Mercurial > hg > truffle
annotate src/share/vm/opto/escape.cpp @ 247:02a35ad4adf8
6723160: Nightly failure: Error: meet not symmetric
Summary: Add missing _instance_id settings and other EA fixes.
Reviewed-by: rasbold
| author | kvn |
|---|---|
| date | Wed, 16 Jul 2008 16:04:39 -0700 |
| parents | 4a4c365f777d |
| children | b0fe4deeb9fb |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 196 | 2 * Copyright 2005-2008 Sun Microsystems, Inc. All Rights Reserved. |
| 0 | 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 } | |
| 247 | 720 if (result->is_Phi()) { |
| 65 | 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) { | |
| 247 | 725 // Create a new Phi with the specified alias index type. |
| 65 | 726 result = split_memory_phi(mphi, alias_idx, orig_phis, phase); |
| 247 | 727 } else if (!is_instance) { |
| 728 // Push all non-instance Phis on the orig_phis worklist to update inputs | |
| 729 // during Phase 4 if needed. | |
| 730 orig_phis.append_if_missing(mphi); | |
| 65 | 731 } |
| 732 } | |
| 733 // the result is either MemNode, PhiNode, InitializeNode. | |
| 734 return result; | |
| 735 } | |
| 736 | |
| 737 | |
| 0 | 738 // |
| 739 // Convert the types of unescaped object to instance types where possible, | |
| 740 // propagate the new type information through the graph, and update memory | |
| 741 // edges and MergeMem inputs to reflect the new type. | |
| 742 // | |
| 743 // We start with allocations (and calls which may be allocations) on alloc_worklist. | |
| 744 // The processing is done in 4 phases: | |
| 745 // | |
| 746 // Phase 1: Process possible allocations from alloc_worklist. Create instance | |
| 747 // types for the CheckCastPP for allocations where possible. | |
| 748 // Propagate the the new types through users as follows: | |
| 749 // casts and Phi: push users on alloc_worklist | |
| 750 // AddP: cast Base and Address inputs to the instance type | |
| 751 // push any AddP users on alloc_worklist and push any memnode | |
| 752 // users onto memnode_worklist. | |
| 753 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and | |
| 754 // search the Memory chain for a store with the appropriate type | |
| 755 // address type. If a Phi is found, create a new version with | |
| 756 // the approriate memory slices from each of the Phi inputs. | |
| 757 // For stores, process the users as follows: | |
| 758 // MemNode: push on memnode_worklist | |
| 759 // MergeMem: push on mergemem_worklist | |
| 760 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice | |
| 761 // moving the first node encountered of each instance type to the | |
| 762 // the input corresponding to its alias index. | |
| 763 // appropriate memory slice. | |
| 764 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. | |
| 765 // | |
| 766 // In the following example, the CheckCastPP nodes are the cast of allocation | |
| 767 // results and the allocation of node 29 is unescaped and eligible to be an | |
| 768 // instance type. | |
| 769 // | |
| 770 // We start with: | |
| 771 // | |
| 772 // 7 Parm #memory | |
| 773 // 10 ConI "12" | |
| 774 // 19 CheckCastPP "Foo" | |
| 775 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 | |
| 776 // 29 CheckCastPP "Foo" | |
| 777 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 | |
| 778 // | |
| 779 // 40 StoreP 25 7 20 ... alias_index=4 | |
| 780 // 50 StoreP 35 40 30 ... alias_index=4 | |
| 781 // 60 StoreP 45 50 20 ... alias_index=4 | |
| 782 // 70 LoadP _ 60 30 ... alias_index=4 | |
| 783 // 80 Phi 75 50 60 Memory alias_index=4 | |
| 784 // 90 LoadP _ 80 30 ... alias_index=4 | |
| 785 // 100 LoadP _ 80 20 ... alias_index=4 | |
| 786 // | |
| 787 // | |
| 788 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 | |
| 789 // and creating a new alias index for node 30. This gives: | |
| 790 // | |
| 791 // 7 Parm #memory | |
| 792 // 10 ConI "12" | |
| 793 // 19 CheckCastPP "Foo" | |
| 794 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 | |
| 795 // 29 CheckCastPP "Foo" iid=24 | |
| 796 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 | |
| 797 // | |
| 798 // 40 StoreP 25 7 20 ... alias_index=4 | |
| 799 // 50 StoreP 35 40 30 ... alias_index=6 | |
| 800 // 60 StoreP 45 50 20 ... alias_index=4 | |
| 801 // 70 LoadP _ 60 30 ... alias_index=6 | |
| 802 // 80 Phi 75 50 60 Memory alias_index=4 | |
| 803 // 90 LoadP _ 80 30 ... alias_index=6 | |
| 804 // 100 LoadP _ 80 20 ... alias_index=4 | |
| 805 // | |
| 806 // In phase 2, new memory inputs are computed for the loads and stores, | |
| 807 // And a new version of the phi is created. In phase 4, the inputs to | |
| 808 // node 80 are updated and then the memory nodes are updated with the | |
| 809 // values computed in phase 2. This results in: | |
| 810 // | |
| 811 // 7 Parm #memory | |
| 812 // 10 ConI "12" | |
| 813 // 19 CheckCastPP "Foo" | |
| 814 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 | |
| 815 // 29 CheckCastPP "Foo" iid=24 | |
| 816 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 | |
| 817 // | |
| 818 // 40 StoreP 25 7 20 ... alias_index=4 | |
| 819 // 50 StoreP 35 7 30 ... alias_index=6 | |
| 820 // 60 StoreP 45 40 20 ... alias_index=4 | |
| 821 // 70 LoadP _ 50 30 ... alias_index=6 | |
| 822 // 80 Phi 75 40 60 Memory alias_index=4 | |
| 823 // 120 Phi 75 50 50 Memory alias_index=6 | |
| 824 // 90 LoadP _ 120 30 ... alias_index=6 | |
| 825 // 100 LoadP _ 80 20 ... alias_index=4 | |
| 826 // | |
| 827 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { | |
| 828 GrowableArray<Node *> memnode_worklist; | |
| 829 GrowableArray<Node *> mergemem_worklist; | |
| 830 GrowableArray<PhiNode *> orig_phis; | |
| 831 PhaseGVN *igvn = _compile->initial_gvn(); | |
| 832 uint new_index_start = (uint) _compile->num_alias_types(); | |
| 833 VectorSet visited(Thread::current()->resource_area()); | |
| 834 VectorSet ptset(Thread::current()->resource_area()); | |
| 835 | |
| 65 | 836 |
| 837 // Phase 1: Process possible allocations from alloc_worklist. | |
| 838 // Create instance types for the CheckCastPP for allocations where possible. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
839 // |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
840 // (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
|
841 // 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
|
842 // see the comment in find_second_addp().) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
843 // |
| 0 | 844 while (alloc_worklist.length() != 0) { |
| 845 Node *n = alloc_worklist.pop(); | |
| 846 uint ni = n->_idx; | |
| 65 | 847 const TypeOopPtr* tinst = NULL; |
| 0 | 848 if (n->is_Call()) { |
| 849 CallNode *alloc = n->as_Call(); | |
| 850 // copy escape information to call node | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
851 PointsToNode* ptn = ptnode_adr(alloc->_idx); |
| 0 | 852 PointsToNode::EscapeState es = escape_state(alloc, igvn); |
| 65 | 853 // We have an allocation or call which returns a Java object, |
| 854 // see if it is unescaped. | |
| 855 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable) | |
| 0 | 856 continue; |
|
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
38
diff
changeset
|
857 if (alloc->is_Allocate()) { |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
38
diff
changeset
|
858 // 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
|
859 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
|
860 } |
| 65 | 861 // find CheckCastPP of call return value |
| 862 n = alloc->result_cast(); | |
| 863 if (n == NULL || // No uses accept Initialize or | |
| 864 !n->is_CheckCastPP()) // not unique CheckCastPP. | |
| 865 continue; | |
| 866 // The inline code for Object.clone() casts the allocation result to | |
| 247 | 867 // java.lang.Object and then to the actual type of the allocated |
| 65 | 868 // object. Detect this case and use the second cast. |
| 247 | 869 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when |
| 870 // the allocation result is cast to java.lang.Object and then | |
| 871 // to the actual Array type. | |
| 65 | 872 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL |
| 247 | 873 && (alloc->is_AllocateArray() || |
| 874 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { | |
| 65 | 875 Node *cast2 = NULL; |
| 876 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 877 Node *use = n->fast_out(i); | |
| 878 if (use->is_CheckCastPP()) { | |
| 879 cast2 = use; | |
| 880 break; | |
| 881 } | |
| 882 } | |
| 883 if (cast2 != NULL) { | |
| 884 n = cast2; | |
| 885 } else { | |
| 886 continue; | |
| 887 } | |
| 888 } | |
| 889 set_escape_state(n->_idx, es); | |
| 247 | 890 // in order for an object to be scalar-replaceable, it must be: |
| 65 | 891 // - a direct allocation (not a call returning an object) |
| 892 // - non-escaping | |
| 893 // - eligible to be a unique type | |
| 894 // - not determined to be ineligible by escape analysis | |
| 0 | 895 set_map(alloc->_idx, n); |
| 896 set_map(n->_idx, alloc); | |
| 65 | 897 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); |
| 898 if (t == NULL) | |
| 0 | 899 continue; // not a TypeInstPtr |
| 247 | 900 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni); |
| 0 | 901 igvn->hash_delete(n); |
| 902 igvn->set_type(n, tinst); | |
| 903 n->raise_bottom_type(tinst); | |
| 904 igvn->hash_insert(n); | |
| 65 | 905 record_for_optimizer(n); |
| 906 if (alloc->is_Allocate() && ptn->_scalar_replaceable && | |
| 907 (t->isa_instptr() || t->isa_aryptr())) { | |
| 163 | 908 |
| 909 // First, put on the worklist all Field edges from Connection Graph | |
| 910 // which is more accurate then putting immediate users from Ideal Graph. | |
| 911 for (uint e = 0; e < ptn->edge_count(); e++) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
912 Node *use = ptnode_adr(ptn->edge_target(e))->_node; |
| 163 | 913 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(), |
| 914 "only AddP nodes are Field edges in CG"); | |
| 915 if (use->outcnt() > 0) { // Don't process dead nodes | |
| 916 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); | |
| 917 if (addp2 != NULL) { | |
| 918 assert(alloc->is_AllocateArray(),"array allocation was expected"); | |
| 919 alloc_worklist.append_if_missing(addp2); | |
| 920 } | |
| 921 alloc_worklist.append_if_missing(use); | |
| 922 } | |
| 923 } | |
| 924 | |
| 65 | 925 // An allocation may have an Initialize which has raw stores. Scan |
| 926 // the users of the raw allocation result and push AddP users | |
| 927 // on alloc_worklist. | |
| 928 Node *raw_result = alloc->proj_out(TypeFunc::Parms); | |
| 929 assert (raw_result != NULL, "must have an allocation result"); | |
| 930 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { | |
| 931 Node *use = raw_result->fast_out(i); | |
| 932 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes | |
| 933 Node* addp2 = find_second_addp(use, raw_result); | |
| 934 if (addp2 != NULL) { | |
| 935 assert(alloc->is_AllocateArray(),"array allocation was expected"); | |
| 936 alloc_worklist.append_if_missing(addp2); | |
| 937 } | |
| 938 alloc_worklist.append_if_missing(use); | |
| 939 } else if (use->is_Initialize()) { | |
| 940 memnode_worklist.append_if_missing(use); | |
| 941 } | |
| 942 } | |
| 943 } | |
| 0 | 944 } else if (n->is_AddP()) { |
| 945 ptset.Clear(); | |
| 65 | 946 PointsTo(ptset, get_addp_base(n), igvn); |
| 0 | 947 assert(ptset.Size() == 1, "AddP address is unique"); |
| 65 | 948 uint elem = ptset.getelem(); // Allocation node's index |
| 949 if (elem == _phantom_object) | |
| 950 continue; // Assume the value was set outside this method. | |
| 951 Node *base = get_map(elem); // CheckCastPP node | |
| 0 | 952 split_AddP(n, base, igvn); |
| 65 | 953 tinst = igvn->type(base)->isa_oopptr(); |
| 954 } else if (n->is_Phi() || | |
| 955 n->is_CheckCastPP() || | |
|
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
956 n->is_EncodeP() || |
|
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
957 n->is_DecodeN() || |
| 65 | 958 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { |
| 0 | 959 if (visited.test_set(n->_idx)) { |
| 960 assert(n->is_Phi(), "loops only through Phi's"); | |
| 961 continue; // already processed | |
| 962 } | |
| 963 ptset.Clear(); | |
| 964 PointsTo(ptset, n, igvn); | |
| 965 if (ptset.Size() == 1) { | |
| 65 | 966 uint elem = ptset.getelem(); // Allocation node's index |
| 967 if (elem == _phantom_object) | |
| 968 continue; // Assume the value was set outside this method. | |
| 969 Node *val = get_map(elem); // CheckCastPP node | |
| 0 | 970 TypeNode *tn = n->as_Type(); |
| 65 | 971 tinst = igvn->type(val)->isa_oopptr(); |
| 223 | 972 assert(tinst != NULL && tinst->is_known_instance() && |
| 973 (uint)tinst->instance_id() == elem , "instance type expected."); | |
| 163 | 974 |
| 975 const Type *tn_type = igvn->type(tn); | |
| 223 | 976 const TypeOopPtr *tn_t; |
| 977 if (tn_type->isa_narrowoop()) { | |
| 978 tn_t = tn_type->make_ptr()->isa_oopptr(); | |
| 979 } else { | |
| 980 tn_t = tn_type->isa_oopptr(); | |
| 981 } | |
| 0 | 982 |
| 65 | 983 if (tn_t != NULL && |
| 223 | 984 tinst->cast_to_instance_id(TypeOopPtr::InstanceBot)->higher_equal(tn_t)) { |
| 163 | 985 if (tn_type->isa_narrowoop()) { |
| 986 tn_type = tinst->make_narrowoop(); | |
| 987 } else { | |
| 988 tn_type = tinst; | |
| 989 } | |
| 0 | 990 igvn->hash_delete(tn); |
| 163 | 991 igvn->set_type(tn, tn_type); |
| 992 tn->set_type(tn_type); | |
| 0 | 993 igvn->hash_insert(tn); |
| 65 | 994 record_for_optimizer(n); |
| 0 | 995 } |
| 996 } | |
| 997 } else { | |
| 998 continue; | |
| 999 } | |
| 1000 // push users on appropriate worklist | |
| 1001 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1002 Node *use = n->fast_out(i); | |
| 1003 if(use->is_Mem() && use->in(MemNode::Address) == n) { | |
| 65 | 1004 memnode_worklist.append_if_missing(use); |
| 1005 } else if (use->is_Initialize()) { | |
| 1006 memnode_worklist.append_if_missing(use); | |
| 1007 } else if (use->is_MergeMem()) { | |
| 1008 mergemem_worklist.append_if_missing(use); | |
| 1009 } else if (use->is_Call() && tinst != NULL) { | |
| 1010 // Look for MergeMem nodes for calls which reference unique allocation | |
| 1011 // (through CheckCastPP nodes) even for debug info. | |
| 1012 Node* m = use->in(TypeFunc::Memory); | |
| 1013 uint iid = tinst->instance_id(); | |
| 1014 while (m->is_Proj() && m->in(0)->is_Call() && | |
| 1015 m->in(0) != use && !m->in(0)->_idx != iid) { | |
| 1016 m = m->in(0)->in(TypeFunc::Memory); | |
| 1017 } | |
| 1018 if (m->is_MergeMem()) { | |
| 1019 mergemem_worklist.append_if_missing(m); | |
| 1020 } | |
| 1021 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes | |
| 1022 Node* addp2 = find_second_addp(use, n); | |
| 1023 if (addp2 != NULL) { | |
| 1024 alloc_worklist.append_if_missing(addp2); | |
| 1025 } | |
| 1026 alloc_worklist.append_if_missing(use); | |
| 1027 } else if (use->is_Phi() || | |
| 1028 use->is_CheckCastPP() || | |
|
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1029 use->is_EncodeP() || |
|
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1030 use->is_DecodeN() || |
| 65 | 1031 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { |
| 1032 alloc_worklist.append_if_missing(use); | |
| 0 | 1033 } |
| 1034 } | |
| 1035 | |
| 1036 } | |
| 65 | 1037 // New alias types were created in split_AddP(). |
| 0 | 1038 uint new_index_end = (uint) _compile->num_alias_types(); |
| 1039 | |
| 1040 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and | |
| 1041 // compute new values for Memory inputs (the Memory inputs are not | |
| 1042 // actually updated until phase 4.) | |
| 1043 if (memnode_worklist.length() == 0) | |
| 1044 return; // nothing to do | |
| 1045 | |
| 1046 while (memnode_worklist.length() != 0) { | |
| 1047 Node *n = memnode_worklist.pop(); | |
| 65 | 1048 if (visited.test_set(n->_idx)) |
| 1049 continue; | |
| 0 | 1050 if (n->is_Phi()) { |
| 1051 assert(n->as_Phi()->adr_type() != TypePtr::BOTTOM, "narrow memory slice required"); | |
| 1052 // we don't need to do anything, but the users must be pushed if we haven't processed | |
| 1053 // this Phi before | |
| 65 | 1054 } else if (n->is_Initialize()) { |
| 1055 // we don't need to do anything, but the users of the memory projection must be pushed | |
| 1056 n = n->as_Initialize()->proj_out(TypeFunc::Memory); | |
| 1057 if (n == NULL) | |
| 0 | 1058 continue; |
| 1059 } else { | |
| 1060 assert(n->is_Mem(), "memory node required."); | |
| 1061 Node *addr = n->in(MemNode::Address); | |
| 65 | 1062 assert(addr->is_AddP(), "AddP required"); |
| 0 | 1063 const Type *addr_t = igvn->type(addr); |
| 1064 if (addr_t == Type::TOP) | |
| 1065 continue; | |
| 1066 assert (addr_t->isa_ptr() != NULL, "pointer type required."); | |
| 1067 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); | |
| 65 | 1068 assert ((uint)alias_idx < new_index_end, "wrong alias index"); |
| 1069 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
|
1070 if (_compile->failing()) { |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
1071 return; |
|
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
1072 } |
| 65 | 1073 if (mem != n->in(MemNode::Memory)) { |
| 0 | 1074 set_map(n->_idx, mem); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1075 ptnode_adr(n->_idx)->_node = n; |
| 65 | 1076 } |
| 0 | 1077 if (n->is_Load()) { |
| 1078 continue; // don't push users | |
| 1079 } else if (n->is_LoadStore()) { | |
| 1080 // get the memory projection | |
| 1081 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1082 Node *use = n->fast_out(i); | |
| 1083 if (use->Opcode() == Op_SCMemProj) { | |
| 1084 n = use; | |
| 1085 break; | |
| 1086 } | |
| 1087 } | |
| 1088 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); | |
| 1089 } | |
| 1090 } | |
| 1091 // push user on appropriate worklist | |
| 1092 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1093 Node *use = n->fast_out(i); | |
| 1094 if (use->is_Phi()) { | |
| 65 | 1095 memnode_worklist.append_if_missing(use); |
| 0 | 1096 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) { |
| 65 | 1097 memnode_worklist.append_if_missing(use); |
| 1098 } else if (use->is_Initialize()) { | |
| 1099 memnode_worklist.append_if_missing(use); | |
| 0 | 1100 } else if (use->is_MergeMem()) { |
| 65 | 1101 mergemem_worklist.append_if_missing(use); |
| 0 | 1102 } |
| 1103 } | |
| 1104 } | |
| 1105 | |
| 65 | 1106 // Phase 3: Process MergeMem nodes from mergemem_worklist. |
| 1107 // Walk each memory moving the first node encountered of each | |
| 1108 // instance type to the the input corresponding to its alias index. | |
| 0 | 1109 while (mergemem_worklist.length() != 0) { |
| 1110 Node *n = mergemem_worklist.pop(); | |
| 1111 assert(n->is_MergeMem(), "MergeMem node required."); | |
| 65 | 1112 if (visited.test_set(n->_idx)) |
| 1113 continue; | |
| 0 | 1114 MergeMemNode *nmm = n->as_MergeMem(); |
| 1115 // Note: we don't want to use MergeMemStream here because we only want to | |
| 65 | 1116 // scan inputs which exist at the start, not ones we add during processing. |
| 0 | 1117 uint nslices = nmm->req(); |
| 1118 igvn->hash_delete(nmm); | |
| 1119 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { | |
| 65 | 1120 Node* mem = nmm->in(i); |
| 1121 Node* cur = NULL; | |
| 0 | 1122 if (mem == NULL || mem->is_top()) |
| 1123 continue; | |
| 1124 while (mem->is_Mem()) { | |
| 1125 const Type *at = igvn->type(mem->in(MemNode::Address)); | |
| 1126 if (at != Type::TOP) { | |
| 1127 assert (at->isa_ptr() != NULL, "pointer type required."); | |
| 1128 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); | |
| 1129 if (idx == i) { | |
| 1130 if (cur == NULL) | |
| 1131 cur = mem; | |
| 1132 } else { | |
| 1133 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { | |
| 1134 nmm->set_memory_at(idx, mem); | |
| 1135 } | |
| 1136 } | |
| 1137 } | |
| 1138 mem = mem->in(MemNode::Memory); | |
| 1139 } | |
| 1140 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); | |
| 65 | 1141 // Find any instance of the current type if we haven't encountered |
| 1142 // a value of the instance along the chain. | |
| 1143 for (uint ni = new_index_start; ni < new_index_end; ni++) { | |
| 1144 if((uint)_compile->get_general_index(ni) == i) { | |
| 1145 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); | |
| 1146 if (nmm->is_empty_memory(m)) { | |
| 1147 Node* result = find_inst_mem(mem, ni, orig_phis, igvn); | |
| 1148 if (_compile->failing()) { | |
| 1149 return; | |
| 1150 } | |
| 1151 nmm->set_memory_at(ni, result); | |
| 1152 } | |
| 1153 } | |
| 1154 } | |
| 1155 } | |
| 1156 // Find the rest of instances values | |
| 1157 for (uint ni = new_index_start; ni < new_index_end; ni++) { | |
| 1158 const TypeOopPtr *tinst = igvn->C->get_adr_type(ni)->isa_oopptr(); | |
| 1159 Node* result = step_through_mergemem(nmm, ni, tinst); | |
| 1160 if (result == nmm->base_memory()) { | |
| 1161 // Didn't find instance memory, search through general slice recursively. | |
| 1162 result = nmm->memory_at(igvn->C->get_general_index(ni)); | |
| 1163 result = find_inst_mem(result, ni, orig_phis, igvn); | |
| 1164 if (_compile->failing()) { | |
| 1165 return; | |
| 1166 } | |
| 1167 nmm->set_memory_at(ni, result); | |
| 1168 } | |
| 1169 } | |
| 1170 igvn->hash_insert(nmm); | |
| 1171 record_for_optimizer(nmm); | |
| 1172 | |
| 1173 // Propagate new memory slices to following MergeMem nodes. | |
| 1174 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1175 Node *use = n->fast_out(i); | |
| 1176 if (use->is_Call()) { | |
| 1177 CallNode* in = use->as_Call(); | |
| 1178 if (in->proj_out(TypeFunc::Memory) != NULL) { | |
| 1179 Node* m = in->proj_out(TypeFunc::Memory); | |
| 1180 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { | |
| 1181 Node* mm = m->fast_out(j); | |
| 1182 if (mm->is_MergeMem()) { | |
| 1183 mergemem_worklist.append_if_missing(mm); | |
| 1184 } | |
| 1185 } | |
| 1186 } | |
| 1187 if (use->is_Allocate()) { | |
| 1188 use = use->as_Allocate()->initialization(); | |
| 1189 if (use == NULL) { | |
| 1190 continue; | |
| 1191 } | |
| 1192 } | |
| 1193 } | |
| 1194 if (use->is_Initialize()) { | |
| 1195 InitializeNode* in = use->as_Initialize(); | |
| 1196 if (in->proj_out(TypeFunc::Memory) != NULL) { | |
| 1197 Node* m = in->proj_out(TypeFunc::Memory); | |
| 1198 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { | |
| 1199 Node* mm = m->fast_out(j); | |
| 1200 if (mm->is_MergeMem()) { | |
| 1201 mergemem_worklist.append_if_missing(mm); | |
| 0 | 1202 } |
| 1203 } | |
| 1204 } | |
| 1205 } | |
| 1206 } | |
| 1207 } | |
| 1208 | |
| 65 | 1209 // Phase 4: Update the inputs of non-instance memory Phis and |
| 1210 // the Memory input of memnodes | |
| 0 | 1211 // First update the inputs of any non-instance Phi's from |
| 1212 // which we split out an instance Phi. Note we don't have | |
| 1213 // to recursively process Phi's encounted on the input memory | |
| 1214 // chains as is done in split_memory_phi() since they will | |
| 1215 // also be processed here. | |
| 247 | 1216 for (int j = 0; j < orig_phis.length(); j++) { |
| 1217 PhiNode *phi = orig_phis.at(j); | |
| 0 | 1218 int alias_idx = _compile->get_alias_index(phi->adr_type()); |
| 1219 igvn->hash_delete(phi); | |
| 1220 for (uint i = 1; i < phi->req(); i++) { | |
| 1221 Node *mem = phi->in(i); | |
| 65 | 1222 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn); |
| 1223 if (_compile->failing()) { | |
| 1224 return; | |
| 1225 } | |
| 0 | 1226 if (mem != new_mem) { |
| 1227 phi->set_req(i, new_mem); | |
| 1228 } | |
| 1229 } | |
| 1230 igvn->hash_insert(phi); | |
| 1231 record_for_optimizer(phi); | |
| 1232 } | |
| 1233 | |
| 1234 // Update the memory inputs of MemNodes with the value we computed | |
| 1235 // in Phase 2. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1236 for (uint i = 0; i < nodes_size(); i++) { |
| 0 | 1237 Node *nmem = get_map(i); |
| 1238 if (nmem != NULL) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1239 Node *n = ptnode_adr(i)->_node; |
| 0 | 1240 if (n != NULL && n->is_Mem()) { |
| 1241 igvn->hash_delete(n); | |
| 1242 n->set_req(MemNode::Memory, nmem); | |
| 1243 igvn->hash_insert(n); | |
| 1244 record_for_optimizer(n); | |
| 1245 } | |
| 1246 } | |
| 1247 } | |
| 1248 } | |
| 1249 | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1250 bool ConnectionGraph::has_candidates(Compile *C) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1251 // 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
|
1252 // are represented by ideal Macro nodes. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1253 int cnt = C->macro_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1254 for( int i=0; i < cnt; i++ ) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1255 Node *n = C->macro_node(i); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1256 if ( n->is_Allocate() ) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1257 return true; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1258 if( n->is_Lock() ) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1259 Node* obj = n->as_Lock()->obj_node()->uncast(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1260 if( !(obj->is_Parm() || obj->is_Con()) ) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1261 return true; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1262 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1263 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1264 return false; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1265 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1266 |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1267 bool ConnectionGraph::compute_escape() { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1268 Compile* C = _compile; |
| 65 | 1269 |
| 163 | 1270 // 1. Populate Connection Graph (CG) with Ideal nodes. |
| 65 | 1271 |
| 1272 Unique_Node_List worklist_init; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1273 worklist_init.map(C->unique(), NULL); // preallocate space |
| 65 | 1274 |
| 1275 // Initialize worklist | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1276 if (C->root() != NULL) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1277 worklist_init.push(C->root()); |
| 65 | 1278 } |
| 1279 | |
| 1280 GrowableArray<int> cg_worklist; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1281 PhaseGVN* igvn = C->initial_gvn(); |
| 65 | 1282 bool has_allocations = false; |
| 1283 | |
| 1284 // Push all useful nodes onto CG list and set their type. | |
| 1285 for( uint next = 0; next < worklist_init.size(); ++next ) { | |
| 1286 Node* n = worklist_init.at(next); | |
| 1287 record_for_escape_analysis(n, igvn); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1288 // Only allocations and java static calls results are checked |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1289 // for an escape status. See process_call_result() below. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1290 if (n->is_Allocate() || n->is_CallStaticJava() && |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1291 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) { |
| 65 | 1292 has_allocations = true; |
| 1293 } | |
| 1294 if(n->is_AddP()) | |
| 1295 cg_worklist.append(n->_idx); | |
| 1296 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 1297 Node* m = n->fast_out(i); // Get user | |
| 1298 worklist_init.push(m); | |
| 1299 } | |
| 1300 } | |
| 0 | 1301 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1302 if (!has_allocations) { |
| 65 | 1303 _collecting = false; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1304 return false; // Nothing to do. |
| 65 | 1305 } |
| 1306 | |
| 1307 // 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
|
1308 uint delayed_size = _delayed_worklist.size(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1309 for( uint next = 0; next < delayed_size; ++next ) { |
| 65 | 1310 Node* n = _delayed_worklist.at(next); |
| 1311 build_connection_graph(n, igvn); | |
| 1312 } | |
| 0 | 1313 |
| 65 | 1314 // 3. Pass to create fields edges (Allocate -F-> AddP). |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1315 uint cg_length = cg_worklist.length(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1316 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1317 int ni = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1318 build_connection_graph(ptnode_adr(ni)->_node, igvn); |
| 65 | 1319 } |
| 1320 | |
| 1321 cg_worklist.clear(); | |
| 1322 cg_worklist.append(_phantom_object); | |
| 1323 | |
| 1324 // 4. Build Connection Graph which need | |
| 1325 // to walk the connection graph. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1326 for (uint ni = 0; ni < nodes_size(); ni++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1327 PointsToNode* ptn = ptnode_adr(ni); |
| 65 | 1328 Node *n = ptn->_node; |
| 1329 if (n != NULL) { // Call, AddP, LoadP, StoreP | |
| 1330 build_connection_graph(n, igvn); | |
| 1331 if (ptn->node_type() != PointsToNode::UnknownType) | |
| 1332 cg_worklist.append(n->_idx); // Collect CG nodes | |
| 1333 } | |
| 0 | 1334 } |
| 1335 | |
| 1336 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
|
1337 GrowableArray<uint> deferred_edges; |
|
a6cb86dd209b
6681577: PIT: some VM tests fails with -XX:+AggressiveOpts in 6u5p b01
kvn
parents:
65
diff
changeset
|
1338 VectorSet visited(Thread::current()->resource_area()); |
| 0 | 1339 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1340 // 5. Remove deferred edges from the graph and collect |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1341 // information needed for type splitting. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1342 cg_length = cg_worklist.length(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1343 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1344 int ni = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1345 PointsToNode* ptn = ptnode_adr(ni); |
| 0 | 1346 PointsToNode::NodeType nt = ptn->node_type(); |
| 1347 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
|
1348 remove_deferred(ni, &deferred_edges, &visited); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1349 Node *n = ptn->_node; |
| 0 | 1350 if (n->is_AddP()) { |
| 65 | 1351 // If this AddP computes an address which may point to more that one |
| 163 | 1352 // object or more then one field (array's element), nothing the address |
| 1353 // points to can be scalar replaceable. | |
| 65 | 1354 Node *base = get_addp_base(n); |
| 0 | 1355 ptset.Clear(); |
| 1356 PointsTo(ptset, base, igvn); | |
| 163 | 1357 if (ptset.Size() > 1 || |
| 1358 (ptset.Size() != 0 && ptn->offset() == Type::OffsetBot)) { | |
| 0 | 1359 for( VectorSetI j(&ptset); j.test(); ++j ) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1360 ptnode_adr(j.elem)->_scalar_replaceable = false; |
| 0 | 1361 } |
| 1362 } | |
| 1363 } | |
| 1364 } | |
| 1365 } | |
| 65 | 1366 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1367 // 6. Propagate escape states. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1368 GrowableArray<int> worklist; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1369 bool has_non_escaping_obj = false; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1370 |
| 0 | 1371 // push all GlobalEscape nodes on the worklist |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1372 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1373 int nk = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1374 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1375 worklist.push(nk); |
| 0 | 1376 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1377 // mark all nodes reachable from GlobalEscape nodes |
| 0 | 1378 while(worklist.length() > 0) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1379 PointsToNode* ptn = ptnode_adr(worklist.pop()); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1380 uint e_cnt = ptn->edge_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1381 for (uint ei = 0; ei < e_cnt; ei++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1382 uint npi = ptn->edge_target(ei); |
| 0 | 1383 PointsToNode *np = ptnode_adr(npi); |
| 65 | 1384 if (np->escape_state() < PointsToNode::GlobalEscape) { |
| 0 | 1385 np->set_escape_state(PointsToNode::GlobalEscape); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1386 worklist.push(npi); |
| 0 | 1387 } |
| 1388 } | |
| 1389 } | |
| 1390 | |
| 1391 // push all ArgEscape nodes on the worklist | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1392 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1393 int nk = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1394 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape) |
| 0 | 1395 worklist.push(nk); |
| 1396 } | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1397 // mark all nodes reachable from ArgEscape nodes |
| 0 | 1398 while(worklist.length() > 0) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1399 PointsToNode* ptn = ptnode_adr(worklist.pop()); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1400 if (ptn->node_type() == PointsToNode::JavaObject) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1401 has_non_escaping_obj = true; // Non GlobalEscape |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1402 uint e_cnt = ptn->edge_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1403 for (uint ei = 0; ei < e_cnt; ei++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1404 uint npi = ptn->edge_target(ei); |
| 0 | 1405 PointsToNode *np = ptnode_adr(npi); |
| 65 | 1406 if (np->escape_state() < PointsToNode::ArgEscape) { |
| 0 | 1407 np->set_escape_state(PointsToNode::ArgEscape); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1408 worklist.push(npi); |
| 0 | 1409 } |
| 1410 } | |
| 1411 } | |
| 65 | 1412 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1413 GrowableArray<Node*> alloc_worklist; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1414 |
| 65 | 1415 // push all NoEscape nodes on the worklist |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1416 for( uint next = 0; next < cg_length; ++next ) { |
| 65 | 1417 int nk = cg_worklist.at(next); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1418 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape) |
| 65 | 1419 worklist.push(nk); |
| 1420 } | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1421 // mark all nodes reachable from NoEscape nodes |
| 65 | 1422 while(worklist.length() > 0) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1423 PointsToNode* ptn = ptnode_adr(worklist.pop()); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1424 if (ptn->node_type() == PointsToNode::JavaObject) |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1425 has_non_escaping_obj = true; // Non GlobalEscape |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1426 Node* n = ptn->_node; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1427 if (n->is_Allocate() && ptn->_scalar_replaceable ) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1428 // Push scalar replaceable alocations on alloc_worklist |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1429 // for processing in split_unique_types(). |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1430 alloc_worklist.append(n); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1431 } |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1432 uint e_cnt = ptn->edge_count(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1433 for (uint ei = 0; ei < e_cnt; ei++) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1434 uint npi = ptn->edge_target(ei); |
| 65 | 1435 PointsToNode *np = ptnode_adr(npi); |
| 1436 if (np->escape_state() < PointsToNode::NoEscape) { | |
| 1437 np->set_escape_state(PointsToNode::NoEscape); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1438 worklist.push(npi); |
| 65 | 1439 } |
| 1440 } | |
| 1441 } | |
| 1442 | |
| 0 | 1443 _collecting = false; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1444 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build"); |
| 0 | 1445 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1446 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1447 if ( has_scalar_replaceable_candidates && |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1448 C->AliasLevel() >= 3 && EliminateAllocations ) { |
| 0 | 1449 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1450 // Now use the escape information to create unique types for |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1451 // scalar replaceable objects. |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1452 split_unique_types(alloc_worklist); |
| 0 | 1453 |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1454 if (C->failing()) return false; |
| 0 | 1455 |
| 65 | 1456 // Clean up after split unique types. |
| 1457 ResourceMark rm; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1458 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn()); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1459 |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1460 C->print_method("After Escape Analysis", 2); |
| 0 | 1461 |
| 65 | 1462 #ifdef ASSERT |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1463 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { |
| 65 | 1464 tty->print("=== No allocations eliminated for "); |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1465 C->method()->print_short_name(); |
| 65 | 1466 if(!EliminateAllocations) { |
| 1467 tty->print(" since EliminateAllocations is off ==="); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1468 } else if(!has_scalar_replaceable_candidates) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1469 tty->print(" since there are no scalar replaceable candidates ==="); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1470 } else if(C->AliasLevel() < 3) { |
| 65 | 1471 tty->print(" since AliasLevel < 3 ==="); |
| 0 | 1472 } |
| 65 | 1473 tty->cr(); |
| 1474 #endif | |
| 0 | 1475 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1476 return has_non_escaping_obj; |
| 0 | 1477 } |
| 1478 | |
| 1479 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) { | |
| 1480 | |
| 1481 switch (call->Opcode()) { | |
| 65 | 1482 #ifdef ASSERT |
| 0 | 1483 case Op_Allocate: |
| 1484 case Op_AllocateArray: | |
| 1485 case Op_Lock: | |
| 1486 case Op_Unlock: | |
| 65 | 1487 assert(false, "should be done already"); |
| 0 | 1488 break; |
| 65 | 1489 #endif |
| 1490 case Op_CallLeafNoFP: | |
| 1491 { | |
| 1492 // Stub calls, objects do not escape but they are not scale replaceable. | |
| 1493 // Adjust escape state for outgoing arguments. | |
| 1494 const TypeTuple * d = call->tf()->domain(); | |
| 1495 VectorSet ptset(Thread::current()->resource_area()); | |
| 1496 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { | |
| 1497 const Type* at = d->field_at(i); | |
| 1498 Node *arg = call->in(i)->uncast(); | |
| 1499 const Type *aat = phase->type(arg); | |
| 1500 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr()) { | |
| 1501 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || | |
| 1502 aat->isa_ptr() != NULL, "expecting an Ptr"); | |
| 1503 set_escape_state(arg->_idx, PointsToNode::ArgEscape); | |
| 1504 if (arg->is_AddP()) { | |
| 1505 // | |
| 1506 // The inline_native_clone() case when the arraycopy stub is called | |
| 1507 // after the allocation before Initialize and CheckCastPP nodes. | |
| 1508 // | |
| 1509 // Set AddP's base (Allocate) as not scalar replaceable since | |
| 1510 // pointer to the base (with offset) is passed as argument. | |
| 1511 // | |
| 1512 arg = get_addp_base(arg); | |
| 1513 } | |
| 1514 ptset.Clear(); | |
| 1515 PointsTo(ptset, arg, phase); | |
| 1516 for( VectorSetI j(&ptset); j.test(); ++j ) { | |
| 1517 uint pt = j.elem; | |
| 1518 set_escape_state(pt, PointsToNode::ArgEscape); | |
| 1519 } | |
| 1520 } | |
| 1521 } | |
| 1522 break; | |
| 1523 } | |
| 0 | 1524 |
| 1525 case Op_CallStaticJava: | |
| 1526 // For a static call, we know exactly what method is being called. | |
| 1527 // Use bytecode estimator to record the call's escape affects | |
| 1528 { | |
| 1529 ciMethod *meth = call->as_CallJava()->method(); | |
| 65 | 1530 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; |
| 1531 // fall-through if not a Java method or no analyzer information | |
| 1532 if (call_analyzer != NULL) { | |
| 0 | 1533 const TypeTuple * d = call->tf()->domain(); |
| 1534 VectorSet ptset(Thread::current()->resource_area()); | |
| 65 | 1535 bool copy_dependencies = false; |
| 0 | 1536 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { |
| 1537 const Type* at = d->field_at(i); | |
| 1538 int k = i - TypeFunc::Parms; | |
| 1539 | |
| 1540 if (at->isa_oopptr() != NULL) { | |
| 65 | 1541 Node *arg = call->in(i)->uncast(); |
| 0 | 1542 |
| 65 | 1543 bool global_escapes = false; |
| 1544 bool fields_escapes = false; | |
| 1545 if (!call_analyzer->is_arg_stack(k)) { | |
| 1546 // The argument global escapes, mark everything it could point to | |
| 1547 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); | |
| 1548 global_escapes = true; | |
| 1549 } else { | |
| 1550 if (!call_analyzer->is_arg_local(k)) { | |
| 1551 // The argument itself doesn't escape, but any fields might | |
| 1552 fields_escapes = true; | |
| 0 | 1553 } |
| 65 | 1554 set_escape_state(arg->_idx, PointsToNode::ArgEscape); |
| 1555 copy_dependencies = true; | |
| 1556 } | |
| 1557 | |
| 1558 ptset.Clear(); | |
| 1559 PointsTo(ptset, arg, phase); | |
| 1560 for( VectorSetI j(&ptset); j.test(); ++j ) { | |
| 1561 uint pt = j.elem; | |
| 1562 if (global_escapes) { | |
| 1563 //The argument global escapes, mark everything it could point to | |
| 1564 set_escape_state(pt, PointsToNode::GlobalEscape); | |
| 1565 } else { | |
| 1566 if (fields_escapes) { | |
| 1567 // The argument itself doesn't escape, but any fields might | |
| 1568 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot); | |
| 1569 } | |
| 1570 set_escape_state(pt, PointsToNode::ArgEscape); | |
| 0 | 1571 } |
| 1572 } | |
| 1573 } | |
| 1574 } | |
| 65 | 1575 if (copy_dependencies) |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1576 call_analyzer->copy_dependencies(_compile->dependencies()); |
| 0 | 1577 break; |
| 1578 } | |
| 1579 } | |
| 1580 | |
| 1581 default: | |
| 65 | 1582 // Fall-through here if not a Java method or no analyzer information |
| 1583 // or some other type of call, assume the worst case: all arguments | |
| 0 | 1584 // globally escape. |
| 1585 { | |
| 1586 // adjust escape state for outgoing arguments | |
| 1587 const TypeTuple * d = call->tf()->domain(); | |
| 1588 VectorSet ptset(Thread::current()->resource_area()); | |
| 1589 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { | |
| 1590 const Type* at = d->field_at(i); | |
| 1591 if (at->isa_oopptr() != NULL) { | |
| 65 | 1592 Node *arg = call->in(i)->uncast(); |
| 1593 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); | |
| 0 | 1594 ptset.Clear(); |
| 1595 PointsTo(ptset, arg, phase); | |
| 1596 for( VectorSetI j(&ptset); j.test(); ++j ) { | |
| 1597 uint pt = j.elem; | |
| 1598 set_escape_state(pt, PointsToNode::GlobalEscape); | |
| 1599 } | |
| 1600 } | |
| 1601 } | |
| 1602 } | |
| 1603 } | |
| 1604 } | |
| 1605 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1606 CallNode *call = resproj->in(0)->as_Call(); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1607 uint call_idx = call->_idx; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1608 uint resproj_idx = resproj->_idx; |
| 0 | 1609 |
| 1610 switch (call->Opcode()) { | |
| 1611 case Op_Allocate: | |
| 1612 { | |
| 1613 Node *k = call->in(AllocateNode::KlassNode); | |
| 1614 const TypeKlassPtr *kt; | |
| 1615 if (k->Opcode() == Op_LoadKlass) { | |
| 1616 kt = k->as_Load()->type()->isa_klassptr(); | |
| 1617 } else { | |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1618 // Also works for DecodeN(LoadNKlass). |
| 0 | 1619 kt = k->as_Type()->type()->isa_klassptr(); |
| 1620 } | |
| 1621 assert(kt != NULL, "TypeKlassPtr required."); | |
| 1622 ciKlass* cik = kt->klass(); | |
| 1623 ciInstanceKlass* ciik = cik->as_instance_klass(); | |
| 1624 | |
| 65 | 1625 PointsToNode::EscapeState es; |
| 1626 uint edge_to; | |
| 0 | 1627 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || ciik->has_finalizer()) { |
| 65 | 1628 es = PointsToNode::GlobalEscape; |
| 1629 edge_to = _phantom_object; // Could not be worse | |
| 0 | 1630 } else { |
| 65 | 1631 es = PointsToNode::NoEscape; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1632 edge_to = call_idx; |
| 0 | 1633 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1634 set_escape_state(call_idx, es); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1635 add_pointsto_edge(resproj_idx, edge_to); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1636 _processed.set(resproj_idx); |
| 0 | 1637 break; |
| 1638 } | |
| 1639 | |
| 1640 case Op_AllocateArray: | |
| 1641 { | |
| 65 | 1642 int length = call->in(AllocateNode::ALength)->find_int_con(-1); |
| 1643 if (length < 0 || length > EliminateAllocationArraySizeLimit) { | |
| 1644 // 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
|
1645 ptnode_adr(call_idx)->_scalar_replaceable = false; |
| 65 | 1646 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1647 set_escape_state(call_idx, PointsToNode::NoEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1648 add_pointsto_edge(resproj_idx, call_idx); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1649 _processed.set(resproj_idx); |
| 0 | 1650 break; |
| 1651 } | |
| 1652 | |
| 1653 case Op_CallStaticJava: | |
| 1654 // For a static call, we know exactly what method is being called. | |
| 1655 // Use bytecode estimator to record whether the call's return value escapes | |
| 1656 { | |
| 65 | 1657 bool done = true; |
| 0 | 1658 const TypeTuple *r = call->tf()->range(); |
| 1659 const Type* ret_type = NULL; | |
| 1660 | |
| 1661 if (r->cnt() > TypeFunc::Parms) | |
| 1662 ret_type = r->field_at(TypeFunc::Parms); | |
| 1663 | |
| 1664 // Note: we use isa_ptr() instead of isa_oopptr() here because the | |
| 1665 // _multianewarray functions return a TypeRawPtr. | |
| 65 | 1666 if (ret_type == NULL || ret_type->isa_ptr() == NULL) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1667 _processed.set(resproj_idx); |
| 0 | 1668 break; // doesn't return a pointer type |
| 65 | 1669 } |
| 0 | 1670 ciMethod *meth = call->as_CallJava()->method(); |
| 65 | 1671 const TypeTuple * d = call->tf()->domain(); |
| 0 | 1672 if (meth == NULL) { |
| 1673 // not a Java method, assume global escape | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1674 set_escape_state(call_idx, PointsToNode::GlobalEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1675 add_pointsto_edge(resproj_idx, _phantom_object); |
| 0 | 1676 } else { |
| 65 | 1677 BCEscapeAnalyzer *call_analyzer = meth->get_bcea(); |
| 1678 bool copy_dependencies = false; | |
| 0 | 1679 |
| 65 | 1680 if (call_analyzer->is_return_allocated()) { |
| 1681 // Returns a newly allocated unescaped object, simply | |
| 1682 // update dependency information. | |
| 1683 // Mark it as NoEscape so that objects referenced by | |
| 1684 // it's fields will be marked as NoEscape at least. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1685 set_escape_state(call_idx, PointsToNode::NoEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1686 add_pointsto_edge(resproj_idx, call_idx); |
| 65 | 1687 copy_dependencies = true; |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1688 } else if (call_analyzer->is_return_local()) { |
| 0 | 1689 // determine whether any arguments are returned |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1690 set_escape_state(call_idx, PointsToNode::NoEscape); |
| 0 | 1691 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { |
| 1692 const Type* at = d->field_at(i); | |
| 1693 | |
| 1694 if (at->isa_oopptr() != NULL) { | |
| 65 | 1695 Node *arg = call->in(i)->uncast(); |
| 0 | 1696 |
| 65 | 1697 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1698 PointsToNode *arg_esp = ptnode_adr(arg->_idx); |
| 65 | 1699 if (arg_esp->node_type() == PointsToNode::UnknownType) |
| 1700 done = false; | |
| 1701 else if (arg_esp->node_type() == PointsToNode::JavaObject) | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1702 add_pointsto_edge(resproj_idx, arg->_idx); |
| 0 | 1703 else |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1704 add_deferred_edge(resproj_idx, arg->_idx); |
| 0 | 1705 arg_esp->_hidden_alias = true; |
| 1706 } | |
| 1707 } | |
| 1708 } | |
| 65 | 1709 copy_dependencies = true; |
| 0 | 1710 } else { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1711 set_escape_state(call_idx, PointsToNode::GlobalEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1712 add_pointsto_edge(resproj_idx, _phantom_object); |
| 65 | 1713 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { |
| 1714 const Type* at = d->field_at(i); | |
| 1715 if (at->isa_oopptr() != NULL) { | |
| 1716 Node *arg = call->in(i)->uncast(); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1717 PointsToNode *arg_esp = ptnode_adr(arg->_idx); |
| 65 | 1718 arg_esp->_hidden_alias = true; |
| 1719 } | |
| 1720 } | |
| 0 | 1721 } |
| 65 | 1722 if (copy_dependencies) |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1723 call_analyzer->copy_dependencies(_compile->dependencies()); |
| 0 | 1724 } |
| 65 | 1725 if (done) |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1726 _processed.set(resproj_idx); |
| 0 | 1727 break; |
| 1728 } | |
| 1729 | |
| 1730 default: | |
| 1731 // Some other type of call, assume the worst case that the | |
| 1732 // returned value, if any, globally escapes. | |
| 1733 { | |
| 1734 const TypeTuple *r = call->tf()->range(); | |
| 1735 if (r->cnt() > TypeFunc::Parms) { | |
| 1736 const Type* ret_type = r->field_at(TypeFunc::Parms); | |
| 1737 | |
| 1738 // Note: we use isa_ptr() instead of isa_oopptr() here because the | |
| 1739 // _multianewarray functions return a TypeRawPtr. | |
| 1740 if (ret_type->isa_ptr() != NULL) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1741 set_escape_state(call_idx, PointsToNode::GlobalEscape); |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1742 add_pointsto_edge(resproj_idx, _phantom_object); |
| 0 | 1743 } |
| 1744 } | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1745 _processed.set(resproj_idx); |
| 0 | 1746 } |
| 1747 } | |
| 1748 } | |
| 1749 | |
| 65 | 1750 // Populate Connection Graph with Ideal nodes and create simple |
| 1751 // connection graph edges (do not need to check the node_type of inputs | |
| 1752 // or to call PointsTo() to walk the connection graph). | |
| 1753 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) { | |
| 1754 if (_processed.test(n->_idx)) | |
| 1755 return; // No need to redefine node's state. | |
| 1756 | |
| 1757 if (n->is_Call()) { | |
| 1758 // Arguments to allocation and locking don't escape. | |
| 1759 if (n->is_Allocate()) { | |
| 1760 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true); | |
| 1761 record_for_optimizer(n); | |
| 1762 } else if (n->is_Lock() || n->is_Unlock()) { | |
| 1763 // Put Lock and Unlock nodes on IGVN worklist to process them during | |
| 1764 // the first IGVN optimization when escape information is still available. | |
| 1765 record_for_optimizer(n); | |
| 1766 _processed.set(n->_idx); | |
| 1767 } else { | |
| 1768 // Have to process call's arguments first. | |
| 1769 PointsToNode::NodeType nt = PointsToNode::UnknownType; | |
| 1770 | |
| 1771 // Check if a call returns an object. | |
| 1772 const TypeTuple *r = n->as_Call()->tf()->range(); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1773 if (n->is_CallStaticJava() && r->cnt() > TypeFunc::Parms && |
| 65 | 1774 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) { |
| 1775 // Note: use isa_ptr() instead of isa_oopptr() here because | |
| 1776 // the _multianewarray functions return a TypeRawPtr. | |
| 1777 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { | |
| 1778 nt = PointsToNode::JavaObject; | |
| 1779 } | |
| 1780 } | |
| 1781 add_node(n, nt, PointsToNode::UnknownEscape, false); | |
| 1782 } | |
| 1783 return; | |
| 1784 } | |
| 1785 | |
| 1786 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because | |
| 1787 // ThreadLocal has RawPrt type. | |
| 1788 switch (n->Opcode()) { | |
| 1789 case Op_AddP: | |
| 1790 { | |
| 1791 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false); | |
| 1792 break; | |
| 1793 } | |
| 1794 case Op_CastX2P: | |
| 1795 { // "Unsafe" memory access. | |
| 1796 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); | |
| 1797 break; | |
| 1798 } | |
| 1799 case Op_CastPP: | |
| 1800 case Op_CheckCastPP: | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1801 case Op_EncodeP: |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1802 case Op_DecodeN: |
| 65 | 1803 { |
| 1804 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); | |
| 1805 int ti = n->in(1)->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1806 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); |
| 65 | 1807 if (nt == PointsToNode::UnknownType) { |
| 1808 _delayed_worklist.push(n); // Process it later. | |
| 1809 break; | |
| 1810 } else if (nt == PointsToNode::JavaObject) { | |
| 1811 add_pointsto_edge(n->_idx, ti); | |
| 1812 } else { | |
| 1813 add_deferred_edge(n->_idx, ti); | |
| 1814 } | |
| 1815 _processed.set(n->_idx); | |
| 1816 break; | |
| 1817 } | |
| 1818 case Op_ConP: | |
| 1819 { | |
| 1820 // assume all pointer constants globally escape except for null | |
| 1821 PointsToNode::EscapeState es; | |
| 1822 if (phase->type(n) == TypePtr::NULL_PTR) | |
| 1823 es = PointsToNode::NoEscape; | |
| 1824 else | |
| 1825 es = PointsToNode::GlobalEscape; | |
| 0 | 1826 |
| 65 | 1827 add_node(n, PointsToNode::JavaObject, es, true); |
| 1828 break; | |
| 1829 } | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1830 case Op_ConN: |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1831 { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1832 // 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
|
1833 PointsToNode::EscapeState es; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1834 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
|
1835 es = PointsToNode::NoEscape; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1836 else |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1837 es = PointsToNode::GlobalEscape; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1838 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1839 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
|
1840 break; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1841 } |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1842 case Op_CreateEx: |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1843 { |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1844 // assume that all exception objects globally escape |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1845 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1846 break; |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1847 } |
| 65 | 1848 case Op_LoadKlass: |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1849 case Op_LoadNKlass: |
| 65 | 1850 { |
| 1851 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); | |
| 1852 break; | |
| 1853 } | |
| 1854 case Op_LoadP: | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1855 case Op_LoadN: |
| 65 | 1856 { |
| 1857 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
|
1858 if (!t->isa_narrowoop() && t->isa_ptr() == NULL) { |
| 65 | 1859 _processed.set(n->_idx); |
| 1860 return; | |
| 1861 } | |
| 1862 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); | |
| 1863 break; | |
| 1864 } | |
| 1865 case Op_Parm: | |
| 1866 { | |
| 1867 _processed.set(n->_idx); // No need to redefine it state. | |
| 1868 uint con = n->as_Proj()->_con; | |
| 1869 if (con < TypeFunc::Parms) | |
| 1870 return; | |
| 1871 const Type *t = n->in(0)->as_Start()->_domain->field_at(con); | |
| 1872 if (t->isa_ptr() == NULL) | |
| 1873 return; | |
| 1874 // We have to assume all input parameters globally escape | |
| 1875 // (Note: passing 'false' since _processed is already set). | |
| 1876 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false); | |
| 1877 break; | |
| 1878 } | |
| 1879 case Op_Phi: | |
| 1880 { | |
| 1881 if (n->as_Phi()->type()->isa_ptr() == NULL) { | |
| 1882 // nothing to do if not an oop | |
| 1883 _processed.set(n->_idx); | |
| 1884 return; | |
| 1885 } | |
| 1886 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); | |
| 1887 uint i; | |
| 1888 for (i = 1; i < n->req() ; i++) { | |
| 1889 Node* in = n->in(i); | |
| 1890 if (in == NULL) | |
| 1891 continue; // ignore NULL | |
| 1892 in = in->uncast(); | |
| 1893 if (in->is_top() || in == n) | |
| 1894 continue; // ignore top or inputs which go back this node | |
| 1895 int ti = in->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1896 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); |
| 65 | 1897 if (nt == PointsToNode::UnknownType) { |
| 1898 break; | |
| 1899 } else if (nt == PointsToNode::JavaObject) { | |
| 1900 add_pointsto_edge(n->_idx, ti); | |
| 1901 } else { | |
| 1902 add_deferred_edge(n->_idx, ti); | |
| 1903 } | |
| 1904 } | |
| 1905 if (i >= n->req()) | |
| 1906 _processed.set(n->_idx); | |
| 1907 else | |
| 1908 _delayed_worklist.push(n); | |
| 1909 break; | |
| 1910 } | |
| 1911 case Op_Proj: | |
| 1912 { | |
| 1913 // we are only interested in the result projection from a call | |
| 1914 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { | |
| 1915 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); | |
| 1916 process_call_result(n->as_Proj(), phase); | |
| 1917 if (!_processed.test(n->_idx)) { | |
| 1918 // The call's result may need to be processed later if the call | |
| 1919 // returns it's argument and the argument is not processed yet. | |
| 1920 _delayed_worklist.push(n); | |
| 1921 } | |
| 1922 } else { | |
| 1923 _processed.set(n->_idx); | |
| 1924 } | |
| 1925 break; | |
| 1926 } | |
| 1927 case Op_Return: | |
| 1928 { | |
| 1929 if( n->req() > TypeFunc::Parms && | |
| 1930 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { | |
| 1931 // Treat Return value as LocalVar with GlobalEscape escape state. | |
| 1932 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false); | |
| 1933 int ti = n->in(TypeFunc::Parms)->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1934 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); |
| 65 | 1935 if (nt == PointsToNode::UnknownType) { |
| 1936 _delayed_worklist.push(n); // Process it later. | |
| 1937 break; | |
| 1938 } else if (nt == PointsToNode::JavaObject) { | |
| 1939 add_pointsto_edge(n->_idx, ti); | |
| 1940 } else { | |
| 1941 add_deferred_edge(n->_idx, ti); | |
| 1942 } | |
| 1943 } | |
| 1944 _processed.set(n->_idx); | |
| 1945 break; | |
| 1946 } | |
| 1947 case Op_StoreP: | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1948 case Op_StoreN: |
| 65 | 1949 { |
| 1950 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
|
1951 adr_type = adr_type->make_ptr(); |
| 65 | 1952 if (adr_type->isa_oopptr()) { |
| 1953 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); | |
| 1954 } else { | |
| 1955 Node* adr = n->in(MemNode::Address); | |
| 1956 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL && | |
| 1957 adr->in(AddPNode::Address)->is_Proj() && | |
| 1958 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { | |
| 1959 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); | |
| 1960 // We are computing a raw address for a store captured | |
| 1961 // by an Initialize compute an appropriate address type. | |
| 1962 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); | |
| 1963 assert(offs != Type::OffsetBot, "offset must be a constant"); | |
| 1964 } else { | |
| 1965 _processed.set(n->_idx); | |
| 1966 return; | |
| 1967 } | |
| 1968 } | |
| 1969 break; | |
| 1970 } | |
| 1971 case Op_StorePConditional: | |
| 1972 case Op_CompareAndSwapP: | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
1973 case Op_CompareAndSwapN: |
| 65 | 1974 { |
| 1975 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
|
1976 adr_type = adr_type->make_ptr(); |
| 65 | 1977 if (adr_type->isa_oopptr()) { |
| 1978 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); | |
| 1979 } else { | |
| 1980 _processed.set(n->_idx); | |
| 1981 return; | |
| 1982 } | |
| 1983 break; | |
| 1984 } | |
| 1985 case Op_ThreadLocal: | |
| 1986 { | |
| 1987 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true); | |
| 1988 break; | |
| 1989 } | |
| 1990 default: | |
| 1991 ; | |
| 1992 // nothing to do | |
| 1993 } | |
| 1994 return; | |
| 1995 } | |
| 1996 | |
| 1997 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1998 uint n_idx = n->_idx; |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
1999 |
| 65 | 2000 // Don't set processed bit for AddP, LoadP, StoreP since |
| 2001 // they may need more then one pass to process. | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2002 if (_processed.test(n_idx)) |
| 65 | 2003 return; // No need to redefine node's state. |
| 2004 | |
| 0 | 2005 if (n->is_Call()) { |
| 2006 CallNode *call = n->as_Call(); | |
| 2007 process_call_arguments(call, phase); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2008 _processed.set(n_idx); |
| 0 | 2009 return; |
| 2010 } | |
| 2011 | |
| 65 | 2012 switch (n->Opcode()) { |
| 0 | 2013 case Op_AddP: |
| 2014 { | |
| 65 | 2015 Node *base = get_addp_base(n); |
| 2016 // Create a field edge to this node from everything base could point to. | |
| 0 | 2017 VectorSet ptset(Thread::current()->resource_area()); |
| 2018 PointsTo(ptset, base, phase); | |
| 2019 for( VectorSetI i(&ptset); i.test(); ++i ) { | |
| 2020 uint pt = i.elem; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2021 add_field_edge(pt, n_idx, address_offset(n, phase)); |
| 65 | 2022 } |
| 2023 break; | |
| 2024 } | |
| 2025 case Op_CastX2P: | |
| 2026 { | |
| 2027 assert(false, "Op_CastX2P"); | |
| 2028 break; | |
| 2029 } | |
| 2030 case Op_CastPP: | |
| 2031 case Op_CheckCastPP: | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
2032 case Op_EncodeP: |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
102
diff
changeset
|
2033 case Op_DecodeN: |
| 65 | 2034 { |
| 2035 int ti = n->in(1)->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2036 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2037 add_pointsto_edge(n_idx, ti); |
| 65 | 2038 } else { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2039 add_deferred_edge(n_idx, ti); |
| 65 | 2040 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2041 _processed.set(n_idx); |
| 65 | 2042 break; |
| 2043 } | |
| 2044 case Op_ConP: | |
| 2045 { | |
| 2046 assert(false, "Op_ConP"); | |
| 2047 break; | |
| 2048 } | |
| 163 | 2049 case Op_ConN: |
| 2050 { | |
| 2051 assert(false, "Op_ConN"); | |
| 2052 break; | |
| 2053 } | |
| 65 | 2054 case Op_CreateEx: |
| 2055 { | |
| 2056 assert(false, "Op_CreateEx"); | |
| 2057 break; | |
| 2058 } | |
| 2059 case Op_LoadKlass: | |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2060 case Op_LoadNKlass: |
| 65 | 2061 { |
| 2062 assert(false, "Op_LoadKlass"); | |
| 2063 break; | |
| 2064 } | |
| 2065 case Op_LoadP: | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
2066 case Op_LoadN: |
| 65 | 2067 { |
| 2068 const Type *t = phase->type(n); | |
| 2069 #ifdef ASSERT | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
2070 if (!t->isa_narrowoop() && t->isa_ptr() == NULL) |
| 65 | 2071 assert(false, "Op_LoadP"); |
| 2072 #endif | |
| 2073 | |
| 2074 Node* adr = n->in(MemNode::Address)->uncast(); | |
| 2075 const Type *adr_type = phase->type(adr); | |
| 2076 Node* adr_base; | |
| 2077 if (adr->is_AddP()) { | |
| 2078 adr_base = get_addp_base(adr); | |
| 2079 } else { | |
| 2080 adr_base = adr; | |
| 2081 } | |
| 2082 | |
| 2083 // For everything "adr_base" could point to, create a deferred edge from | |
| 2084 // this node to each field with the same offset. | |
| 2085 VectorSet ptset(Thread::current()->resource_area()); | |
| 2086 PointsTo(ptset, adr_base, phase); | |
| 2087 int offset = address_offset(adr, phase); | |
| 2088 for( VectorSetI i(&ptset); i.test(); ++i ) { | |
| 2089 uint pt = i.elem; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2090 add_deferred_edge_to_fields(n_idx, pt, offset); |
| 0 | 2091 } |
| 2092 break; | |
| 2093 } | |
| 2094 case Op_Parm: | |
| 2095 { | |
| 65 | 2096 assert(false, "Op_Parm"); |
| 0 | 2097 break; |
| 2098 } | |
| 2099 case Op_Phi: | |
| 2100 { | |
| 65 | 2101 #ifdef ASSERT |
| 2102 if (n->as_Phi()->type()->isa_ptr() == NULL) | |
| 2103 assert(false, "Op_Phi"); | |
| 2104 #endif | |
| 2105 for (uint i = 1; i < n->req() ; i++) { | |
| 2106 Node* in = n->in(i); | |
| 2107 if (in == NULL) | |
| 2108 continue; // ignore NULL | |
| 2109 in = in->uncast(); | |
| 2110 if (in->is_top() || in == n) | |
| 2111 continue; // ignore top or inputs which go back this node | |
| 2112 int ti = in->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2113 if (ptnode_adr(in->_idx)->node_type() == PointsToNode::JavaObject) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2114 add_pointsto_edge(n_idx, ti); |
| 65 | 2115 } else { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2116 add_deferred_edge(n_idx, ti); |
| 65 | 2117 } |
| 2118 } | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2119 _processed.set(n_idx); |
| 0 | 2120 break; |
| 2121 } | |
| 65 | 2122 case Op_Proj: |
| 0 | 2123 { |
| 65 | 2124 // we are only interested in the result projection from a call |
| 2125 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { | |
| 2126 process_call_result(n->as_Proj(), phase); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2127 assert(_processed.test(n_idx), "all call results should be processed"); |
| 65 | 2128 } else { |
| 2129 assert(false, "Op_Proj"); | |
| 2130 } | |
| 0 | 2131 break; |
| 2132 } | |
| 65 | 2133 case Op_Return: |
| 0 | 2134 { |
| 65 | 2135 #ifdef ASSERT |
| 2136 if( n->req() <= TypeFunc::Parms || | |
| 2137 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { | |
| 2138 assert(false, "Op_Return"); | |
| 0 | 2139 } |
| 65 | 2140 #endif |
| 2141 int ti = n->in(TypeFunc::Parms)->_idx; | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2142 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { |
|
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2143 add_pointsto_edge(n_idx, ti); |
| 65 | 2144 } else { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2145 add_deferred_edge(n_idx, ti); |
| 65 | 2146 } |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2147 _processed.set(n_idx); |
| 0 | 2148 break; |
| 2149 } | |
| 2150 case Op_StoreP: | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
2151 case Op_StoreN: |
| 0 | 2152 case Op_StorePConditional: |
| 2153 case Op_CompareAndSwapP: | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
2154 case Op_CompareAndSwapN: |
| 0 | 2155 { |
| 2156 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
|
2157 const Type *adr_type = phase->type(adr)->make_ptr(); |
| 65 | 2158 #ifdef ASSERT |
| 0 | 2159 if (!adr_type->isa_oopptr()) |
| 65 | 2160 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP"); |
| 2161 #endif | |
| 0 | 2162 |
| 65 | 2163 assert(adr->is_AddP(), "expecting an AddP"); |
| 2164 Node *adr_base = get_addp_base(adr); | |
| 2165 Node *val = n->in(MemNode::ValueIn)->uncast(); | |
| 2166 // For everything "adr_base" could point to, create a deferred edge | |
| 2167 // to "val" from each field with the same offset. | |
| 0 | 2168 VectorSet ptset(Thread::current()->resource_area()); |
| 2169 PointsTo(ptset, adr_base, phase); | |
| 2170 for( VectorSetI i(&ptset); i.test(); ++i ) { | |
| 2171 uint pt = i.elem; | |
| 65 | 2172 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase)); |
| 0 | 2173 } |
| 2174 break; | |
| 2175 } | |
| 65 | 2176 case Op_ThreadLocal: |
| 0 | 2177 { |
| 65 | 2178 assert(false, "Op_ThreadLocal"); |
| 0 | 2179 break; |
| 2180 } | |
| 2181 default: | |
| 2182 ; | |
| 2183 // nothing to do | |
| 2184 } | |
| 2185 } | |
| 2186 | |
| 2187 #ifndef PRODUCT | |
| 2188 void ConnectionGraph::dump() { | |
| 2189 PhaseGVN *igvn = _compile->initial_gvn(); | |
| 2190 bool first = true; | |
| 2191 | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2192 uint size = nodes_size(); |
| 65 | 2193 for (uint ni = 0; ni < size; ni++) { |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2194 PointsToNode *ptn = ptnode_adr(ni); |
| 65 | 2195 PointsToNode::NodeType ptn_type = ptn->node_type(); |
| 2196 | |
| 2197 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL) | |
| 0 | 2198 continue; |
| 65 | 2199 PointsToNode::EscapeState es = escape_state(ptn->_node, igvn); |
| 2200 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) { | |
| 2201 if (first) { | |
| 2202 tty->cr(); | |
| 2203 tty->print("======== Connection graph for "); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2204 _compile->method()->print_short_name(); |
| 65 | 2205 tty->cr(); |
| 2206 first = false; | |
| 2207 } | |
| 2208 tty->print("%6d ", ni); | |
| 2209 ptn->dump(); | |
| 2210 // Print all locals which reference this allocation | |
| 2211 for (uint li = ni; li < size; li++) { | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2212 PointsToNode *ptn_loc = ptnode_adr(li); |
| 65 | 2213 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type(); |
| 2214 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL && | |
| 2215 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) { | |
| 2216 tty->print("%6d LocalVar [[%d]]", li, ni); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2217 ptnode_adr(li)->_node->dump(); |
| 0 | 2218 } |
| 2219 } | |
| 65 | 2220 if (Verbose) { |
| 2221 // Print all fields which reference this allocation | |
| 2222 for (uint i = 0; i < ptn->edge_count(); i++) { | |
| 2223 uint ei = ptn->edge_target(i); | |
| 2224 tty->print("%6d Field [[%d]]", ei, ni); | |
|
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
2225 ptnode_adr(ei)->_node->dump(); |
| 65 | 2226 } |
| 2227 } | |
| 2228 tty->cr(); | |
| 0 | 2229 } |
| 2230 } | |
| 2231 } | |
| 2232 #endif |