 1093 } |
| 1094 } | |
| 1095 } | |
| 1096 | |
| 65 | 1097 // Phase 3: Process MergeMem nodes from mergemem_worklist. |
| 1098 // Walk each memory moving the first node encountered of each | |
| 1099 // instance type to the the input corresponding to its alias index. | |
| 0 | 1100 while (mergemem_worklist.length() != 0) { |
| 1101 Node *n = mergemem_worklist.pop(); | |
| 1102 assert(n->is_MergeMem(), "MergeMem node required."); | |
| 65 | 1103 if (visited.test_set(n->_idx)) |
| 1104 continue; | |
| 0 | 1105 MergeMemNode *nmm = n->as_MergeMem(); |
| 1106 // Note: we don't want to use MergeMemStream here because we only want to | |
| 65 | 1107 // scan inputs which exist at the start, not ones we add during processing. |
| 0 | 1108 uint nslices = nmm->req(); |
| 1109 igvn->hash_delete(nmm); | |
| 1110 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { | |
| 65 | 1111 Node* mem = nmm->in(i); |
| 1112 Node* cur = NULL; | |
| 0 | 1113 if (mem == NULL || mem->is_top()) |
| 1114 continue; | |
| 1115 while (mem->is_Mem()) { | |
| 1116 const Type *at = igvn->type(mem->in(MemNode::Address)); | |
| 1117 if (at != Type::TOP) { | |
| 1118 assert (at->isa_ptr() != NULL, "pointer type required."); | |
| 1119 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); | |
| 1120 if (idx == i) { | |
| 1121 if (cur == NULL) | |
| 1122 cur = mem; | |
| 1123 } else { | |
| 1124 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { | |
| 1125 nmm->set_memory_at(idx, mem); | |
| 1126 } | |
| 1127 } | |
| 1128 } | |
| 1129 mem = mem->in(MemNode::Memory); | |
| 1130 } | |
| 1131 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); | |
| 65 | 1132 // Find any instance of the current type if we haven't encountered |
| 1133 // a value of the instance along the chain. | |
| 1134 for (uint ni = new_index_start; ni < new_index_end; ni++) { | |
| 1135 if((uint)_compile->get_general_index(ni) == i) { | |
| 1136 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); | |
| 1137 if (nmm->is_empty_memory(m)) { | |
| 1138 Node* result = find_inst_mem(mem, ni, orig_phis, igvn); | |
| 1139 if (_compile->failing()) { | |
| 1140 return; | |
| 1141 } | |
| 1142 nmm->set_memory_at(ni, result); | |
| 1143 } | |
| 1144 } | |
| 1145 } | |
| 1146 } | |
| 1147 // Find the rest of instances values | |
| 1148 for (uint ni = new_index_start; ni < new_index_end; ni++) { | |
| 1149 const TypeOopPtr *tinst = igvn->C->get_adr_type(ni)->isa_oopptr(); | |
| 1150 Node* result = step_through_mergemem(nmm, ni, tinst); | |
| 1151 if (result == nmm->base_memory()) { | |
| 1152 // Didn't find instance memory, search through general slice recursively. | |
| 1153 result = nmm->memory_at(igvn->C->get_general_index(ni)); | |
| 1154 result = find_inst_mem(result, ni, orig_phis, igvn); | |
| 1155 if (_compile->failing()) { | |
| 1156 return; | |
| 1157 } | |
| 1158 nmm->set_memory_at(ni, result); | |
| 1159 } | |
| 1160 } | |
| 1161 igvn->hash_insert(nmm); | |
| 1162 record_for_optimizer(nmm); | |
| 1163 | |
| 1164 // Propagate new memory slices to following MergeMem nodes. | |
| 1165 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1166 Node *use = n->fast_out(i); | |
| 1167 if (use->is_Call()) { | |
| 1168 CallNode* in = use->as_Call(); | |
| 1169 if (in->proj_out(TypeFunc::Memory) != NULL) { | |
| 1170 Node* m = in->proj_out(TypeFunc::Memory); | |
| 1171 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { | |
| 1172 Node* mm = m->fast_out(j); | |
| 1173 if (mm->is_MergeMem()) { | |
| 1174 mergemem_worklist.append_if_missing(mm); | |
| 1175 } | |
| 1176 } | |
| 1177 } | |
| 1178 if (use->is_Allocate()) { | |
| 1179 use = use->as_Allocate()->initialization(); | |
| 1180 if (use == NULL) { | |
| 1181 continue; | |
| 1182 } | |
| 1183 } | |
| 1184 } | |
| 1185 if (use->is_Initialize()) { | |
| 1186 InitializeNode* in = use->as_Initialize(); | |
| 1187 if (in->proj_out(TypeFunc::Memory) != NULL) { | |
| 1188 Node* m = in->proj_out(TypeFunc::Memory); | |
| 1189 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { | |
| 1190 Node* mm = m->fast_out(j); | |
| 1191 if (mm->is_MergeMem()) { | |
| 1192 mergemem_worklist.append_if_missing(mm); | |
| 0 | 1193 } |
| 1194 } | |
| 1195 } | |
| 1196 } | |
| 1197 } | |
| 1198 } | |
| 1199 | |
| 65 | 1200 // Phase 4: Update the inputs of non-instance memory Phis and |
| 1201 // the Memory input of memnodes | |
| 0 | 1202 // First update the inputs of any non-instance Phi's from |
| 1203 // which we split out an instance Phi. Note we don't have | |
| 1204 // to recursively process Phi's encounted on the input memory | |
| 1205 // chains as is done in split_memory_phi() since they will | |
| 1206 // also be processed here. | |
| 1207 while (orig_phis.length() != 0) { | |
| 1208 PhiNode *phi = orig_phis.pop(); | |
| 1209 int alias_idx = _compile->get_alias_index(phi->adr_type()); | |
| 1210 igvn->hash_delete(phi); | |
| 1211 for (uint i = 1; i < phi->req(); i++) { | |
| 1212 Node *mem = phi->in(i); | |
| 65 | 1213 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn); |
| 1214 if (_compile->failing()) { | |
| 1215 return; | |
| 1216 } | |
| 0 | 1217 if (mem != new_mem) { |
| 1218 phi->set_req(i, new_mem); | |
| 1219 } | |
| 1220 } | |
| 1221 igvn->hash_insert(phi); | |
| 1222 record_for_optimizer(phi); | |
| 1223 } | |
| 1224 | |
| 1225 // Update the memory inputs of MemNodes with the value we computed | |
| 1226 // in Phase 2. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1227 for (uint i = 0; i < nodes_size(); i++) { |
| 0 | 1228 Node *nmem = get_map(i); |
| 1229 if (nmem != NULL) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1230 Node *n = ptnode_adr(i)->_node; |
| 0 | 1231 if (n != NULL && n->is_Mem()) { |
| 1232 igvn->hash_delete(n); | |
| 1233 n->set_req(MemNode::Memory, nmem); | |
| 1234 igvn->hash_insert(n); | |
| 1235 record_for_optimizer(n); | |
| 1236 } | |
| 1237 } | |
| 1238 } | |
| 1239 } | |
| 1240 | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1241 bool ConnectionGraph::has_candidates(Compile *C) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1242 // EA brings benefits only when the code has allocations and/or locks which |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1243 // are represented by ideal Macro nodes. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1244 int cnt = C->macro_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1245 for( int i=0; i < cnt; i++ ) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1246 Node *n = C->macro_node(i); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1247 if ( n->is_Allocate() ) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1248 return true; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1249 if( n->is_Lock() ) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1250 Node* obj = n->as_Lock()->obj_node()->uncast(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1251 if( !(obj->is_Parm() || obj->is_Con()) ) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1252 return true; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1253 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1254 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1255 return false; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1256 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1257 |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1258 bool ConnectionGraph::compute_escape() { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1259 Compile* C = _compile; |
| 65 | 1260 |
| 163 | 1261 // 1. Populate Connection Graph (CG) with Ideal nodes. |
| 65 | 1262 |
| 1263 Unique_Node_List worklist_init; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1264 worklist_init.map(C->unique(), NULL); // preallocate space |
| 65 | 1265 |
| 1266 // Initialize worklist | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1267 if (C->root() != NULL) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1268 worklist_init.push(C->root()); |
| 65 | 1269 } |
| 1270 | |
| 1271 GrowableArray<int> cg_worklist; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1272 PhaseGVN* igvn = C->initial_gvn(); |
| 65 | 1273 bool has_allocations = false; |
| 1274 | |
| 1275 // Push all useful nodes onto CG list and set their type. | |
| 1276 for( uint next = 0; next < worklist_init.size(); ++next ) { | |
| 1277 Node* n = worklist_init.at(next); | |
| 1278 record_for_escape_analysis(n, igvn); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1279 // Only allocations and java static calls results are checked |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1280 // for an escape status. See process_call_result() below. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1281 if (n->is_Allocate() || n->is_CallStaticJava() && |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1282 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) { |
| 65 | 1283 has_allocations = true; |
| 1284 } | |
| 1285 if(n->is_AddP()) | |
| 1286 cg_worklist.append(n->_idx); | |
| 1287 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1288 Node* m = n->fast_out(i); // Get user | |
| 1289 worklist_init.push(m); | |
| 1290 } | |
| 1291 } | |
| 0 | 1292 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1293 if (!has_allocations) { |
| 65 | 1294 _collecting = false; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1295 return false; // Nothing to do. |
| 65 | 1296 } |
| 1297 | |
| 1298 // 2. First pass to create simple CG edges (doesn't require to walk CG). | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1299 uint delayed_size = _delayed_worklist.size(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1300 for( uint next = 0; next < delayed_size; ++next ) { |
| 65 | 1301 Node* n = _delayed_worklist.at(next); |
| 1302 build_connection_graph(n, igvn); | |
| 1303 } | |
| 0 | 1304 |
| 65 | 1305 // 3. Pass to create fields edges (Allocate -F-> AddP). |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1306 uint cg_length = cg_worklist.length(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1307 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1308 int ni = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1309 build_connection_graph(ptnode_adr(ni)->_node, igvn); |
| 65 | 1310 } |
| 1311 | |
| 1312 cg_worklist.clear(); | |
| 1313 cg_worklist.append(_phantom_object); | |
| 1314 | |
| 1315 // 4. Build Connection Graph which need | |
| 1316 // to walk the connection graph. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1317 for (uint ni = 0; ni < nodes_size(); ni++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1318 PointsToNode* ptn = ptnode_adr(ni); |
| 65 | 1319 Node *n = ptn->_node; |
| 1320 if (n != NULL) { // Call, AddP, LoadP, StoreP | |
| 1321 build_connection_graph(n, igvn); | |
| 1322 if (ptn->node_type() != PointsToNode::UnknownType) | |
| 1323 cg_worklist.append(n->_idx); // Collect CG nodes | |
| 1324 } | |
| 0 | 1325 } |
| 1326 | |
| 1327 VectorSet ptset(Thread::current()->resource_area()); | |
|
101
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
1328 GrowableArray<uint> deferred_edges; |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
1329 VectorSet visited(Thread::current()->resource_area()); |
| 0 | 1330 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1331 // 5. Remove deferred edges from the graph and collect |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1332 // information needed for type splitting. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1333 cg_length = cg_worklist.length(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1334 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1335 int ni = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1336 PointsToNode* ptn = ptnode_adr(ni); |
| 0 | 1337 PointsToNode::NodeType nt = ptn->node_type(); |
| 1338 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) { | |
|
101
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
1339 remove_deferred(ni, &deferred_edges, &visited); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1340 Node *n = ptn->_node; |
| 0 | 1341 if (n->is_AddP()) { |
| 65 | 1342 // If this AddP computes an address which may point to more that one |
| 163 | 1343 // object or more then one field (array's element), nothing the address |
| 1344 // points to can be scalar replaceable. | |
| 65 | 1345 Node *base = get_addp_base(n); |
| 0 | 1346 ptset.Clear(); |
| 1347 PointsTo(ptset, base, igvn); | |
| 163 | 1348 if (ptset.Size() > 1 || |
| 1349 (ptset.Size() != 0 && ptn->offset() == Type::OffsetBot)) { | |
| 0 | 1350 for( VectorSetI j(&ptset); j.test(); ++j ) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1351 ptnode_adr(j.elem)->_scalar_replaceable = false; |
| 0 | 1352 } |
| 1353 } | |
| 1354 } | |
| 1355 } | |
| 1356 } | |
| 65 | 1357 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1358 // 6. Propagate escape states. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1359 GrowableArray<int> worklist; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1360 bool has_non_escaping_obj = false; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1361 |
| 0 | 1362 // push all GlobalEscape nodes on the worklist |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1363 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1364 int nk = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1365 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1366 worklist.push(nk); |
| 0 | 1367 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1368 // mark all nodes reachable from GlobalEscape nodes |
| 0 | 1369 while(worklist.length() > 0) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1370 PointsToNode* ptn = ptnode_adr(worklist.pop()); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1371 uint e_cnt = ptn->edge_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1372 for (uint ei = 0; ei < e_cnt; ei++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1373 uint npi = ptn->edge_target(ei); |
| 0 | 1374 PointsToNode *np = ptnode_adr(npi); |
| 65 | 1375 if (np->escape_state() < PointsToNode::GlobalEscape) { |
| 0 | 1376 np->set_escape_state(PointsToNode::GlobalEscape); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1377 worklist.push(npi); |
| 0 | 1378 } |
| 1379 } | |
| 1380 } | |
| 1381 | |
| 1382 // push all ArgEscape nodes on the worklist | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1383 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1384 int nk = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1385 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape) |
| 0 | 1386 worklist.push(nk); |
| 1387 } | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1388 // mark all nodes reachable from ArgEscape nodes |
| 0 | 1389 while(worklist.length() > 0) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1390 PointsToNode* ptn = ptnode_adr(worklist.pop()); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1391 if (ptn->node_type() == PointsToNode::JavaObject) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1392 has_non_escaping_obj = true; // Non GlobalEscape |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1393 uint e_cnt = ptn->edge_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1394 for (uint ei = 0; ei < e_cnt; ei++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1395 uint npi = ptn->edge_target(ei); |
| 0 | 1396 PointsToNode *np = ptnode_adr(npi); |
| 65 | 1397 if (np->escape_state() < PointsToNode::ArgEscape) { |
| 0 | 1398 np->set_escape_state(PointsToNode::ArgEscape); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1399 worklist.push(npi); |
| 0 | 1400 } |
| 1401 } | |
| 1402 } | |
| 65 | 1403 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1404 GrowableArray<Node*> alloc_worklist; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1405 |
| 65 | 1406 // push all NoEscape nodes on the worklist |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1407 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1408 int nk = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1409 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape) |
| 65 | 1410 worklist.push(nk); |
| 1411 } | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1412 // mark all nodes reachable from NoEscape nodes |
| 65 | 1413 while(worklist.length() > 0) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1414 PointsToNode* ptn = ptnode_adr(worklist.pop()); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1415 if (ptn->node_type() == PointsToNode::JavaObject) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1416 has_non_escaping_obj = true; // Non GlobalEscape |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1417 Node* n = ptn->_node; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1418 if (n->is_Allocate() && ptn->_scalar_replaceable ) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1419 // Push scalar replaceable alocations on alloc_worklist |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1420 // for processing in split_unique_types(). |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1421 alloc_worklist.append(n); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1422 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1423 uint e_cnt = ptn->edge_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1424 for (uint ei = 0; ei < e_cnt; ei++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1425 uint npi = ptn->edge_target(ei); |
| 65 | 1426 PointsToNode *np = ptnode_adr(npi); |
| 1427 if (np->escape_state() < PointsToNode::NoEscape) { | |
| 1428 np->set_escape_state(PointsToNode::NoEscape); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1429 worklist.push(npi); |
| 65 | 1430 } |
| 1431 } | |
| 1432 } | |
| 1433 | |
| 0 | 1434 _collecting = false; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1435 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build"); |
| 0 | 1436 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1437 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1438 if ( has_scalar_replaceable_candidates && |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1439 C->AliasLevel() >= 3 && EliminateAllocations ) { |
| 0 | 1440 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1441 // Now use the escape information to create unique types for |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1442 // scalar replaceable objects. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1443 split_unique_types(alloc_worklist); |
| 0 | 1444 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1445 if (C->failing()) return false; |
| 0 | 1446 |
| 65 | 1447 // Clean up after split unique types. |
| 1448 ResourceMark rm; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1449 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn()); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1450 |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1451 C->print_method("After Escape Analysis", 2); |
| 0 | 1452 |
| 65 | 1453 #ifdef ASSERT |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1454 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { |
| 65 | 1455 tty->print("=== No allocations eliminated for "); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1456 C->method()->print_short_name(); |
| 65 | 1457 if(!EliminateAllocations) { |
| 1458 tty->print(" since EliminateAllocations is off ==="); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1459 } else if(!has_scalar_replaceable_candidates) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1460 tty->print(" since there are no scalar replaceable candidates ==="); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1461 } else if(C->AliasLevel() < 3) { |
| 65 | 1462 tty->print(" since AliasLevel < 3 ==="); |
| 0 | 1463 } |
| 65 | 1464 tty->cr(); |
| 1465 #endif | |
| 0 | 1466 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1467 return has_non_escaping_obj; |
| 0 | 1468 } |
| 1469 | |
| 1470 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) { | |
| 1471 | |
| 1472 switch (call->Opcode()) { | |
| 65 | 1473 #ifdef ASSERT |
| 0 | 1474 case Op_Allocate: |
| 1475 case Op_AllocateArray: | |
| 1476 case Op_Lock: | |
| 1477 case Op_Unlock: | |
| 65 | 1478 assert(false, "should be done already"); |
| 0 | 1479 break; |
| 65 | 1480 #endif |
| 1481 case Op_CallLeafNoFP: | |
| 1482 { | |
| 1483 // Stub calls, objects do not escape but they are not scale replaceable. | |
| 1484 // Adjust escape state for outgoing arguments. | |
| 1485 const TypeTuple * d = call->tf()->domain(); | |
| 1486 VectorSet ptset(Thread::current()->resource_area()); | |
| 1487 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { | |
| 1488 const Type* at = d->field_at(i); | |
| 1489 Node *arg = call->in(i)->uncast(); | |
| 1490 const Type *aat = phase->type(arg); | |
| 1491 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr()) { | |
| 1492 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || | |
| 1493 aat->isa_ptr() != NULL, "expecting an Ptr"); | |
| 1494 set_escape_state(arg->_idx, PointsToNode::ArgEscape); | |
| 1495 if (arg->is_AddP()) { | |
| 1496 // | |
| 1497 // The inline_native_clone() case when the arraycopy stub is called | |
| 1498 // after the allocation before Initialize and CheckCastPP nodes. | |
| 1499 // | |
| 1500 // Set AddP's base (Allocate) as not scalar replaceable since | |
| 1501 // pointer to the base (with offset) is passed as argument. | |
| 1502 // | |
| 1503 arg = get_addp_base(arg); | |
| 1504 } | |
| 1505 ptset.Clear(); | |
| 1506 PointsTo(ptset, arg, phase); | |
| 1507 for( VectorSetI j(&ptset); j.test(); ++j ) { | |
| 1508 uint pt = j.elem; | |
| 1509 set_escape_state(pt, PointsToNode::ArgEscape); | |
| 1510 } | |
| 1511 } | |
| 1512 } | |
| 1513 break; | |
| 1514 } | |
| 0 | 1515 |
| 1516 case Op_CallStaticJava: | |
| 1517 // For a static call, we know exactly what method is being called. | |
| 1518 // Use bytecode estimator to record the call's escape affects | |
| 1519 { | |
| 1520 ciMethod *meth = call->as_CallJava()->method(); | |
| 65 | 1521 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; |
| 1522 // fall-through if not a Java method or no analyzer information | |
| 1523 if (call_analyzer != NULL) { | |
| 0 | 1524 const TypeTuple * d = call->tf()->domain(); |
| 1525 VectorSet ptset(Thread::current()->resource_area()); | |
| 65 | 1526 bool copy_dependencies = false; |
| 0 | 1527 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { |
| 1528 const Type* at = d->field_at(i); | |
| 1529 int k = i - TypeFunc::Parms; | |
| 1530 | |
| 1531 if (at->isa_oopptr() != NULL) { | |
| 65 | 1532 Node *arg = call->in(i)->uncast(); |
| 0 | 1533 |
| 65 | 1534 bool global_escapes = false; |
| 1535 bool fields_escapes = false; | |
| 1536 if (!call_analyzer->is_arg_stack(k)) { | |
| 1537 // The argument global escapes, mark everything it could point to | |
| 1538 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); | |
| 1539 global_escapes = true; | |
| 1540 } else { | |
| 1541 if (!call_analyzer->is_arg_local(k)) { | |
| 1542 // The argument itself doesn't escape, but any fields might | |
| 1543 fields_escapes = true; | |
| 0 | 1544 } |
| 65 | 1545 set_escape_state(arg->_idx, PointsToNode::ArgEscape); |
| 1546 copy_dependencies = true; | |
| 1547 } | |
| 1548 | |
| 1549 ptset.Clear(); | |
| 1550 PointsTo(ptset, arg, phase); | |
| 1551 for( VectorSetI j(&ptset); j.test(); ++j ) { | |
| 1552 uint pt = j.elem; | |
| 1553 if (global_escapes) { | |
| 1554 //The argument global escapes, mark everything it could point to | |
| 1555 set_escape_state(pt, PointsToNode::GlobalEscape); | |
| 1556 } else { | |
| 1557 if (fields_escapes) { | |
| 1558 // The argument itself doesn't escape, but any fields might | |
| 1559 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot); | |
| 1560 } | |
| 1561 set_escape_state(pt, PointsToNode::ArgEscape); | |
| 0 | 1562 } |
| 1563 } | |
| 1564 } | |
| 1565 } | |
| 65 | 1566 if (copy_dependencies) |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1567 call_analyzer->copy_dependencies(_compile->dependencies()); |
| 0 | 1568 break; |
| 1569 } | |
| 1570 } | |
| 1571 | |
| 1572 default: | |
| 65 | 1573 // Fall-through here if not a Java method or no analyzer information |
| 1574 // or some other type of call, assume the worst case: all arguments | |
| 0 | 1575 // globally escape. |
| 1576 { | |
| 1577 // adjust escape state for outgoing arguments | |
| 1578 const TypeTuple * d = call->tf()->domain(); | |
| 1579 VectorSet ptset(Thread::current()->resource_area()); | |
| 1580 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { | |
| 1581 const Type* at = d->field_at(i); | |
| 1582 if (at->isa_oopptr() != NULL) { | |
| 65 | 1583 Node *arg = call->in(i)->uncast(); |
| 1584 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); | |
| 0 | 1585 ptset.Clear(); |
| 1586 PointsTo(ptset, arg, phase); | |
| 1587 for( VectorSetI j(&ptset); j.test(); ++j ) { | |
| 1588 uint pt = j.elem; | |
| 1589 set_escape_state(pt, PointsToNode::GlobalEscape); | |
| 1590 } | |
| 1591 } | |
| 1592 } | |
| 1593 } | |
| 1594 } | |
| 1595 } | |
| 1596 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1597 CallNode *call = resproj->in(0)->as_Call(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1598 uint call_idx = call->_idx; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1599 uint resproj_idx = resproj->_idx; |
| 0 | 1600 |
| 1601 switch (call->Opcode()) { | |
| 1602 case Op_Allocate: | |
| 1603 { | |
| 1604 Node *k = call->in(AllocateNode::KlassNode); | |
| 1605 const TypeKlassPtr *kt; | |
| 1606 if (k->Opcode() == Op_LoadKlass) { | |
| 1607 kt = k->as_Load()->type()->isa_klassptr(); | |
| 1608 } else { | |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1609 // Also works for DecodeN(LoadNKlass). |
| 0 | 1610 kt = k->as_Type()->type()->isa_klassptr(); |
| 1611 } | |
| 1612 assert(kt != NULL, "TypeKlassPtr required."); | |
| 1613 ciKlass* cik = kt->klass(); | |
| 1614 ciInstanceKlass* ciik = cik->as_instance_klass(); | |
| 1615 | |
| 65 | 1616 PointsToNode::EscapeState es; |
| 1617 uint edge_to; | |
| 0 | 1618 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || ciik->has_finalizer()) { |
| 65 | 1619 es = PointsToNode::GlobalEscape; |
| 1620 edge_to = _phantom_object; // Could not be worse | |
| 0 | 1621 } else { |
| 65 | 1622 es = PointsToNode::NoEscape; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1623 edge_to = call_idx; |
| 0 | 1624 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1625 set_escape_state(call_idx, es); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1626 add_pointsto_edge(resproj_idx, edge_to); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1627 _processed.set(resproj_idx); |
| 0 | 1628 break; |
| 1629 } | |
| 1630 | |
| 1631 case Op_AllocateArray: | |
| 1632 { | |
| 65 | 1633 int length = call->in(AllocateNode::ALength)->find_int_con(-1); |
| 1634 if (length < 0 || length > EliminateAllocationArraySizeLimit) { | |
| 1635 // Not scalar replaceable if the length is not constant or too big. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1636 ptnode_adr(call_idx)->_scalar_replaceable = false; |
| 65 | 1637 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1638 set_escape_state(call_idx, PointsToNode::NoEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1639 add_pointsto_edge(resproj_idx, call_idx); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1640 _processed.set(resproj_idx); |
| 0 | 1641 break; |
| 1642 } | |
| 1643 | |
| 1644 case Op_CallStaticJava: | |
| 1645 // For a static call, we know exactly what method is being called. | |
| 1646 // Use bytecode estimator to record whether the call's return value escapes | |
| 1647 { | |
| 65 | 1648 bool done = true; |
| 0 | 1649 const TypeTuple *r = call->tf()->range(); |
| 1650 const Type* ret_type = NULL; | |
| 1651 | |
| 1652 if (r->cnt() > TypeFunc::Parms) | |
| 1653 ret_type = r->field_at(TypeFunc::Parms); | |
| 1654 | |
| 1655 // Note: we use isa_ptr() instead of isa_oopptr() here because the | |
| 1656 // _multianewarray functions return a TypeRawPtr. | |
| 65 | 1657 if (ret_type == NULL || ret_type->isa_ptr() == NULL) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1658 _processed.set(resproj_idx); |
| 0 | 1659 break; // doesn't return a pointer type |
| 65 | 1660 } |
| 0 | 1661 ciMethod *meth = call->as_CallJava()->method(); |
| 65 | 1662 const TypeTuple * d = call->tf()->domain(); |
| 0 | 1663 if (meth == NULL) { |
| 1664 // not a Java method, assume global escape | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1665 set_escape_state(call_idx, PointsToNode::GlobalEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1666 add_pointsto_edge(resproj_idx, _phantom_object); |
| 0 | 1667 } else { |
| 65 | 1668 BCEscapeAnalyzer *call_analyzer = meth->get_bcea(); |
| 1669 bool copy_dependencies = false; | |
| 0 | 1670 |
| 65 | 1671 if (call_analyzer->is_return_allocated()) { |
| 1672 // Returns a newly allocated unescaped object, simply | |
| 1673 // update dependency information. | |
| 1674 // Mark it as NoEscape so that objects referenced by | |
| 1675 // it's fields will be marked as NoEscape at least. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1676 set_escape_state(call_idx, PointsToNode::NoEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1677 add_pointsto_edge(resproj_idx, call_idx); |
| 65 | 1678 copy_dependencies = true; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1679 } else if (call_analyzer->is_return_local()) { |
| 0 | 1680 // determine whether any arguments are returned |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1681 set_escape_state(call_idx, PointsToNode::NoEscape); |
| 0 | 1682 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { |
| 1683 const Type* at = d->field_at(i); | |
| 1684 | |
| 1685 if (at->isa_oopptr() != NULL) { | |
| 65 | 1686 Node *arg = call->in(i)->uncast(); |
| 0 | 1687 |
| 65 | 1688 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1689 PointsToNode *arg_esp = ptnode_adr(arg->_idx); |
| 65 | 1690 if (arg_esp->node_type() == PointsToNode::UnknownType) |
| 1691 done = false; | |
| 1692 else if (arg_esp->node_type() == PointsToNode::JavaObject) | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1693 add_pointsto_edge(resproj_idx, arg->_idx); |
| 0 | 1694 else |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1695 add_deferred_edge(resproj_idx, arg->_idx); |
| 0 | 1696 arg_esp->_hidden_alias = true; |
| 1697 } | |
| 1698 } | |
| 1699 } | |
| 65 | 1700 copy_dependencies = true; |
| 0 | 1701 } else { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1702 set_escape_state(call_idx, PointsToNode::GlobalEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1703 add_pointsto_edge(resproj_idx, _phantom_object); |
| 65 | 1704 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { |
| 1705 const Type* at = d->field_at(i); | |
| 1706 if (at->isa_oopptr() != NULL) { | |
| 1707 Node *arg = call->in(i)->uncast(); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1708 PointsToNode *arg_esp = ptnode_adr(arg->_idx); |
| 65 | 1709 arg_esp->_hidden_alias = true; |
| 1710 } | |
| 1711 } | |
| 0 | 1712 } |
| 65 | 1713 if (copy_dependencies) |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1714 call_analyzer->copy_dependencies(_compile->dependencies()); |
| 0 | 1715 } |
| 65 | 1716 if (done) |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1717 _processed.set(resproj_idx); |
| 0 | 1718 break; |
| 1719 } | |
| 1720 | |
| 1721 default: | |
| 1722 // Some other type of call, assume the worst case that the | |
| 1723 // returned value, if any, globally escapes. | |
| 1724 { | |
| 1725 const TypeTuple *r = call->tf()->range(); | |
| 1726 if (r->cnt() > TypeFunc::Parms) { | |
| 1727 const Type* ret_type = r->field_at(TypeFunc::Parms); | |
| 1728 | |
| 1729 // Note: we use isa_ptr() instead of isa_oopptr() here because the | |
| 1730 // _multianewarray functions return a TypeRawPtr. | |
| 1731 if (ret_type->isa_ptr() != NULL) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1732 set_escape_state(call_idx, PointsToNode::GlobalEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1733 add_pointsto_edge(resproj_idx, _phantom_object); |
| 0 | 1734 } |
| 1735 } | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1736 _processed.set(resproj_idx); |
| 0 | 1737 } |
| 1738 } | |
| 1739 } | |
| 1740 | |
| 65 | 1741 // Populate Connection Graph with Ideal nodes and create simple |
| 1742 // connection graph edges (do not need to check the node_type of inputs | |
| 1743 // or to call PointsTo() to walk the connection graph). | |
| 1744 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) { | |
| 1745 if (_processed.test(n->_idx)) | |
| 1746 return; // No need to redefine node's state. | |
| 1747 | |
| 1748 if (n->is_Call()) { | |
| 1749 // Arguments to allocation and locking don't escape. | |
| 1750 if (n->is_Allocate()) { | |
| 1751 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true); | |
| 1752 record_for_optimizer(n); | |
| 1753 } else if (n->is_Lock() || n->is_Unlock()) { | |
| 1754 // Put Lock and Unlock nodes on IGVN worklist to process them during | |
| 1755 // the first IGVN optimization when escape information is still available. | |
| 1756 record_for_optimizer(n); | |
| 1757 _processed.set(n->_idx); | |
| 1758 } else { | |
| 1759 // Have to process call's arguments first. | |
| 1760 PointsToNode::NodeType nt = PointsToNode::UnknownType; | |
| 1761 | |
| 1762 // Check if a call returns an object. | |
| 1763 const TypeTuple *r = n->as_Call()->tf()->range(); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1764 if (n->is_CallStaticJava() && r->cnt() > TypeFunc::Parms && |
| 65 | 1765 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) { |
| 1766 // Note: use isa_ptr() instead of isa_oopptr() here because | |
| 1767 // the _multianewarray functions return a TypeRawPtr. | |
| 1768 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { | |
| 1769 nt = PointsToNode::JavaObject; | |
| 1770 } | |
| 1771 } | |
| 1772 add_node(n, nt, PointsToNode::UnknownEscape, false); | |
| 1773 } | |
| 1774 return; | |
| 1775 } | |
| 1776 | |
| 1777 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because | |
| 1778 // ThreadLocal has RawPrt type. | |
| 1779 switch (n->Opcode()) { | |
| 1780 case Op_AddP: | |
| 1781 { | |
| 1782 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false); | |
| 1783 break; | |
| 1784 } | |
| 1785 case Op_CastX2P: | |
| 1786 { // "Unsafe" memory access. | |
| 1787 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); | |
| 1788 break; | |
| 1789 } | |
| 1790 case Op_CastPP: | |
| 1791 case Op_CheckCastPP: | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1792 case Op_EncodeP: |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1793 case Op_DecodeN: |
| 65 | 1794 { |
| 1795 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); | |
| 1796 int ti = n->in(1)->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1797 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); |
| 65 | 1798 if (nt == PointsToNode::UnknownType) { |
| 1799 _delayed_worklist.push(n); // Process it later. | |
| 1800 break; | |
| 1801 } else if (nt == PointsToNode::JavaObject) { | |
| 1802 add_pointsto_edge(n->_idx, ti); | |
| 1803 } else { | |
| 1804 add_deferred_edge(n->_idx, ti); | |
| 1805 } | |
| 1806 _processed.set(n->_idx); | |
| 1807 break; | |
| 1808 } | |
| 1809 case Op_ConP: | |
| 1810 { | |
| 1811 // assume all pointer constants globally escape except for null | |
| 1812 PointsToNode::EscapeState es; | |
| 1813 if (phase->type(n) == TypePtr::NULL_PTR) | |
| 1814 es = PointsToNode::NoEscape; | |
| 1815 else | |
| 1816 es = PointsToNode::GlobalEscape; | |
| 0 | 1817 |
| 65 | 1818 add_node(n, PointsToNode::JavaObject, es, true); |
| 1819 break; | |
| 1820 } | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1821 case Op_ConN: |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1822 { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1823 // assume all narrow oop constants globally escape except for null |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1824 PointsToNode::EscapeState es; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1825 if (phase->type(n) == TypeNarrowOop::NULL_PTR) |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1826 es = PointsToNode::NoEscape; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1827 else |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1828 es = PointsToNode::GlobalEscape; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1829 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1830 add_node(n, PointsToNode::JavaObject, es, true); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1831 break; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1832 } |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1833 case Op_CreateEx: |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1834 { |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1835 // assume that all exception objects globally escape |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1836 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1837 break; |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1838 } |
| 65 | 1839 case Op_LoadKlass: |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1840 case Op_LoadNKlass: |
| 65 | 1841 { |
| 1842 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); | |
| 1843 break; | |
| 1844 } | |
| 1845 case Op_LoadP: | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1846 case Op_LoadN: |
| 65 | 1847 { |
| 1848 const Type *t = phase->type(n); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1849 if (!t->isa_narrowoop() && t->isa_ptr() == NULL) { |
| 65 | 1850 _processed.set(n->_idx); |
| 1851 return; | |
| 1852 } | |
| 1853 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); | |
| 1854 break; | |
| 1855 } | |
| 1856 case Op_Parm: | |
| 1857 { | |
| 1858 _processed.set(n->_idx); // No need to redefine it state. | |
| 1859 uint con = n->as_Proj()->_con; | |
| 1860 if (con < TypeFunc::Parms) | |
| 1861 return; | |
| 1862 const Type *t = n->in(0)->as_Start()->_domain->field_at(con); | |
| 1863 if (t->isa_ptr() == NULL) | |
| 1864 return; | |
| 1865 // We have to assume all input parameters globally escape | |
| 1866 // (Note: passing 'false' since _processed is already set). | |
| 1867 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false); | |
| 1868 break; | |
| 1869 } | |
| 1870 case Op_Phi: | |
| 1871 { | |
| 1872 if (n->as_Phi()->type()->isa_ptr() == NULL) { | |
| 1873 // nothing to do if not an oop | |
| 1874 _processed.set(n->_idx); | |
| 1875 return; | |
| 1876 } | |
| 1877 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); | |
| 1878 uint i; | |
| 1879 for (i = 1; i < n->req() ; i++) { | |
| 1880 Node* in = n->in(i); | |
| 1881 if (in == NULL) | |
| 1882 continue; // ignore NULL | |
| 1883 in = in->uncast(); | |
| 1884 if (in->is_top() || in == n) | |
| 1885 continue; // ignore top or inputs which go back this node | |
| 1886 int ti = in->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1887 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); |
| 65 | 1888 if (nt == PointsToNode::UnknownType) { |
| 1889 break; | |
| 1890 } else if (nt == PointsToNode::JavaObject) { | |
| 1891 add_pointsto_edge(n->_idx, ti); | |
| 1892 } else { | |
| 1893 add_deferred_edge(n->_idx, ti); | |
| 1894 } | |
| 1895 } | |
| 1896 if (i >= n->req()) | |
| 1897 _processed.set(n->_idx); | |
| 1898 else | |
| 1899 _delayed_worklist.push(n); | |
| 1900 break; | |
| 1901 } | |
| 1902 case Op_Proj: | |
| 1903 { | |
| 1904 // we are only interested in the result projection from a call | |
| 1905 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { | |
| 1906 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); | |
| 1907 process_call_result(n->as_Proj(), phase); | |
| 1908 if (!_processed.test(n->_idx)) { | |
| 1909 // The call's result may need to be processed later if the call | |
| 1910 // returns it's argument and the argument is not processed yet. | |
| 1911 _delayed_worklist.push(n); | |
| 1912 } | |
| 1913 } else { | |
| 1914 _processed.set(n->_idx); | |
| 1915 } | |
| 1916 break; | |
| 1917 } | |
| 1918 case Op_Return: | |
| 1919 { | |
| 1920 if( n->req() > TypeFunc::Parms && | |
| 1921 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { | |
| 1922 // Treat Return value as LocalVar with GlobalEscape escape state. | |
| 1923 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false); | |
| 1924 int ti = n->in(TypeFunc::Parms)->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1925 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); |
| 65 | 1926 if (nt == PointsToNode::UnknownType) { |
| 1927 _delayed_worklist.push(n); // Process it later. | |
| 1928 break; | |
| 1929 } else if (nt == PointsToNode::JavaObject) { | |
| 1930 add_pointsto_edge(n->_idx, ti); | |
| 1931 } else { | |
| 1932 add_deferred_edge(n->_idx, ti); | |
| 1933 } | |
| 1934 } | |
| 1935 _processed.set(n->_idx); | |
| 1936 break; | |
| 1937 } | |
| 1938 case Op_StoreP: | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1939 case Op_StoreN: |
| 65 | 1940 { |
| 1941 const Type *adr_type = phase->type(n->in(MemNode::Address)); | |
|
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
168
diff
changeset
|
1942 adr_type = adr_type->make_ptr(); |
| 65 | 1943 if (adr_type->isa_oopptr()) { |
| 1944 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); | |
| 1945 } else { | |
| 1946 Node* adr = n->in(MemNode::Address); | |
| 1947 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL && | |
| 1948 adr->in(AddPNode::Address)->is_Proj() && | |
| 1949 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { | |
| 1950 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); | |
| 1951 // We are computing a raw address for a store captured | |
| 1952 // by an Initialize compute an appropriate address type. | |
| 1953 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); | |
| 1954 assert(offs != Type::OffsetBot, "offset must be a constant"); | |
| 1955 } else { | |
| 1956 _processed.set(n->_idx); | |
| 1957 return; | |
| 1958 } | |
| 1959 } | |
| 1960 break; | |
| 1961 } | |
| 1962 case Op_StorePConditional: | |
| 1963 case Op_CompareAndSwapP: | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1964 case Op_CompareAndSwapN: |
| 65 | 1965 { |
| 1966 const Type *adr_type = phase->type(n->in(MemNode::Address)); | |
|
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
168
diff
changeset
|
1967 adr_type = adr_type->make_ptr(); |
| 65 | 1968 if (adr_type->isa_oopptr()) { |
| 1969 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); | |
| 1970 } else { | |
| 1971 _processed.set(n->_idx); | |
| 1972 return; | |
| 1973 } | |
| 1974 break; | |
| 1975 } | |
| 1976 case Op_ThreadLocal: | |
| 1977 { | |
| 1978 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true); | |
| 1979 break; | |
| 1980 } | |
| 1981 default: | |
| 1982 ; | |
| 1983 // nothing to do | |
| 1984 } | |
| 1985 return; | |
| 1986 } | |
| 1987 | |
| 1988 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1989 uint n_idx = n->_idx; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1990 |
| 65 | 1991 // Don't set processed bit for AddP, LoadP, StoreP since |
| 1992 // they may need more then one pass to process. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1993 if (_processed.test(n_idx)) |
| 65 | 1994 return; // No need to redefine node's state. |
| 1995 | |
| 0 | 1996 if (n->is_Call()) { |
| 1997 CallNode *call = n->as_Call(); | |
| 1998 process_call_arguments(call, phase); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1999 _processed.set(n_idx); |
| 0 | 2000 return; |
| 2001 } | |
| 2002 | |
| 65 | 2003 switch (n->Opcode()) { |
| 0 | 2004 case Op_AddP: |
| 2005 { | |
| 65 | 2006 Node *base = get_addp_base(n); |
| 2007 // Create a field edge to this node from everything base could point to. | |
| 0 | 2008 VectorSet ptset(Thread::current()->resource_area()); |
| 2009 PointsTo(ptset, base, phase); | |
| 2010 for( VectorSetI i(&ptset); i.test(); ++i ) { | |
| 2011 uint pt = i.elem; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2012 add_field_edge(pt, n_idx, address_offset(n, phase)); |
| 65 | 2013 } |
| 2014 break; | |
| 2015 } | |
| 2016 case Op_CastX2P: | |
| 2017 { | |
| 2018 assert(false, "Op_CastX2P"); | |
| 2019 break; | |
| 2020 } | |
| 2021 case Op_CastPP: | |
| 2022 case Op_CheckCastPP: | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
2023 case Op_EncodeP: |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
2024 case Op_DecodeN: |
| 65 | 2025 { |
| 2026 int ti = n->in(1)->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2027 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2028 add_pointsto_edge(n_idx, ti); |
| 65 | 2029 } else { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2030 add_deferred_edge(n_idx, ti); |
| 65 | 2031 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2032 _processed.set(n_idx); |
| 65 | 2033 break; |
| 2034 } | |
| 2035 case Op_ConP: | |
| 2036 { | |
| 2037 assert(false, "Op_ConP"); | |
| 2038 break; | |
| 2039 } | |
| 163 | 2040 case Op_ConN: |
| 2041 { | |
| 2042 assert(false, "Op_ConN"); | |
| 2043 break; | |
| 2044 } | |
| 65 | 2045 case Op_CreateEx: |
| 2046 { | |
| 2047 assert(false, "Op_CreateEx"); | |
| 2048 break; | |
| 2049 } | |
| 2050 case Op_LoadKlass: | |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2051 case Op_LoadNKlass: |
| 65 | 2052 { |
| 2053 assert(false, "Op_LoadKlass"); | |
| 2054 break; | |
| 2055 } | |
| 2056 case Op_LoadP: | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
2057 case Op_LoadN: |
| 65 | 2058 { |
| 2059 const Type *t = phase->type(n); | |
| 2060 #ifdef ASSERT | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
2061 if (!t->isa_narrowoop() && t->isa_ptr() == NULL) |
| 65 | 2062 assert(false, "Op_LoadP"); |
| 2063 #endif | |
| 2064 | |
| 2065 Node* adr = n->in(MemNode::Address)->uncast(); | |
| 2066 const Type *adr_type = phase->type(adr); | |
| 2067 Node* adr_base; | |
| 2068 if (adr->is_AddP()) { | |
| 2069 adr_base = get_addp_base(adr); | |
| 2070 } else { | |
| 2071 adr_base = adr; | |
| 2072 } | |
| 2073 | |
| 2074 // For everything "adr_base" could point to, create a deferred edge from | |
| 2075 // this node to each field with the same offset. | |
| 2076 VectorSet ptset(Thread::current()->resource_area()); | |
| 2077 PointsTo(ptset, adr_base, phase); | |
| 2078 int offset = address_offset(adr, phase); | |
| 2079 for( VectorSetI i(&ptset); i.test(); ++i ) { | |
| 2080 uint pt = i.elem; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2081 add_deferred_edge_to_fields(n_idx, pt, offset); |
| 0 | 2082 } |
| 2083 break; | |
| 2084 } | |
| 2085 case Op_Parm: | |
| 2086 { | |
| 65 | 2087 assert(false, "Op_Parm"); |
| 0 | 2088 break; |
| 2089 } | |
| 2090 case Op_Phi: | |
| 2091 { | |
| 65 | 2092 #ifdef ASSERT |
| 2093 if (n->as_Phi()->type()->isa_ptr() == NULL) | |
| 2094 assert(false, "Op_Phi"); | |
| 2095 #endif | |
| 2096 for (uint i = 1; i < n->req() ; i++) { | |
| 2097 Node* in = n->in(i); | |
| 2098 if (in == NULL) | |
| 2099 continue; // ignore NULL | |
| 2100 in = in->uncast(); | |
| 2101 if (in->is_top() || in == n) | |
| 2102 continue; // ignore top or inputs which go back this node | |
| 2103 int ti = in->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2104 if (ptnode_adr(in->_idx)->node_type() == PointsToNode::JavaObject) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2105 add_pointsto_edge(n_idx, ti); |
| 65 | 2106 } else { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2107 add_deferred_edge(n_idx, ti); |
| 65 | 2108 } |
| 2109 } | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2110 _processed.set(n_idx); |
| 0 | 2111 break; |
| 2112 } | |
| 65 | 2113 case Op_Proj: |
| 0 | 2114 { |
| 65 | 2115 // we are only interested in the result projection from a call |
| 2116 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { | |
| 2117 process_call_result(n->as_Proj(), phase); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2118 assert(_processed.test(n_idx), "all call results should be processed"); |
| 65 | 2119 } else { |
| 2120 assert(false, "Op_Proj"); | |
| 2121 } | |
| 0 | 2122 break; |
| 2123 } | |
| 65 | 2124 case Op_Return: |
| 0 | 2125 { |
| 65 | 2126 #ifdef ASSERT |
| 2127 if( n->req() <= TypeFunc::Parms || | |
| 2128 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { | |
| 2129 assert(false, "Op_Return"); | |
| 0 | 2130 } |
| 65 | 2131 #endif |
| 2132 int ti = n->in(TypeFunc::Parms)->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2133 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2134 add_pointsto_edge(n_idx, ti); |
| 65 | 2135 } else { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2136 add_deferred_edge(n_idx, ti); |
| 65 | 2137 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2138 _processed.set(n_idx); |
| 0 | 2139 break; |
| 2140 } | |
| 2141 case Op_StoreP: | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
2142 case Op_StoreN: |
| 0 | 2143 case Op_StorePConditional: |
| 2144 case Op_CompareAndSwapP: | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
2145 case Op_CompareAndSwapN: |
| 0 | 2146 { |
| 2147 Node *adr = n->in(MemNode::Address); | |
|
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
168
diff
changeset
|
2148 const Type *adr_type = phase->type(adr)->make_ptr(); |
| 65 | 2149 #ifdef ASSERT |
| 0 | 2150 if (!adr_type->isa_oopptr()) |
| 65 | 2151 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP"); |
| 2152 #endif | |
| 0 | 2153 |
| 65 | 2154 assert(adr->is_AddP(), "expecting an AddP"); |
| 2155 Node *adr_base = get_addp_base(adr); | |
| 2156 Node *val = n->in(MemNode::ValueIn)->uncast(); | |
| 2157 // For everything "adr_base" could point to, create a deferred edge | |
| 2158 // to "val" from each field with the same offset. | |
| 0 | 2159 VectorSet ptset(Thread::current()->resource_area()); |
| 2160 PointsTo(ptset, adr_base, phase); | |
| 2161 for( VectorSetI i(&ptset); i.test(); ++i ) { | |
| 2162 uint pt = i.elem; | |
| 65 | 2163 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase)); |
| 0 | 2164 } |
| 2165 break; | |
| 2166 } | |
| 65 | 2167 case Op_ThreadLocal: |
| 0 | 2168 { |
| 65 | 2169 assert(false, "Op_ThreadLocal"); |
| 0 | 2170 break; |
| 2171 } | |
| 2172 default: | |
| 2173 ; | |
| 2174 // nothing to do | |
| 2175 } | |
| 2176 } | |
| 2177 | |
| 2178 #ifndef PRODUCT | |
| 2179 void ConnectionGraph::dump() { | |
| 2180 PhaseGVN *igvn = _compile->initial_gvn(); | |
| 2181 bool first = true; | |
| 2182 | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2183 uint size = nodes_size(); |
| 65 | 2184 for (uint ni = 0; ni < size; ni++) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2185 PointsToNode *ptn = ptnode_adr(ni); |
| 65 | 2186 PointsToNode::NodeType ptn_type = ptn->node_type(); |
| 2187 | |
| 2188 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL) | |
| 0 | 2189 continue; |
| 65 | 2190 PointsToNode::EscapeState es = escape_state(ptn->_node, igvn); |
| 2191 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) { | |
| 2192 if (first) { | |
| 2193 tty->cr(); | |
| 2194 tty->print("======== Connection graph for "); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2195 _compile->method()->print_short_name(); |
| 65 | 2196 tty->cr(); |
| 2197 first = false; | |
| 2198 } | |
| 2199 tty->print("%6d ", ni); | |
| 2200 ptn->dump(); | |
| 2201 // Print all locals which reference this allocation | |
| 2202 for (uint li = ni; li < size; li++) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2203 PointsToNode *ptn_loc = ptnode_adr(li); |
| 65 | 2204 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type(); |
| 2205 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL && | |
| 2206 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) { | |
| 2207 tty->print("%6d LocalVar [[%d]]", li, ni); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2208 ptnode_adr(li)->_node->dump(); |
| 0 | 2209 } |
| 2210 } | |
| 65 | 2211 if (Verbose) { |
| 2212 // Print all fields which reference this allocation | |
| 2213 for (uint i = 0; i < ptn->edge_count(); i++) { | |
| 2214 uint ei = ptn->edge_target(i); | |
| 2215 tty->print("%6d Field [[%d]]", ei, ni); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2216 ptnode_adr(ei)->_node->dump(); |
| 65 | 2217 } |
| 2218 } | |
| 2219 tty->cr(); | |
| 0 | 2220 } |
| 2221 } | |
| 2222 } | |
| 2223 #endif |