Mercurial > hg > graal-compiler
annotate src/share/vm/opto/loopopts.cpp @ 6216:161ae369407d
7181632: nsk classLoad001_14 failure and CompileTheWorld crash after 7178145.
Summary: Need to copy the inlined exception table to the new constMethodOop during method rewriting.
Reviewed-by: coleenp, dholmes
author | jiangli |
---|---|
date | Thu, 05 Jul 2012 20:54:45 -0400 |
parents | 5e990493719e |
children | e626685e9f6c |
rev | line source |
---|---|
0 | 1 /* |
2426
1d1603768966
7010070: Update all 2010 Oracle-changed OpenJDK files to have the proper copyright dates - second pass
trims
parents:
2383
diff
changeset
|
2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1254
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1254
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1254
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 25 #include "precompiled.hpp" |
26 #include "memory/allocation.inline.hpp" | |
27 #include "opto/addnode.hpp" | |
28 #include "opto/connode.hpp" | |
29 #include "opto/divnode.hpp" | |
30 #include "opto/loopnode.hpp" | |
4047 | 31 #include "opto/matcher.hpp" |
1972 | 32 #include "opto/mulnode.hpp" |
33 #include "opto/rootnode.hpp" | |
34 #include "opto/subnode.hpp" | |
0 | 35 |
36 //============================================================================= | |
37 //------------------------------split_thru_phi--------------------------------- | |
38 // Split Node 'n' through merge point if there is enough win. | |
39 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { | |
69 | 40 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { |
41 // ConvI2L may have type information on it which is unsafe to push up | |
42 // so disable this for now | |
43 return NULL; | |
44 } | |
0 | 45 int wins = 0; |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
46 assert(!n->is_CFG(), ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
47 assert(region->is_Region(), ""); |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
48 |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
49 const Type* type = n->bottom_type(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
50 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
51 Node *phi; |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
52 if (t_oop != NULL && t_oop->is_known_instance_field()) { |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
53 int iid = t_oop->instance_id(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
54 int index = C->get_alias_index(t_oop); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
55 int offset = t_oop->offset(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
56 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
57 } else { |
1254
4ee1c645110e
6924097: assert((_type == Type::MEMORY) == (_adr_type != 0),"adr_type for memory phis only")
kvn
parents:
866
diff
changeset
|
58 phi = PhiNode::make_blank(region, n); |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
35
diff
changeset
|
59 } |
0 | 60 uint old_unique = C->unique(); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
61 for (uint i = 1; i < region->req(); i++) { |
0 | 62 Node *x; |
63 Node* the_clone = NULL; | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
64 if (region->in(i) == C->top()) { |
0 | 65 x = C->top(); // Dead path? Use a dead data op |
66 } else { | |
67 x = n->clone(); // Else clone up the data op | |
68 the_clone = x; // Remember for possible deletion. | |
69 // Alter data node to use pre-phi inputs | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
70 if (n->in(0) == region) |
0 | 71 x->set_req( 0, region->in(i) ); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
72 for (uint j = 1; j < n->req(); j++) { |
0 | 73 Node *in = n->in(j); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
74 if (in->is_Phi() && in->in(0) == region) |
0 | 75 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone |
76 } | |
77 } | |
78 // Check for a 'win' on some paths | |
79 const Type *t = x->Value(&_igvn); | |
80 | |
81 bool singleton = t->singleton(); | |
82 | |
83 // A TOP singleton indicates that there are no possible values incoming | |
84 // along a particular edge. In most cases, this is OK, and the Phi will | |
85 // be eliminated later in an Ideal call. However, we can't allow this to | |
86 // happen if the singleton occurs on loop entry, as the elimination of | |
87 // the PhiNode may cause the resulting node to migrate back to a previous | |
88 // loop iteration. | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
89 if (singleton && t == Type::TOP) { |
0 | 90 // Is_Loop() == false does not confirm the absence of a loop (e.g., an |
91 // irreducible loop may not be indicated by an affirmative is_Loop()); | |
92 // therefore, the only top we can split thru a phi is on a backedge of | |
93 // a loop. | |
94 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); | |
95 } | |
96 | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
97 if (singleton) { |
0 | 98 wins++; |
99 x = ((PhaseGVN&)_igvn).makecon(t); | |
100 } else { | |
101 // We now call Identity to try to simplify the cloned node. | |
102 // Note that some Identity methods call phase->type(this). | |
103 // Make sure that the type array is big enough for | |
104 // our new node, even though we may throw the node away. | |
105 // (Note: This tweaking with igvn only works because x is a new node.) | |
106 _igvn.set_type(x, t); | |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
107 // If x is a TypeNode, capture any more-precise type permanently into Node |
605 | 108 // otherwise it will be not updated during igvn->transform since |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
109 // igvn->type(x) is set to x->Value() already. |
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
110 x->raise_bottom_type(t); |
0 | 111 Node *y = x->Identity(&_igvn); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
112 if (y != x) { |
0 | 113 wins++; |
114 x = y; | |
115 } else { | |
116 y = _igvn.hash_find(x); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
117 if (y) { |
0 | 118 wins++; |
119 x = y; | |
120 } else { | |
121 // Else x is a new node we are keeping | |
122 // We do not need register_new_node_with_optimizer | |
123 // because set_type has already been called. | |
124 _igvn._worklist.push(x); | |
125 } | |
126 } | |
127 } | |
128 if (x != the_clone && the_clone != NULL) | |
129 _igvn.remove_dead_node(the_clone); | |
130 phi->set_req( i, x ); | |
131 } | |
132 // Too few wins? | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
133 if (wins <= policy) { |
0 | 134 _igvn.remove_dead_node(phi); |
135 return NULL; | |
136 } | |
137 | |
138 // Record Phi | |
139 register_new_node( phi, region ); | |
140 | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
141 for (uint i2 = 1; i2 < phi->req(); i2++) { |
0 | 142 Node *x = phi->in(i2); |
143 // If we commoned up the cloned 'x' with another existing Node, | |
144 // the existing Node picks up a new use. We need to make the | |
145 // existing Node occur higher up so it dominates its uses. | |
146 Node *old_ctrl; | |
147 IdealLoopTree *old_loop; | |
148 | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
149 if (x->is_Con()) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
150 // Constant's control is always root. |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
151 set_ctrl(x, C->root()); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
152 continue; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
153 } |
0 | 154 // The occasional new node |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
155 if (x->_idx >= old_unique) { // Found a new, unplaced node? |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
156 old_ctrl = NULL; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
157 old_loop = NULL; // Not in any prior loop |
0 | 158 } else { |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
159 old_ctrl = get_ctrl(x); |
0 | 160 old_loop = get_loop(old_ctrl); // Get prior loop |
161 } | |
162 // New late point must dominate new use | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
163 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
164 if (new_ctrl == old_ctrl) // Nothing is changed |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
165 continue; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
166 |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
167 IdealLoopTree *new_loop = get_loop(new_ctrl); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
168 |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
169 // Don't move x into a loop if its uses are |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
170 // outside of loop. Otherwise x will be cloned |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
171 // for each use outside of this loop. |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
172 IdealLoopTree *use_loop = get_loop(region); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
173 if (!new_loop->is_member(use_loop) && |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
174 (old_loop == NULL || !new_loop->is_member(old_loop))) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
175 // Take early control, later control will be recalculated |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
176 // during next iteration of loop optimizations. |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
177 new_ctrl = get_early_ctrl(x); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
178 new_loop = get_loop(new_ctrl); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
179 } |
0 | 180 // Set new location |
181 set_ctrl(x, new_ctrl); | |
182 // If changing loop bodies, see if we need to collect into new body | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
183 if (old_loop != new_loop) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
184 if (old_loop && !old_loop->_child) |
0 | 185 old_loop->_body.yank(x); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
186 if (!new_loop->_child) |
0 | 187 new_loop->_body.push(x); // Collect body info |
188 } | |
189 } | |
190 | |
191 return phi; | |
192 } | |
193 | |
194 //------------------------------dominated_by------------------------------------ | |
195 // Replace the dominated test with an obvious true or false. Place it on the | |
196 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the | |
197 // live path up to the dominating control. | |
3840
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
198 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { |
0 | 199 #ifndef PRODUCT |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
200 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test"); |
0 | 201 #endif |
202 | |
203 | |
204 // prevdom is the dominating projection of the dominating test. | |
205 assert( iff->is_If(), "" ); | |
206 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); | |
207 int pop = prevdom->Opcode(); | |
208 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
209 if (flip) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
210 if (pop == Op_IfTrue) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
211 pop = Op_IfFalse; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
212 else |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
213 pop = Op_IfTrue; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
214 } |
0 | 215 // 'con' is set to true or false to kill the dominated test. |
216 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); | |
217 set_ctrl(con, C->root()); // Constant gets a new use | |
218 // Hack the dominated test | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
219 _igvn.replace_input_of(iff, 1, con); |
0 | 220 |
221 // If I dont have a reachable TRUE and FALSE path following the IfNode then | |
222 // I can assume this path reaches an infinite loop. In this case it's not | |
223 // important to optimize the data Nodes - either the whole compilation will | |
224 // be tossed or this path (and all data Nodes) will go dead. | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
225 if (iff->outcnt() != 2) return; |
0 | 226 |
227 // Make control-dependent data Nodes on the live path (path that will remain | |
228 // once the dominated IF is removed) become control-dependent on the | |
229 // dominating projection. | |
3840
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
230 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
231 |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
232 // Loop predicates may have depending checks which should not |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
233 // be skipped. For example, range check predicate has two checks |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
234 // for lower and upper bounds. |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
235 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp->as_Proj()->_con)->as_Proj(); |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
236 if (exclude_loop_predicate && |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
237 is_uncommon_trap_proj(unc_proj, Deoptimization::Reason_predicate)) |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
238 return; // Let IGVN transformation change control dependence. |
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
239 |
0 | 240 IdealLoopTree *old_loop = get_loop(dp); |
241 | |
242 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { | |
243 Node* cd = dp->fast_out(i); // Control-dependent node | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
244 if (cd->depends_only_on_test()) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
245 assert(cd->in(0) == dp, ""); |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
246 _igvn.replace_input_of(cd, 0, prevdom); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
247 set_early_ctrl(cd); |
0 | 248 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
249 if (old_loop != new_loop) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
250 if (!old_loop->_child) old_loop->_body.yank(cd); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
251 if (!new_loop->_child) new_loop->_body.push(cd); |
0 | 252 } |
253 --i; | |
254 --imax; | |
255 } | |
256 } | |
257 } | |
258 | |
259 //------------------------------has_local_phi_input---------------------------- | |
260 // Return TRUE if 'n' has Phi inputs from its local block and no other | |
261 // block-local inputs (all non-local-phi inputs come from earlier blocks) | |
262 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { | |
263 Node *n_ctrl = get_ctrl(n); | |
264 // See if some inputs come from a Phi in this block, or from before | |
265 // this block. | |
266 uint i; | |
267 for( i = 1; i < n->req(); i++ ) { | |
268 Node *phi = n->in(i); | |
269 if( phi->is_Phi() && phi->in(0) == n_ctrl ) | |
270 break; | |
271 } | |
272 if( i >= n->req() ) | |
273 return NULL; // No Phi inputs; nowhere to clone thru | |
274 | |
275 // Check for inputs created between 'n' and the Phi input. These | |
276 // must split as well; they have already been given the chance | |
277 // (courtesy of a post-order visit) and since they did not we must | |
278 // recover the 'cost' of splitting them by being very profitable | |
279 // when splitting 'n'. Since this is unlikely we simply give up. | |
280 for( i = 1; i < n->req(); i++ ) { | |
281 Node *m = n->in(i); | |
282 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { | |
283 // We allow the special case of AddP's with no local inputs. | |
284 // This allows us to split-up address expressions. | |
285 if (m->is_AddP() && | |
286 get_ctrl(m->in(2)) != n_ctrl && | |
287 get_ctrl(m->in(3)) != n_ctrl) { | |
288 // Move the AddP up to dominating point | |
289 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); | |
290 continue; | |
291 } | |
292 return NULL; | |
293 } | |
294 } | |
295 | |
296 return n_ctrl; | |
297 } | |
298 | |
299 //------------------------------remix_address_expressions---------------------- | |
300 // Rework addressing expressions to get the most loop-invariant stuff | |
301 // moved out. We'd like to do all associative operators, but it's especially | |
302 // important (common) to do address expressions. | |
303 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { | |
304 if (!has_ctrl(n)) return NULL; | |
305 Node *n_ctrl = get_ctrl(n); | |
306 IdealLoopTree *n_loop = get_loop(n_ctrl); | |
307 | |
308 // See if 'n' mixes loop-varying and loop-invariant inputs and | |
309 // itself is loop-varying. | |
310 | |
311 // Only interested in binary ops (and AddP) | |
312 if( n->req() < 3 || n->req() > 4 ) return NULL; | |
313 | |
314 Node *n1_ctrl = get_ctrl(n->in( 1)); | |
315 Node *n2_ctrl = get_ctrl(n->in( 2)); | |
316 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); | |
317 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); | |
318 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); | |
319 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); | |
320 | |
321 // Does one of my inputs spin in a tighter loop than self? | |
322 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || | |
323 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || | |
324 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) | |
325 return NULL; // Leave well enough alone | |
326 | |
327 // Is at least one of my inputs loop-invariant? | |
328 if( n1_loop == n_loop && | |
329 n2_loop == n_loop && | |
330 n3_loop == n_loop ) | |
331 return NULL; // No loop-invariant inputs | |
332 | |
333 | |
334 int n_op = n->Opcode(); | |
335 | |
336 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). | |
337 if( n_op == Op_LShiftI ) { | |
338 // Scale is loop invariant | |
339 Node *scale = n->in(2); | |
340 Node *scale_ctrl = get_ctrl(scale); | |
341 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); | |
342 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) | |
343 return NULL; | |
344 const TypeInt *scale_t = scale->bottom_type()->isa_int(); | |
345 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) | |
346 return NULL; // Dont bother with byte/short masking | |
347 // Add must vary with loop (else shift would be loop-invariant) | |
348 Node *add = n->in(1); | |
349 Node *add_ctrl = get_ctrl(add); | |
350 IdealLoopTree *add_loop = get_loop(add_ctrl); | |
351 //assert( n_loop == add_loop, "" ); | |
352 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops | |
353 | |
354 // Convert I-V into I+ (0-V); same for V-I | |
355 if( add->Opcode() == Op_SubI && | |
356 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { | |
357 Node *zero = _igvn.intcon(0); | |
358 set_ctrl(zero, C->root()); | |
359 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) ); | |
360 register_new_node( neg, get_ctrl(add->in(2) ) ); | |
361 add = new (C, 3) AddINode( add->in(1), neg ); | |
362 register_new_node( add, add_ctrl ); | |
363 } | |
364 if( add->Opcode() != Op_AddI ) return NULL; | |
365 // See if one add input is loop invariant | |
366 Node *add_var = add->in(1); | |
367 Node *add_var_ctrl = get_ctrl(add_var); | |
368 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); | |
369 Node *add_invar = add->in(2); | |
370 Node *add_invar_ctrl = get_ctrl(add_invar); | |
371 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); | |
372 if( add_var_loop == n_loop ) { | |
373 } else if( add_invar_loop == n_loop ) { | |
374 // Swap to find the invariant part | |
375 add_invar = add_var; | |
376 add_invar_ctrl = add_var_ctrl; | |
377 add_invar_loop = add_var_loop; | |
378 add_var = add->in(2); | |
379 Node *add_var_ctrl = get_ctrl(add_var); | |
380 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); | |
381 } else // Else neither input is loop invariant | |
382 return NULL; | |
383 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) | |
384 return NULL; // No invariant part of the add? | |
385 | |
386 // Yes! Reshape address expression! | |
387 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale ); | |
850
fd50a67f97d1
6860469: remix_address_expressions sets incorrect control causing crash in split_if_with_block_post
never
parents:
834
diff
changeset
|
388 Node *inv_scale_ctrl = |
fd50a67f97d1
6860469: remix_address_expressions sets incorrect control causing crash in split_if_with_block_post
never
parents:
834
diff
changeset
|
389 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? |
fd50a67f97d1
6860469: remix_address_expressions sets incorrect control causing crash in split_if_with_block_post
never
parents:
834
diff
changeset
|
390 add_invar_ctrl : scale_ctrl; |
fd50a67f97d1
6860469: remix_address_expressions sets incorrect control causing crash in split_if_with_block_post
never
parents:
834
diff
changeset
|
391 register_new_node( inv_scale, inv_scale_ctrl ); |
0 | 392 Node *var_scale = new (C, 3) LShiftINode( add_var, scale ); |
393 register_new_node( var_scale, n_ctrl ); | |
394 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale ); | |
395 register_new_node( var_add, n_ctrl ); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
396 _igvn.replace_node( n, var_add ); |
0 | 397 return var_add; |
398 } | |
399 | |
400 // Replace (I+V) with (V+I) | |
401 if( n_op == Op_AddI || | |
402 n_op == Op_AddL || | |
403 n_op == Op_AddF || | |
404 n_op == Op_AddD || | |
405 n_op == Op_MulI || | |
406 n_op == Op_MulL || | |
407 n_op == Op_MulF || | |
408 n_op == Op_MulD ) { | |
409 if( n2_loop == n_loop ) { | |
410 assert( n1_loop != n_loop, "" ); | |
411 n->swap_edges(1, 2); | |
412 } | |
413 } | |
414 | |
415 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), | |
416 // but not if I2 is a constant. | |
417 if( n_op == Op_AddP ) { | |
418 if( n2_loop == n_loop && n3_loop != n_loop ) { | |
419 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { | |
420 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); | |
421 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); | |
422 IdealLoopTree *n22loop = get_loop( n22_ctrl ); | |
423 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); | |
424 if( n22loop != n_loop && n22loop->is_member(n_loop) && | |
425 n23_loop == n_loop ) { | |
426 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); | |
427 // Stuff new AddP in the loop preheader | |
428 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); | |
429 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); | |
430 register_new_node( add2, n_ctrl ); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
431 _igvn.replace_node( n, add2 ); |
0 | 432 return add2; |
433 } | |
434 } | |
435 } | |
436 | |
437 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) | |
438 if( n2_loop != n_loop && n3_loop == n_loop ) { | |
439 if( n->in(3)->Opcode() == Op_AddI ) { | |
440 Node *V = n->in(3)->in(1); | |
441 Node *I = n->in(3)->in(2); | |
442 if( is_member(n_loop,get_ctrl(V)) ) { | |
443 } else { | |
444 Node *tmp = V; V = I; I = tmp; | |
445 } | |
446 if( !is_member(n_loop,get_ctrl(I)) ) { | |
447 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); | |
448 // Stuff new AddP in the loop preheader | |
449 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); | |
450 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); | |
451 register_new_node( add2, n_ctrl ); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
452 _igvn.replace_node( n, add2 ); |
0 | 453 return add2; |
454 } | |
455 } | |
456 } | |
457 } | |
458 | |
459 return NULL; | |
460 } | |
461 | |
462 //------------------------------conditional_move------------------------------- | |
463 // Attempt to replace a Phi with a conditional move. We have some pretty | |
464 // strict profitability requirements. All Phis at the merge point must | |
465 // be converted, so we can remove the control flow. We need to limit the | |
466 // number of c-moves to a small handful. All code that was in the side-arms | |
467 // of the CFG diamond is now speculatively executed. This code has to be | |
468 // "cheap enough". We are pretty much limited to CFG diamonds that merge | |
469 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. | |
470 Node *PhaseIdealLoop::conditional_move( Node *region ) { | |
471 | |
4047 | 472 assert(region->is_Region(), "sanity check"); |
473 if (region->req() != 3) return NULL; | |
0 | 474 |
475 // Check for CFG diamond | |
476 Node *lp = region->in(1); | |
477 Node *rp = region->in(2); | |
4047 | 478 if (!lp || !rp) return NULL; |
0 | 479 Node *lp_c = lp->in(0); |
4047 | 480 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; |
0 | 481 IfNode *iff = lp_c->as_If(); |
482 | |
483 // Check for ops pinned in an arm of the diamond. | |
484 // Can't remove the control flow in this case | |
4047 | 485 if (lp->outcnt() > 1) return NULL; |
486 if (rp->outcnt() > 1) return NULL; | |
487 | |
488 IdealLoopTree* r_loop = get_loop(region); | |
489 assert(r_loop == get_loop(iff), "sanity"); | |
490 // Always convert to CMOVE if all results are used only outside this loop. | |
491 bool used_inside_loop = (r_loop == _ltree_root); | |
0 | 492 |
493 // Check profitability | |
494 int cost = 0; | |
35
e2ae28d2ce91
6667588: Don't generate duplicated CMP for float/double values
kvn
parents:
0
diff
changeset
|
495 int phis = 0; |
0 | 496 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
497 Node *out = region->fast_out(i); | |
4047 | 498 if (!out->is_Phi()) continue; // Ignore other control edges, etc |
35
e2ae28d2ce91
6667588: Don't generate duplicated CMP for float/double values
kvn
parents:
0
diff
changeset
|
499 phis++; |
0 | 500 PhiNode* phi = out->as_Phi(); |
4047 | 501 BasicType bt = phi->type()->basic_type(); |
502 switch (bt) { | |
503 case T_FLOAT: | |
504 case T_DOUBLE: { | |
505 cost += Matcher::float_cmove_cost(); // Could be very expensive | |
506 break; | |
507 } | |
508 case T_LONG: { | |
509 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's | |
510 } | |
0 | 511 case T_INT: // These all CMOV fine |
4047 | 512 case T_ADDRESS: { // (RawPtr) |
0 | 513 cost++; |
514 break; | |
4047 | 515 } |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
516 case T_NARROWOOP: // Fall through |
0 | 517 case T_OBJECT: { // Base oops are OK, but not derived oops |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
518 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); |
0 | 519 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a |
520 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus | |
521 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we | |
522 // have a Phi for the base here that we convert to a CMOVE all is well | |
523 // and good. But if the base is dead, we'll not make a CMOVE. Later | |
524 // the allocator will have to produce a base by creating a CMOVE of the | |
525 // relevant bases. This puts the allocator in the business of | |
526 // manufacturing expensive instructions, generally a bad plan. | |
527 // Just Say No to Conditionally-Moved Derived Pointers. | |
4047 | 528 if (tp && tp->offset() != 0) |
0 | 529 return NULL; |
530 cost++; | |
531 break; | |
532 } | |
533 default: | |
534 return NULL; // In particular, can't do memory or I/O | |
535 } | |
536 // Add in cost any speculative ops | |
4047 | 537 for (uint j = 1; j < region->req(); j++) { |
0 | 538 Node *proj = region->in(j); |
539 Node *inp = phi->in(j); | |
540 if (get_ctrl(inp) == proj) { // Found local op | |
541 cost++; | |
542 // Check for a chain of dependent ops; these will all become | |
543 // speculative in a CMOV. | |
4047 | 544 for (uint k = 1; k < inp->req(); k++) |
0 | 545 if (get_ctrl(inp->in(k)) == proj) |
4047 | 546 cost += ConditionalMoveLimit; // Too much speculative goo |
0 | 547 } |
548 } | |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
549 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. |
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
550 // This will likely Split-If, a higher-payoff operation. |
0 | 551 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { |
552 Node* use = phi->fast_out(k); | |
4047 | 553 if (use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP()) |
554 cost += ConditionalMoveLimit; | |
555 // Is there a use inside the loop? | |
556 // Note: check only basic types since CMoveP is pinned. | |
557 if (!used_inside_loop && is_java_primitive(bt)) { | |
558 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); | |
559 if (r_loop == u_loop || r_loop->is_member(u_loop)) { | |
560 used_inside_loop = true; | |
561 } | |
562 } | |
0 | 563 } |
564 } | |
35
e2ae28d2ce91
6667588: Don't generate duplicated CMP for float/double values
kvn
parents:
0
diff
changeset
|
565 Node* bol = iff->in(1); |
4047 | 566 assert(bol->Opcode() == Op_Bool, ""); |
35
e2ae28d2ce91
6667588: Don't generate duplicated CMP for float/double values
kvn
parents:
0
diff
changeset
|
567 int cmp_op = bol->in(1)->Opcode(); |
e2ae28d2ce91
6667588: Don't generate duplicated CMP for float/double values
kvn
parents:
0
diff
changeset
|
568 // It is expensive to generate flags from a float compare. |
e2ae28d2ce91
6667588: Don't generate duplicated CMP for float/double values
kvn
parents:
0
diff
changeset
|
569 // Avoid duplicated float compare. |
4047 | 570 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; |
571 | |
572 float infrequent_prob = PROB_UNLIKELY_MAG(3); | |
573 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. | |
574 if (used_inside_loop) { | |
575 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo | |
576 | |
577 // BlockLayoutByFrequency optimization moves infrequent branch | |
578 // from hot path. No point in CMOV'ing in such case (110 is used | |
579 // instead of 100 to take into account not exactness of float value). | |
580 if (BlockLayoutByFrequency) { | |
581 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); | |
582 } | |
583 } | |
584 // Check for highly predictable branch. No point in CMOV'ing if | |
585 // we are going to predict accurately all the time. | |
586 if (iff->_prob < infrequent_prob || | |
587 iff->_prob > (1.0f - infrequent_prob)) | |
588 return NULL; | |
0 | 589 |
590 // -------------- | |
591 // Now replace all Phis with CMOV's | |
592 Node *cmov_ctrl = iff->in(0); | |
593 uint flip = (lp->Opcode() == Op_IfTrue); | |
4047 | 594 while (1) { |
0 | 595 PhiNode* phi = NULL; |
596 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { | |
597 Node *out = region->fast_out(i); | |
598 if (out->is_Phi()) { | |
599 phi = out->as_Phi(); | |
600 break; | |
601 } | |
602 } | |
603 if (phi == NULL) break; | |
604 #ifndef PRODUCT | |
4047 | 605 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV"); |
0 | 606 #endif |
607 // Move speculative ops | |
4047 | 608 for (uint j = 1; j < region->req(); j++) { |
0 | 609 Node *proj = region->in(j); |
610 Node *inp = phi->in(j); | |
611 if (get_ctrl(inp) == proj) { // Found local op | |
612 #ifndef PRODUCT | |
4047 | 613 if (PrintOpto && VerifyLoopOptimizations) { |
0 | 614 tty->print(" speculate: "); |
615 inp->dump(); | |
616 } | |
617 #endif | |
618 set_ctrl(inp, cmov_ctrl); | |
619 } | |
620 } | |
621 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); | |
622 register_new_node( cmov, cmov_ctrl ); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
623 _igvn.replace_node( phi, cmov ); |
0 | 624 #ifndef PRODUCT |
4047 | 625 if (TraceLoopOpts) { |
626 tty->print("CMOV "); | |
627 r_loop->dump_head(); | |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
4047
diff
changeset
|
628 if (Verbose) { |
4047 | 629 bol->in(1)->dump(1); |
630 cmov->dump(1); | |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
4047
diff
changeset
|
631 } |
4047 | 632 } |
633 if (VerifyLoopOptimizations) verify(); | |
0 | 634 #endif |
635 } | |
636 | |
637 // The useless CFG diamond will fold up later; see the optimization in | |
638 // RegionNode::Ideal. | |
639 _igvn._worklist.push(region); | |
640 | |
641 return iff->in(1); | |
642 } | |
643 | |
644 //------------------------------split_if_with_blocks_pre----------------------- | |
645 // Do the real work in a non-recursive function. Data nodes want to be | |
646 // cloned in the pre-order so they can feed each other nicely. | |
647 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { | |
648 // Cloning these guys is unlikely to win | |
649 int n_op = n->Opcode(); | |
650 if( n_op == Op_MergeMem ) return n; | |
651 if( n->is_Proj() ) return n; | |
652 // Do not clone-up CmpFXXX variations, as these are always | |
653 // followed by a CmpI | |
654 if( n->is_Cmp() ) return n; | |
655 // Attempt to use a conditional move instead of a phi/branch | |
656 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { | |
657 Node *cmov = conditional_move( n ); | |
658 if( cmov ) return cmov; | |
659 } | |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
251
diff
changeset
|
660 if( n->is_CFG() || n->is_LoadStore() ) |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
251
diff
changeset
|
661 return n; |
0 | 662 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd |
663 n_op == Op_Opaque2 ) { | |
664 if( !C->major_progress() ) // If chance of no more loop opts... | |
665 _igvn._worklist.push(n); // maybe we'll remove them | |
666 return n; | |
667 } | |
668 | |
669 if( n->is_Con() ) return n; // No cloning for Con nodes | |
670 | |
671 Node *n_ctrl = get_ctrl(n); | |
672 if( !n_ctrl ) return n; // Dead node | |
673 | |
674 // Attempt to remix address expressions for loop invariants | |
675 Node *m = remix_address_expressions( n ); | |
676 if( m ) return m; | |
677 | |
678 // Determine if the Node has inputs from some local Phi. | |
679 // Returns the block to clone thru. | |
680 Node *n_blk = has_local_phi_input( n ); | |
681 if( !n_blk ) return n; | |
682 // Do not clone the trip counter through on a CountedLoop | |
683 // (messes up the canonical shape). | |
684 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; | |
685 | |
686 // Check for having no control input; not pinned. Allow | |
687 // dominating control. | |
688 if( n->in(0) ) { | |
689 Node *dom = idom(n_blk); | |
690 if( dom_lca( n->in(0), dom ) != n->in(0) ) | |
691 return n; | |
692 } | |
693 // Policy: when is it profitable. You must get more wins than | |
694 // policy before it is considered profitable. Policy is usually 0, | |
695 // so 1 win is considered profitable. Big merges will require big | |
696 // cloning, so get a larger policy. | |
697 int policy = n_blk->req() >> 2; | |
698 | |
699 // If the loop is a candidate for range check elimination, | |
700 // delay splitting through it's phi until a later loop optimization | |
701 if (n_blk->is_CountedLoop()) { | |
702 IdealLoopTree *lp = get_loop(n_blk); | |
703 if (lp && lp->_rce_candidate) { | |
704 return n; | |
705 } | |
706 } | |
707 | |
708 // Use same limit as split_if_with_blocks_post | |
709 if( C->unique() > 35000 ) return n; // Method too big | |
710 | |
711 // Split 'n' through the merge point if it is profitable | |
712 Node *phi = split_thru_phi( n, n_blk, policy ); | |
4047 | 713 if (!phi) return n; |
0 | 714 |
715 // Found a Phi to split thru! | |
716 // Replace 'n' with the new phi | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
717 _igvn.replace_node( n, phi ); |
0 | 718 // Moved a load around the loop, 'en-registering' something. |
4047 | 719 if (n_blk->is_Loop() && n->is_Load() && |
720 !phi->in(LoopNode::LoopBackControl)->is_Load()) | |
0 | 721 C->set_major_progress(); |
722 | |
723 return phi; | |
724 } | |
725 | |
726 static bool merge_point_too_heavy(Compile* C, Node* region) { | |
727 // Bail out if the region and its phis have too many users. | |
728 int weight = 0; | |
729 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { | |
730 weight += region->fast_out(i)->outcnt(); | |
731 } | |
732 int nodes_left = MaxNodeLimit - C->unique(); | |
733 if (weight * 8 > nodes_left) { | |
734 #ifndef PRODUCT | |
735 if (PrintOpto) | |
736 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); | |
737 #endif | |
738 return true; | |
739 } else { | |
740 return false; | |
741 } | |
742 } | |
743 | |
744 static bool merge_point_safe(Node* region) { | |
745 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode | |
746 // having a PhiNode input. This sidesteps the dangerous case where the split | |
747 // ConvI2LNode may become TOP if the input Value() does not | |
748 // overlap the ConvI2L range, leaving a node which may not dominate its | |
749 // uses. | |
750 // A better fix for this problem can be found in the BugTraq entry, but | |
751 // expediency for Mantis demands this hack. | |
834
0f2d888530e7
6855164: SIGSEGV during compilation of method involving loop over CharSequence.
cfang
parents:
605
diff
changeset
|
752 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop |
0f2d888530e7
6855164: SIGSEGV during compilation of method involving loop over CharSequence.
cfang
parents:
605
diff
changeset
|
753 // split_if_with_blocks from splitting a block because we could not move around |
0f2d888530e7
6855164: SIGSEGV during compilation of method involving loop over CharSequence.
cfang
parents:
605
diff
changeset
|
754 // the FastLockNode. |
0 | 755 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
756 Node* n = region->fast_out(i); | |
757 if (n->is_Phi()) { | |
758 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
759 Node* m = n->fast_out(j); | |
834
0f2d888530e7
6855164: SIGSEGV during compilation of method involving loop over CharSequence.
cfang
parents:
605
diff
changeset
|
760 if (m->is_FastLock()) |
0 | 761 return false; |
834
0f2d888530e7
6855164: SIGSEGV during compilation of method involving loop over CharSequence.
cfang
parents:
605
diff
changeset
|
762 #ifdef _LP64 |
0f2d888530e7
6855164: SIGSEGV during compilation of method involving loop over CharSequence.
cfang
parents:
605
diff
changeset
|
763 if (m->Opcode() == Op_ConvI2L) |
0f2d888530e7
6855164: SIGSEGV during compilation of method involving loop over CharSequence.
cfang
parents:
605
diff
changeset
|
764 return false; |
0f2d888530e7
6855164: SIGSEGV during compilation of method involving loop over CharSequence.
cfang
parents:
605
diff
changeset
|
765 #endif |
0 | 766 } |
767 } | |
768 } | |
769 return true; | |
770 } | |
771 | |
772 | |
773 //------------------------------place_near_use--------------------------------- | |
774 // Place some computation next to use but not inside inner loops. | |
775 // For inner loop uses move it to the preheader area. | |
776 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { | |
777 IdealLoopTree *u_loop = get_loop( useblock ); | |
778 return (u_loop->_irreducible || u_loop->_child) | |
779 ? useblock | |
780 : u_loop->_head->in(LoopNode::EntryControl); | |
781 } | |
782 | |
783 | |
784 //------------------------------split_if_with_blocks_post---------------------- | |
785 // Do the real work in a non-recursive function. CFG hackery wants to be | |
786 // in the post-order, so it can dirty the I-DOM info and not use the dirtied | |
787 // info. | |
788 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { | |
789 | |
790 // Cloning Cmp through Phi's involves the split-if transform. | |
791 // FastLock is not used by an If | |
792 if( n->is_Cmp() && !n->is_FastLock() ) { | |
793 if( C->unique() > 35000 ) return; // Method too big | |
794 | |
795 // Do not do 'split-if' if irreducible loops are present. | |
796 if( _has_irreducible_loops ) | |
797 return; | |
798 | |
799 Node *n_ctrl = get_ctrl(n); | |
800 // Determine if the Node has inputs from some local Phi. | |
801 // Returns the block to clone thru. | |
802 Node *n_blk = has_local_phi_input( n ); | |
803 if( n_blk != n_ctrl ) return; | |
804 | |
805 if( merge_point_too_heavy(C, n_ctrl) ) | |
806 return; | |
807 | |
808 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? | |
809 Node *bol = n->unique_out(); | |
810 assert( bol->is_Bool(), "expect a bool here" ); | |
811 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? | |
812 Node *iff = bol->unique_out(); | |
813 | |
814 // Check some safety conditions | |
815 if( iff->is_If() ) { // Classic split-if? | |
816 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if | |
817 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE | |
4779
c8d8e124380c
7064302: JDK7 build 147 crashed after testing my java 6-compiled web app
kvn
parents:
4064
diff
changeset
|
818 // Can't split CMove with different control edge. |
c8d8e124380c
7064302: JDK7 build 147 crashed after testing my java 6-compiled web app
kvn
parents:
4064
diff
changeset
|
819 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return; |
0 | 820 if( get_ctrl(iff->in(2)) == n_ctrl || |
821 get_ctrl(iff->in(3)) == n_ctrl ) | |
822 return; // Inputs not yet split-up | |
823 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { | |
824 return; // Loop-invar test gates loop-varying CMOVE | |
825 } | |
826 } else { | |
827 return; // some other kind of node, such as an Allocate | |
828 } | |
829 | |
830 // Do not do 'split-if' if some paths are dead. First do dead code | |
831 // elimination and then see if its still profitable. | |
832 for( uint i = 1; i < n_ctrl->req(); i++ ) | |
833 if( n_ctrl->in(i) == C->top() ) | |
834 return; | |
835 | |
836 // When is split-if profitable? Every 'win' on means some control flow | |
837 // goes dead, so it's almost always a win. | |
838 int policy = 0; | |
839 // If trying to do a 'Split-If' at the loop head, it is only | |
840 // profitable if the cmp folds up on BOTH paths. Otherwise we | |
841 // risk peeling a loop forever. | |
842 | |
843 // CNC - Disabled for now. Requires careful handling of loop | |
844 // body selection for the cloned code. Also, make sure we check | |
845 // for any input path not being in the same loop as n_ctrl. For | |
846 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' | |
847 // because the alternative loop entry points won't be converted | |
848 // into LoopNodes. | |
849 IdealLoopTree *n_loop = get_loop(n_ctrl); | |
850 for( uint j = 1; j < n_ctrl->req(); j++ ) | |
851 if( get_loop(n_ctrl->in(j)) != n_loop ) | |
852 return; | |
853 | |
854 // Check for safety of the merge point. | |
855 if( !merge_point_safe(n_ctrl) ) { | |
856 return; | |
857 } | |
858 | |
859 // Split compare 'n' through the merge point if it is profitable | |
860 Node *phi = split_thru_phi( n, n_ctrl, policy ); | |
861 if( !phi ) return; | |
862 | |
863 // Found a Phi to split thru! | |
864 // Replace 'n' with the new phi | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
865 _igvn.replace_node( n, phi ); |
0 | 866 |
867 // Now split the bool up thru the phi | |
868 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
869 _igvn.replace_node( bol, bolphi ); |
0 | 870 assert( iff->in(1) == bolphi, "" ); |
871 if( bolphi->Value(&_igvn)->singleton() ) | |
872 return; | |
873 | |
874 // Conditional-move? Must split up now | |
875 if( !iff->is_If() ) { | |
876 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
877 _igvn.replace_node( iff, cmovphi ); |
0 | 878 return; |
879 } | |
880 | |
881 // Now split the IF | |
882 do_split_if( iff ); | |
883 return; | |
884 } | |
885 | |
886 // Check for an IF ready to split; one that has its | |
887 // condition codes input coming from a Phi at the block start. | |
888 int n_op = n->Opcode(); | |
889 | |
890 // Check for an IF being dominated by another IF same test | |
891 if( n_op == Op_If ) { | |
892 Node *bol = n->in(1); | |
893 uint max = bol->outcnt(); | |
894 // Check for same test used more than once? | |
895 if( n_op == Op_If && max > 1 && bol->is_Bool() ) { | |
896 // Search up IDOMs to see if this IF is dominated. | |
897 Node *cutoff = get_ctrl(bol); | |
898 | |
899 // Now search up IDOMs till cutoff, looking for a dominating test | |
900 Node *prevdom = n; | |
901 Node *dom = idom(prevdom); | |
902 while( dom != cutoff ) { | |
903 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { | |
904 // Replace the dominated test with an obvious true or false. | |
905 // Place it on the IGVN worklist for later cleanup. | |
906 C->set_major_progress(); | |
3840
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3281
diff
changeset
|
907 dominated_by( prevdom, n, false, true ); |
0 | 908 #ifndef PRODUCT |
909 if( VerifyLoopOptimizations ) verify(); | |
910 #endif | |
911 return; | |
912 } | |
913 prevdom = dom; | |
914 dom = idom(prevdom); | |
915 } | |
916 } | |
917 } | |
918 | |
919 // See if a shared loop-varying computation has no loop-varying uses. | |
920 // Happens if something is only used for JVM state in uncommon trap exits, | |
921 // like various versions of induction variable+offset. Clone the | |
922 // computation per usage to allow it to sink out of the loop. | |
923 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) | |
924 Node *n_ctrl = get_ctrl(n); | |
925 IdealLoopTree *n_loop = get_loop(n_ctrl); | |
926 if( n_loop != _ltree_root ) { | |
927 DUIterator_Fast imax, i = n->fast_outs(imax); | |
928 for (; i < imax; i++) { | |
929 Node* u = n->fast_out(i); | |
930 if( !has_ctrl(u) ) break; // Found control user | |
931 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); | |
932 if( u_loop == n_loop ) break; // Found loop-varying use | |
933 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop | |
934 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 | |
935 } | |
936 bool did_break = (i < imax); // Did we break out of the previous loop? | |
937 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! | |
4779
c8d8e124380c
7064302: JDK7 build 147 crashed after testing my java 6-compiled web app
kvn
parents:
4064
diff
changeset
|
938 Node *late_load_ctrl = NULL; |
0 | 939 if (n->is_Load()) { |
940 // If n is a load, get and save the result from get_late_ctrl(), | |
941 // to be later used in calculating the control for n's clones. | |
942 clear_dom_lca_tags(); | |
943 late_load_ctrl = get_late_ctrl(n, n_ctrl); | |
944 } | |
945 // If n is a load, and the late control is the same as the current | |
946 // control, then the cloning of n is a pointless exercise, because | |
947 // GVN will ensure that we end up where we started. | |
948 if (!n->is_Load() || late_load_ctrl != n_ctrl) { | |
949 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { | |
950 Node *u = n->last_out(j); // Clone private computation per use | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
951 _igvn.rehash_node_delayed(u); |
0 | 952 Node *x = n->clone(); // Clone computation |
953 Node *x_ctrl = NULL; | |
954 if( u->is_Phi() ) { | |
955 // Replace all uses of normal nodes. Replace Phi uses | |
605 | 956 // individually, so the separate Nodes can sink down |
0 | 957 // different paths. |
958 uint k = 1; | |
959 while( u->in(k) != n ) k++; | |
960 u->set_req( k, x ); | |
961 // x goes next to Phi input path | |
962 x_ctrl = u->in(0)->in(k); | |
963 --j; | |
964 } else { // Normal use | |
965 // Replace all uses | |
966 for( uint k = 0; k < u->req(); k++ ) { | |
967 if( u->in(k) == n ) { | |
968 u->set_req( k, x ); | |
969 --j; | |
970 } | |
971 } | |
972 x_ctrl = get_ctrl(u); | |
973 } | |
974 | |
975 // Find control for 'x' next to use but not inside inner loops. | |
976 // For inner loop uses get the preheader area. | |
977 x_ctrl = place_near_use(x_ctrl); | |
978 | |
979 if (n->is_Load()) { | |
980 // For loads, add a control edge to a CFG node outside of the loop | |
981 // to force them to not combine and return back inside the loop | |
982 // during GVN optimization (4641526). | |
983 // | |
984 // Because we are setting the actual control input, factor in | |
985 // the result from get_late_ctrl() so we respect any | |
986 // anti-dependences. (6233005). | |
987 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); | |
988 | |
989 // Don't allow the control input to be a CFG splitting node. | |
990 // Such nodes should only have ProjNodes as outs, e.g. IfNode | |
991 // should only have IfTrueNode and IfFalseNode (4985384). | |
992 x_ctrl = find_non_split_ctrl(x_ctrl); | |
993 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); | |
994 | |
995 x->set_req(0, x_ctrl); | |
996 } | |
997 register_new_node(x, x_ctrl); | |
998 | |
999 // Some institutional knowledge is needed here: 'x' is | |
1000 // yanked because if the optimizer runs GVN on it all the | |
1001 // cloned x's will common up and undo this optimization and | |
1002 // be forced back in the loop. This is annoying because it | |
1003 // makes +VerifyOpto report false-positives on progress. I | |
1004 // tried setting control edges on the x's to force them to | |
1005 // not combine, but the matching gets worried when it tries | |
1006 // to fold a StoreP and an AddP together (as part of an | |
1007 // address expression) and the AddP and StoreP have | |
1008 // different controls. | |
318
60bc5071073f
6738933: assert with base pointers must match with compressed oops enabled
never
parents:
293
diff
changeset
|
1009 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x); |
0 | 1010 } |
1011 _igvn.remove_dead_node(n); | |
1012 } | |
1013 } | |
1014 } | |
1015 } | |
1016 | |
1017 // Check for Opaque2's who's loop has disappeared - who's input is in the | |
1018 // same loop nest as their output. Remove 'em, they are no longer useful. | |
1019 if( n_op == Op_Opaque2 && | |
1020 n->in(1) != NULL && | |
1021 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
1022 _igvn.replace_node( n, n->in(1) ); |
0 | 1023 } |
1024 } | |
1025 | |
1026 //------------------------------split_if_with_blocks--------------------------- | |
1027 // Check for aggressive application of 'split-if' optimization, | |
1028 // using basic block level info. | |
1029 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { | |
1030 Node *n = C->root(); | |
1031 visited.set(n->_idx); // first, mark node as visited | |
1032 // Do pre-visit work for root | |
1033 n = split_if_with_blocks_pre( n ); | |
1034 uint cnt = n->outcnt(); | |
1035 uint i = 0; | |
1036 while (true) { | |
1037 // Visit all children | |
1038 if (i < cnt) { | |
1039 Node* use = n->raw_out(i); | |
1040 ++i; | |
1041 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { | |
1042 // Now do pre-visit work for this use | |
1043 use = split_if_with_blocks_pre( use ); | |
1044 nstack.push(n, i); // Save parent and next use's index. | |
1045 n = use; // Process all children of current use. | |
1046 cnt = use->outcnt(); | |
1047 i = 0; | |
1048 } | |
1049 } | |
1050 else { | |
1051 // All of n's children have been processed, complete post-processing. | |
1052 if (cnt != 0 && !n->is_Con()) { | |
1053 assert(has_node(n), "no dead nodes"); | |
1054 split_if_with_blocks_post( n ); | |
1055 } | |
1056 if (nstack.is_empty()) { | |
1057 // Finished all nodes on stack. | |
1058 break; | |
1059 } | |
1060 // Get saved parent node and next use's index. Visit the rest of uses. | |
1061 n = nstack.node(); | |
1062 cnt = n->outcnt(); | |
1063 i = nstack.index(); | |
1064 nstack.pop(); | |
1065 } | |
1066 } | |
1067 } | |
1068 | |
1069 | |
1070 //============================================================================= | |
1071 // | |
1072 // C L O N E A L O O P B O D Y | |
1073 // | |
1074 | |
1075 //------------------------------clone_iff-------------------------------------- | |
1076 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. | |
1077 // "Nearly" because all Nodes have been cloned from the original in the loop, | |
1078 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs | |
1079 // through the Phi recursively, and return a Bool. | |
1080 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { | |
1081 | |
1082 // Convert this Phi into a Phi merging Bools | |
1083 uint i; | |
1084 for( i = 1; i < phi->req(); i++ ) { | |
1085 Node *b = phi->in(i); | |
1086 if( b->is_Phi() ) { | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1087 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); |
0 | 1088 } else { |
1089 assert( b->is_Bool(), "" ); | |
1090 } | |
1091 } | |
1092 | |
1093 Node *sample_bool = phi->in(1); | |
1094 Node *sample_cmp = sample_bool->in(1); | |
1095 | |
1096 // Make Phis to merge the Cmp's inputs. | |
1097 int size = phi->in(0)->req(); | |
1098 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); | |
1099 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); | |
1100 for( i = 1; i < phi->req(); i++ ) { | |
1101 Node *n1 = phi->in(i)->in(1)->in(1); | |
1102 Node *n2 = phi->in(i)->in(1)->in(2); | |
1103 phi1->set_req( i, n1 ); | |
1104 phi2->set_req( i, n2 ); | |
1105 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); | |
1106 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); | |
1107 } | |
1108 // See if these Phis have been made before. | |
1109 // Register with optimizer | |
1110 Node *hit1 = _igvn.hash_find_insert(phi1); | |
1111 if( hit1 ) { // Hit, toss just made Phi | |
1112 _igvn.remove_dead_node(phi1); // Remove new phi | |
1113 assert( hit1->is_Phi(), "" ); | |
1114 phi1 = (PhiNode*)hit1; // Use existing phi | |
1115 } else { // Miss | |
1116 _igvn.register_new_node_with_optimizer(phi1); | |
1117 } | |
1118 Node *hit2 = _igvn.hash_find_insert(phi2); | |
1119 if( hit2 ) { // Hit, toss just made Phi | |
1120 _igvn.remove_dead_node(phi2); // Remove new phi | |
1121 assert( hit2->is_Phi(), "" ); | |
1122 phi2 = (PhiNode*)hit2; // Use existing phi | |
1123 } else { // Miss | |
1124 _igvn.register_new_node_with_optimizer(phi2); | |
1125 } | |
1126 // Register Phis with loop/block info | |
1127 set_ctrl(phi1, phi->in(0)); | |
1128 set_ctrl(phi2, phi->in(0)); | |
1129 // Make a new Cmp | |
1130 Node *cmp = sample_cmp->clone(); | |
1131 cmp->set_req( 1, phi1 ); | |
1132 cmp->set_req( 2, phi2 ); | |
1133 _igvn.register_new_node_with_optimizer(cmp); | |
1134 set_ctrl(cmp, phi->in(0)); | |
1135 | |
1136 // Make a new Bool | |
1137 Node *b = sample_bool->clone(); | |
1138 b->set_req(1,cmp); | |
1139 _igvn.register_new_node_with_optimizer(b); | |
1140 set_ctrl(b, phi->in(0)); | |
1141 | |
1142 assert( b->is_Bool(), "" ); | |
1143 return (BoolNode*)b; | |
1144 } | |
1145 | |
1146 //------------------------------clone_bool------------------------------------- | |
1147 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. | |
1148 // "Nearly" because all Nodes have been cloned from the original in the loop, | |
1149 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs | |
1150 // through the Phi recursively, and return a Bool. | |
1151 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { | |
1152 uint i; | |
1153 // Convert this Phi into a Phi merging Bools | |
1154 for( i = 1; i < phi->req(); i++ ) { | |
1155 Node *b = phi->in(i); | |
1156 if( b->is_Phi() ) { | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1157 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); |
0 | 1158 } else { |
1159 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); | |
1160 } | |
1161 } | |
1162 | |
1163 Node *sample_cmp = phi->in(1); | |
1164 | |
1165 // Make Phis to merge the Cmp's inputs. | |
1166 int size = phi->in(0)->req(); | |
1167 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); | |
1168 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); | |
1169 for( uint j = 1; j < phi->req(); j++ ) { | |
1170 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP | |
1171 Node *n1, *n2; | |
1172 if( cmp_top->is_Cmp() ) { | |
1173 n1 = cmp_top->in(1); | |
1174 n2 = cmp_top->in(2); | |
1175 } else { | |
1176 n1 = n2 = cmp_top; | |
1177 } | |
1178 phi1->set_req( j, n1 ); | |
1179 phi2->set_req( j, n2 ); | |
1180 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); | |
1181 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); | |
1182 } | |
1183 | |
1184 // See if these Phis have been made before. | |
1185 // Register with optimizer | |
1186 Node *hit1 = _igvn.hash_find_insert(phi1); | |
1187 if( hit1 ) { // Hit, toss just made Phi | |
1188 _igvn.remove_dead_node(phi1); // Remove new phi | |
1189 assert( hit1->is_Phi(), "" ); | |
1190 phi1 = (PhiNode*)hit1; // Use existing phi | |
1191 } else { // Miss | |
1192 _igvn.register_new_node_with_optimizer(phi1); | |
1193 } | |
1194 Node *hit2 = _igvn.hash_find_insert(phi2); | |
1195 if( hit2 ) { // Hit, toss just made Phi | |
1196 _igvn.remove_dead_node(phi2); // Remove new phi | |
1197 assert( hit2->is_Phi(), "" ); | |
1198 phi2 = (PhiNode*)hit2; // Use existing phi | |
1199 } else { // Miss | |
1200 _igvn.register_new_node_with_optimizer(phi2); | |
1201 } | |
1202 // Register Phis with loop/block info | |
1203 set_ctrl(phi1, phi->in(0)); | |
1204 set_ctrl(phi2, phi->in(0)); | |
1205 // Make a new Cmp | |
1206 Node *cmp = sample_cmp->clone(); | |
1207 cmp->set_req( 1, phi1 ); | |
1208 cmp->set_req( 2, phi2 ); | |
1209 _igvn.register_new_node_with_optimizer(cmp); | |
1210 set_ctrl(cmp, phi->in(0)); | |
1211 | |
1212 assert( cmp->is_Cmp(), "" ); | |
1213 return (CmpNode*)cmp; | |
1214 } | |
1215 | |
1216 //------------------------------sink_use--------------------------------------- | |
1217 // If 'use' was in the loop-exit block, it now needs to be sunk | |
1218 // below the post-loop merge point. | |
1219 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { | |
1220 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { | |
1221 set_ctrl(use, post_loop); | |
1222 for (DUIterator j = use->outs(); use->has_out(j); j++) | |
1223 sink_use(use->out(j), post_loop); | |
1224 } | |
1225 } | |
1226 | |
1227 //------------------------------clone_loop------------------------------------- | |
1228 // | |
1229 // C L O N E A L O O P B O D Y | |
1230 // | |
1231 // This is the basic building block of the loop optimizations. It clones an | |
1232 // entire loop body. It makes an old_new loop body mapping; with this mapping | |
1233 // you can find the new-loop equivalent to an old-loop node. All new-loop | |
1234 // nodes are exactly equal to their old-loop counterparts, all edges are the | |
1235 // same. All exits from the old-loop now have a RegionNode that merges the | |
1236 // equivalent new-loop path. This is true even for the normal "loop-exit" | |
1237 // condition. All uses of loop-invariant old-loop values now come from (one | |
1238 // or more) Phis that merge their new-loop equivalents. | |
1239 // | |
1240 // This operation leaves the graph in an illegal state: there are two valid | |
1241 // control edges coming from the loop pre-header to both loop bodies. I'll | |
1242 // definitely have to hack the graph after running this transform. | |
1243 // | |
1244 // From this building block I will further edit edges to perform loop peeling | |
1245 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. | |
1246 // | |
1247 // Parameter side_by_size_idom: | |
1248 // When side_by_size_idom is NULL, the dominator tree is constructed for | |
1249 // the clone loop to dominate the original. Used in construction of | |
1250 // pre-main-post loop sequence. | |
1251 // When nonnull, the clone and original are side-by-side, both are | |
1252 // dominated by the side_by_side_idom node. Used in construction of | |
1253 // unswitched loops. | |
1254 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, | |
1255 Node* side_by_side_idom) { | |
1256 | |
1257 // Step 1: Clone the loop body. Make the old->new mapping. | |
1258 uint i; | |
1259 for( i = 0; i < loop->_body.size(); i++ ) { | |
1260 Node *old = loop->_body.at(i); | |
1261 Node *nnn = old->clone(); | |
1262 old_new.map( old->_idx, nnn ); | |
1263 _igvn.register_new_node_with_optimizer(nnn); | |
1264 } | |
1265 | |
1266 | |
1267 // Step 2: Fix the edges in the new body. If the old input is outside the | |
1268 // loop use it. If the old input is INside the loop, use the corresponding | |
1269 // new node instead. | |
1270 for( i = 0; i < loop->_body.size(); i++ ) { | |
1271 Node *old = loop->_body.at(i); | |
1272 Node *nnn = old_new[old->_idx]; | |
1273 // Fix CFG/Loop controlling the new node | |
1274 if (has_ctrl(old)) { | |
1275 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); | |
1276 } else { | |
1277 set_loop(nnn, loop->_parent); | |
1278 if (old->outcnt() > 0) { | |
1279 set_idom( nnn, old_new[idom(old)->_idx], dd ); | |
1280 } | |
1281 } | |
1282 // Correct edges to the new node | |
1283 for( uint j = 0; j < nnn->req(); j++ ) { | |
1284 Node *n = nnn->in(j); | |
1285 if( n ) { | |
1286 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); | |
1287 if( loop->is_member( old_in_loop ) ) | |
1288 nnn->set_req(j, old_new[n->_idx]); | |
1289 } | |
1290 } | |
1291 _igvn.hash_find_insert(nnn); | |
1292 } | |
1293 Node *newhead = old_new[loop->_head->_idx]; | |
1294 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); | |
1295 | |
1296 | |
1297 // Step 3: Now fix control uses. Loop varying control uses have already | |
1298 // been fixed up (as part of all input edges in Step 2). Loop invariant | |
1299 // control uses must be either an IfFalse or an IfTrue. Make a merge | |
1300 // point to merge the old and new IfFalse/IfTrue nodes; make the use | |
1301 // refer to this. | |
1302 ResourceArea *area = Thread::current()->resource_area(); | |
1303 Node_List worklist(area); | |
1304 uint new_counter = C->unique(); | |
1305 for( i = 0; i < loop->_body.size(); i++ ) { | |
1306 Node* old = loop->_body.at(i); | |
1307 if( !old->is_CFG() ) continue; | |
1308 Node* nnn = old_new[old->_idx]; | |
1309 | |
1310 // Copy uses to a worklist, so I can munge the def-use info | |
1311 // with impunity. | |
1312 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) | |
1313 worklist.push(old->fast_out(j)); | |
1314 | |
1315 while( worklist.size() ) { // Visit all uses | |
1316 Node *use = worklist.pop(); | |
1317 if (!has_node(use)) continue; // Ignore dead nodes | |
1318 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); | |
1319 if( !loop->is_member( use_loop ) && use->is_CFG() ) { | |
1320 // Both OLD and USE are CFG nodes here. | |
1321 assert( use->is_Proj(), "" ); | |
1322 | |
1323 // Clone the loop exit control projection | |
1324 Node *newuse = use->clone(); | |
1325 newuse->set_req(0,nnn); | |
1326 _igvn.register_new_node_with_optimizer(newuse); | |
1327 set_loop(newuse, use_loop); | |
1328 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); | |
1329 | |
1330 // We need a Region to merge the exit from the peeled body and the | |
1331 // exit from the old loop body. | |
1332 RegionNode *r = new (C, 3) RegionNode(3); | |
1333 // Map the old use to the new merge point | |
1334 old_new.map( use->_idx, r ); | |
1335 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); | |
1336 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); | |
1337 | |
1338 // The original user of 'use' uses 'r' instead. | |
1339 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { | |
1340 Node* useuse = use->last_out(l); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1341 _igvn.rehash_node_delayed(useuse); |
0 | 1342 uint uses_found = 0; |
1343 if( useuse->in(0) == use ) { | |
1344 useuse->set_req(0, r); | |
1345 uses_found++; | |
1346 if( useuse->is_CFG() ) { | |
1347 assert( dom_depth(useuse) > dd_r, "" ); | |
1348 set_idom(useuse, r, dom_depth(useuse)); | |
1349 } | |
1350 } | |
1351 for( uint k = 1; k < useuse->req(); k++ ) { | |
1352 if( useuse->in(k) == use ) { | |
1353 useuse->set_req(k, r); | |
1354 uses_found++; | |
1355 } | |
1356 } | |
1357 l -= uses_found; // we deleted 1 or more copies of this edge | |
1358 } | |
1359 | |
1360 // Now finish up 'r' | |
1361 r->set_req( 1, newuse ); | |
1362 r->set_req( 2, use ); | |
1363 _igvn.register_new_node_with_optimizer(r); | |
1364 set_loop(r, use_loop); | |
1365 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); | |
1366 } // End of if a loop-exit test | |
1367 } | |
1368 } | |
1369 | |
1370 // Step 4: If loop-invariant use is not control, it must be dominated by a | |
1371 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region | |
1372 // there if needed. Make a Phi there merging old and new used values. | |
1373 Node_List *split_if_set = NULL; | |
1374 Node_List *split_bool_set = NULL; | |
1375 Node_List *split_cex_set = NULL; | |
1376 for( i = 0; i < loop->_body.size(); i++ ) { | |
1377 Node* old = loop->_body.at(i); | |
1378 Node* nnn = old_new[old->_idx]; | |
1379 // Copy uses to a worklist, so I can munge the def-use info | |
1380 // with impunity. | |
1381 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) | |
1382 worklist.push(old->fast_out(j)); | |
1383 | |
1384 while( worklist.size() ) { | |
1385 Node *use = worklist.pop(); | |
1386 if (!has_node(use)) continue; // Ignore dead nodes | |
1387 if (use->in(0) == C->top()) continue; | |
1388 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); | |
1389 // Check for data-use outside of loop - at least one of OLD or USE | |
1390 // must not be a CFG node. | |
1391 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { | |
1392 | |
1393 // If the Data use is an IF, that means we have an IF outside of the | |
1394 // loop that is switching on a condition that is set inside of the | |
1395 // loop. Happens if people set a loop-exit flag; then test the flag | |
1396 // in the loop to break the loop, then test is again outside of the | |
1397 // loop to determine which way the loop exited. | |
1398 if( use->is_If() || use->is_CMove() ) { | |
1399 // Since this code is highly unlikely, we lazily build the worklist | |
1400 // of such Nodes to go split. | |
1401 if( !split_if_set ) | |
1402 split_if_set = new Node_List(area); | |
1403 split_if_set->push(use); | |
1404 } | |
1405 if( use->is_Bool() ) { | |
1406 if( !split_bool_set ) | |
1407 split_bool_set = new Node_List(area); | |
1408 split_bool_set->push(use); | |
1409 } | |
1410 if( use->Opcode() == Op_CreateEx ) { | |
1411 if( !split_cex_set ) | |
1412 split_cex_set = new Node_List(area); | |
1413 split_cex_set->push(use); | |
1414 } | |
1415 | |
1416 | |
1417 // Get "block" use is in | |
1418 uint idx = 0; | |
1419 while( use->in(idx) != old ) idx++; | |
1420 Node *prev = use->is_CFG() ? use : get_ctrl(use); | |
1421 assert( !loop->is_member( get_loop( prev ) ), "" ); | |
1422 Node *cfg = prev->_idx >= new_counter | |
1423 ? prev->in(2) | |
1424 : idom(prev); | |
1425 if( use->is_Phi() ) // Phi use is in prior block | |
1426 cfg = prev->in(idx); // NOT in block of Phi itself | |
1427 if (cfg->is_top()) { // Use is dead? | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1428 _igvn.replace_input_of(use, idx, C->top()); |
0 | 1429 continue; |
1430 } | |
1431 | |
1432 while( !loop->is_member( get_loop( cfg ) ) ) { | |
1433 prev = cfg; | |
1434 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); | |
1435 } | |
1436 // If the use occurs after merging several exits from the loop, then | |
1437 // old value must have dominated all those exits. Since the same old | |
1438 // value was used on all those exits we did not need a Phi at this | |
1439 // merge point. NOW we do need a Phi here. Each loop exit value | |
1440 // is now merged with the peeled body exit; each exit gets its own | |
1441 // private Phi and those Phis need to be merged here. | |
1442 Node *phi; | |
1443 if( prev->is_Region() ) { | |
1444 if( idx == 0 ) { // Updating control edge? | |
1445 phi = prev; // Just use existing control | |
1446 } else { // Else need a new Phi | |
1447 phi = PhiNode::make( prev, old ); | |
1448 // Now recursively fix up the new uses of old! | |
1449 for( uint i = 1; i < prev->req(); i++ ) { | |
1450 worklist.push(phi); // Onto worklist once for each 'old' input | |
1451 } | |
1452 } | |
1453 } else { | |
1454 // Get new RegionNode merging old and new loop exits | |
1455 prev = old_new[prev->_idx]; | |
1456 assert( prev, "just made this in step 7" ); | |
1457 if( idx == 0 ) { // Updating control edge? | |
1458 phi = prev; // Just use existing control | |
1459 } else { // Else need a new Phi | |
1460 // Make a new Phi merging data values properly | |
1461 phi = PhiNode::make( prev, old ); | |
1462 phi->set_req( 1, nnn ); | |
1463 } | |
1464 } | |
1465 // If inserting a new Phi, check for prior hits | |
1466 if( idx != 0 ) { | |
1467 Node *hit = _igvn.hash_find_insert(phi); | |
1468 if( hit == NULL ) { | |
1469 _igvn.register_new_node_with_optimizer(phi); // Register new phi | |
1470 } else { // or | |
1471 // Remove the new phi from the graph and use the hit | |
1472 _igvn.remove_dead_node(phi); | |
1473 phi = hit; // Use existing phi | |
1474 } | |
1475 set_ctrl(phi, prev); | |
1476 } | |
1477 // Make 'use' use the Phi instead of the old loop body exit value | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1478 _igvn.replace_input_of(use, idx, phi); |
0 | 1479 if( use->_idx >= new_counter ) { // If updating new phis |
1480 // Not needed for correctness, but prevents a weak assert | |
1481 // in AddPNode from tripping (when we end up with different | |
1482 // base & derived Phis that will become the same after | |
1483 // IGVN does CSE). | |
1484 Node *hit = _igvn.hash_find_insert(use); | |
1485 if( hit ) // Go ahead and re-hash for hits. | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
1486 _igvn.replace_node( use, hit ); |
0 | 1487 } |
1488 | |
1489 // If 'use' was in the loop-exit block, it now needs to be sunk | |
1490 // below the post-loop merge point. | |
1491 sink_use( use, prev ); | |
1492 } | |
1493 } | |
1494 } | |
1495 | |
1496 // Check for IFs that need splitting/cloning. Happens if an IF outside of | |
1497 // the loop uses a condition set in the loop. The original IF probably | |
1498 // takes control from one or more OLD Regions (which in turn get from NEW | |
1499 // Regions). In any case, there will be a set of Phis for each merge point | |
1500 // from the IF up to where the original BOOL def exists the loop. | |
1501 if( split_if_set ) { | |
1502 while( split_if_set->size() ) { | |
1503 Node *iff = split_if_set->pop(); | |
1504 if( iff->in(1)->is_Phi() ) { | |
1505 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1506 _igvn.replace_input_of(iff, 1, b); |
0 | 1507 } |
1508 } | |
1509 } | |
1510 if( split_bool_set ) { | |
1511 while( split_bool_set->size() ) { | |
1512 Node *b = split_bool_set->pop(); | |
1513 Node *phi = b->in(1); | |
1514 assert( phi->is_Phi(), "" ); | |
1515 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1516 _igvn.replace_input_of(b, 1, cmp); |
0 | 1517 } |
1518 } | |
1519 if( split_cex_set ) { | |
1520 while( split_cex_set->size() ) { | |
1521 Node *b = split_cex_set->pop(); | |
1522 assert( b->in(0)->is_Region(), "" ); | |
1523 assert( b->in(1)->is_Phi(), "" ); | |
1524 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); | |
1525 split_up( b, b->in(0), NULL ); | |
1526 } | |
1527 } | |
1528 | |
1529 } | |
1530 | |
1531 | |
1532 //---------------------- stride_of_possible_iv ------------------------------------- | |
1533 // Looks for an iff/bool/comp with one operand of the compare | |
1534 // being a cycle involving an add and a phi, | |
1535 // with an optional truncation (left-shift followed by a right-shift) | |
1536 // of the add. Returns zero if not an iv. | |
1537 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { | |
1538 Node* trunc1 = NULL; | |
1539 Node* trunc2 = NULL; | |
1540 const TypeInt* ttype = NULL; | |
1541 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { | |
1542 return 0; | |
1543 } | |
1544 BoolNode* bl = iff->in(1)->as_Bool(); | |
1545 Node* cmp = bl->in(1); | |
1546 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { | |
1547 return 0; | |
1548 } | |
1549 // Must have an invariant operand | |
1550 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { | |
1551 return 0; | |
1552 } | |
1553 Node* add2 = NULL; | |
1554 Node* cmp1 = cmp->in(1); | |
1555 if (cmp1->is_Phi()) { | |
1556 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) | |
1557 Node* phi = cmp1; | |
1558 for (uint i = 1; i < phi->req(); i++) { | |
1559 Node* in = phi->in(i); | |
1560 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, | |
1561 &trunc1, &trunc2, &ttype); | |
1562 if (add && add->in(1) == phi) { | |
1563 add2 = add->in(2); | |
1564 break; | |
1565 } | |
1566 } | |
1567 } else { | |
1568 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) | |
1569 Node* addtrunc = cmp1; | |
1570 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, | |
1571 &trunc1, &trunc2, &ttype); | |
1572 if (add && add->in(1)->is_Phi()) { | |
1573 Node* phi = add->in(1); | |
1574 for (uint i = 1; i < phi->req(); i++) { | |
1575 if (phi->in(i) == addtrunc) { | |
1576 add2 = add->in(2); | |
1577 break; | |
1578 } | |
1579 } | |
1580 } | |
1581 } | |
1582 if (add2 != NULL) { | |
1583 const TypeInt* add2t = _igvn.type(add2)->is_int(); | |
1584 if (add2t->is_con()) { | |
1585 return add2t->get_con(); | |
1586 } | |
1587 } | |
1588 return 0; | |
1589 } | |
1590 | |
1591 | |
1592 //---------------------- stay_in_loop ------------------------------------- | |
1593 // Return the (unique) control output node that's in the loop (if it exists.) | |
1594 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { | |
1595 Node* unique = NULL; | |
1596 if (!n) return NULL; | |
1597 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
1598 Node* use = n->fast_out(i); | |
1599 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { | |
1600 if (unique != NULL) { | |
1601 return NULL; | |
1602 } | |
1603 unique = use; | |
1604 } | |
1605 } | |
1606 return unique; | |
1607 } | |
1608 | |
1609 //------------------------------ register_node ------------------------------------- | |
1610 // Utility to register node "n" with PhaseIdealLoop | |
1611 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { | |
1612 _igvn.register_new_node_with_optimizer(n); | |
1613 loop->_body.push(n); | |
1614 if (n->is_CFG()) { | |
1615 set_loop(n, loop); | |
1616 set_idom(n, pred, ddepth); | |
1617 } else { | |
1618 set_ctrl(n, pred); | |
1619 } | |
1620 } | |
1621 | |
1622 //------------------------------ proj_clone ------------------------------------- | |
1623 // Utility to create an if-projection | |
1624 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { | |
1625 ProjNode* c = p->clone()->as_Proj(); | |
1626 c->set_req(0, iff); | |
1627 return c; | |
1628 } | |
1629 | |
1630 //------------------------------ short_circuit_if ------------------------------------- | |
1631 // Force the iff control output to be the live_proj | |
1632 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { | |
1633 int proj_con = live_proj->_con; | |
1634 assert(proj_con == 0 || proj_con == 1, "false or true projection"); | |
1635 Node *con = _igvn.intcon(proj_con); | |
1636 set_ctrl(con, C->root()); | |
1637 if (iff) { | |
1638 iff->set_req(1, con); | |
1639 } | |
1640 return con; | |
1641 } | |
1642 | |
1643 //------------------------------ insert_if_before_proj ------------------------------------- | |
1644 // Insert a new if before an if projection (* - new node) | |
1645 // | |
1646 // before | |
1647 // if(test) | |
1648 // / \ | |
1649 // v v | |
1650 // other-proj proj (arg) | |
1651 // | |
1652 // after | |
1653 // if(test) | |
1654 // / \ | |
1655 // / v | |
1656 // | * proj-clone | |
1657 // v | | |
1658 // other-proj v | |
1659 // * new_if(relop(cmp[IU](left,right))) | |
1660 // / \ | |
1661 // v v | |
1662 // * new-proj proj | |
1663 // (returned) | |
1664 // | |
1665 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { | |
1666 IfNode* iff = proj->in(0)->as_If(); | |
1667 IdealLoopTree *loop = get_loop(proj); | |
1668 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); | |
1669 int ddepth = dom_depth(proj); | |
1670 | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1671 _igvn.rehash_node_delayed(iff); |
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1672 _igvn.rehash_node_delayed(proj); |
0 | 1673 |
1674 proj->set_req(0, NULL); // temporary disconnect | |
1675 ProjNode* proj2 = proj_clone(proj, iff); | |
1676 register_node(proj2, loop, iff, ddepth); | |
1677 | |
1678 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); | |
1679 register_node(cmp, loop, proj2, ddepth); | |
1680 | |
1681 BoolNode* bol = new (C,2)BoolNode(cmp, relop); | |
1682 register_node(bol, loop, proj2, ddepth); | |
1683 | |
1684 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); | |
1685 register_node(new_if, loop, proj2, ddepth); | |
1686 | |
1687 proj->set_req(0, new_if); // reattach | |
1688 set_idom(proj, new_if, ddepth); | |
1689 | |
1690 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); | |
1691 register_node(new_exit, get_loop(other_proj), new_if, ddepth); | |
1692 | |
1693 return new_exit; | |
1694 } | |
1695 | |
1696 //------------------------------ insert_region_before_proj ------------------------------------- | |
1697 // Insert a region before an if projection (* - new node) | |
1698 // | |
1699 // before | |
1700 // if(test) | |
1701 // / | | |
1702 // v | | |
1703 // proj v | |
1704 // other-proj | |
1705 // | |
1706 // after | |
1707 // if(test) | |
1708 // / | | |
1709 // v | | |
1710 // * proj-clone v | |
1711 // | other-proj | |
1712 // v | |
1713 // * new-region | |
1714 // | | |
1715 // v | |
1716 // * dum_if | |
1717 // / \ | |
1718 // v \ | |
1719 // * dum-proj v | |
1720 // proj | |
1721 // | |
1722 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { | |
1723 IfNode* iff = proj->in(0)->as_If(); | |
1724 IdealLoopTree *loop = get_loop(proj); | |
1725 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); | |
1726 int ddepth = dom_depth(proj); | |
1727 | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1728 _igvn.rehash_node_delayed(iff); |
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1729 _igvn.rehash_node_delayed(proj); |
0 | 1730 |
1731 proj->set_req(0, NULL); // temporary disconnect | |
1732 ProjNode* proj2 = proj_clone(proj, iff); | |
1733 register_node(proj2, loop, iff, ddepth); | |
1734 | |
1735 RegionNode* reg = new (C,2)RegionNode(2); | |
1736 reg->set_req(1, proj2); | |
1737 register_node(reg, loop, iff, ddepth); | |
1738 | |
1739 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); | |
1740 register_node(dum_if, loop, reg, ddepth); | |
1741 | |
1742 proj->set_req(0, dum_if); // reattach | |
1743 set_idom(proj, dum_if, ddepth); | |
1744 | |
1745 ProjNode* dum_proj = proj_clone(other_proj, dum_if); | |
1746 register_node(dum_proj, loop, dum_if, ddepth); | |
1747 | |
1748 return reg; | |
1749 } | |
1750 | |
1751 //------------------------------ insert_cmpi_loop_exit ------------------------------------- | |
1752 // Clone a signed compare loop exit from an unsigned compare and | |
1753 // insert it before the unsigned cmp on the stay-in-loop path. | |
1754 // All new nodes inserted in the dominator tree between the original | |
1755 // if and it's projections. The original if test is replaced with | |
1756 // a constant to force the stay-in-loop path. | |
1757 // | |
1758 // This is done to make sure that the original if and it's projections | |
1759 // still dominate the same set of control nodes, that the ctrl() relation | |
1760 // from data nodes to them is preserved, and that their loop nesting is | |
1761 // preserved. | |
1762 // | |
1763 // before | |
1764 // if(i <u limit) unsigned compare loop exit | |
1765 // / | | |
1766 // v v | |
1767 // exit-proj stay-in-loop-proj | |
1768 // | |
1769 // after | |
1770 // if(stay-in-loop-const) original if | |
1771 // / | | |
1772 // / v | |
1773 // / if(i < limit) new signed test | |
1774 // / / | | |
1775 // / / v | |
1776 // / / if(i <u limit) new cloned unsigned test | |
1777 // / / / | | |
1778 // v v v | | |
1779 // region | | |
1780 // | | | |
1781 // dum-if | | |
1782 // / | | | |
1783 // ether | | | |
1784 // v v | |
1785 // exit-proj stay-in-loop-proj | |
1786 // | |
1787 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { | |
1788 const bool Signed = true; | |
1789 const bool Unsigned = false; | |
1790 | |
1791 BoolNode* bol = if_cmpu->in(1)->as_Bool(); | |
1792 if (bol->_test._test != BoolTest::lt) return NULL; | |
1793 CmpNode* cmpu = bol->in(1)->as_Cmp(); | |
1794 if (cmpu->Opcode() != Op_CmpU) return NULL; | |
1795 int stride = stride_of_possible_iv(if_cmpu); | |
1796 if (stride == 0) return NULL; | |
1797 | |
1798 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); | |
1799 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); | |
1800 | |
1801 Node* limit = NULL; | |
1802 if (stride > 0) { | |
1803 limit = cmpu->in(2); | |
1804 } else { | |
1805 limit = _igvn.makecon(TypeInt::ZERO); | |
1806 set_ctrl(limit, C->root()); | |
1807 } | |
1808 // Create a new region on the exit path | |
1809 RegionNode* reg = insert_region_before_proj(lp_exit); | |
1810 | |
1811 // Clone the if-cmpu-true-false using a signed compare | |
1812 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; | |
1813 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); | |
1814 reg->add_req(cmpi_exit); | |
1815 | |
1816 // Clone the if-cmpu-true-false | |
1817 BoolTest::mask rel_u = bol->_test._test; | |
1818 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); | |
1819 reg->add_req(cmpu_exit); | |
1820 | |
1821 // Force original if to stay in loop. | |
1822 short_circuit_if(if_cmpu, lp_continue); | |
1823 | |
1824 return cmpi_exit->in(0)->as_If(); | |
1825 } | |
1826 | |
1827 //------------------------------ remove_cmpi_loop_exit ------------------------------------- | |
1828 // Remove a previously inserted signed compare loop exit. | |
1829 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { | |
1830 Node* lp_proj = stay_in_loop(if_cmp, loop); | |
1831 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && | |
1832 stay_in_loop(lp_proj, loop)->is_If() && | |
1833 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); | |
1834 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); | |
1835 set_ctrl(con, C->root()); | |
1836 if_cmp->set_req(1, con); | |
1837 } | |
1838 | |
1839 //------------------------------ scheduled_nodelist ------------------------------------- | |
1840 // Create a post order schedule of nodes that are in the | |
1841 // "member" set. The list is returned in "sched". | |
1842 // The first node in "sched" is the loop head, followed by | |
1843 // nodes which have no inputs in the "member" set, and then | |
1844 // followed by the nodes that have an immediate input dependence | |
1845 // on a node in "sched". | |
1846 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { | |
1847 | |
1848 assert(member.test(loop->_head->_idx), "loop head must be in member set"); | |
1849 Arena *a = Thread::current()->resource_area(); | |
1850 VectorSet visited(a); | |
1851 Node_Stack nstack(a, loop->_body.size()); | |
1852 | |
1853 Node* n = loop->_head; // top of stack is cached in "n" | |
1854 uint idx = 0; | |
1855 visited.set(n->_idx); | |
1856 | |
1857 // Initially push all with no inputs from within member set | |
1858 for(uint i = 0; i < loop->_body.size(); i++ ) { | |
1859 Node *elt = loop->_body.at(i); | |
1860 if (member.test(elt->_idx)) { | |
1861 bool found = false; | |
1862 for (uint j = 0; j < elt->req(); j++) { | |
1863 Node* def = elt->in(j); | |
1864 if (def && member.test(def->_idx) && def != elt) { | |
1865 found = true; | |
1866 break; | |
1867 } | |
1868 } | |
1869 if (!found && elt != loop->_head) { | |
1870 nstack.push(n, idx); | |
1871 n = elt; | |
1872 assert(!visited.test(n->_idx), "not seen yet"); | |
1873 visited.set(n->_idx); | |
1874 } | |
1875 } | |
1876 } | |
1877 | |
1878 // traverse out's that are in the member set | |
1879 while (true) { | |
1880 if (idx < n->outcnt()) { | |
1881 Node* use = n->raw_out(idx); | |
1882 idx++; | |
1883 if (!visited.test_set(use->_idx)) { | |
1884 if (member.test(use->_idx)) { | |
1885 nstack.push(n, idx); | |
1886 n = use; | |
1887 idx = 0; | |
1888 } | |
1889 } | |
1890 } else { | |
1891 // All outputs processed | |
1892 sched.push(n); | |
1893 if (nstack.is_empty()) break; | |
1894 n = nstack.node(); | |
1895 idx = nstack.index(); | |
1896 nstack.pop(); | |
1897 } | |
1898 } | |
1899 } | |
1900 | |
1901 | |
1902 //------------------------------ has_use_in_set ------------------------------------- | |
1903 // Has a use in the vector set | |
1904 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { | |
1905 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
1906 Node* use = n->fast_out(j); | |
1907 if (vset.test(use->_idx)) { | |
1908 return true; | |
1909 } | |
1910 } | |
1911 return false; | |
1912 } | |
1913 | |
1914 | |
1915 //------------------------------ has_use_internal_to_set ------------------------------------- | |
1916 // Has use internal to the vector set (ie. not in a phi at the loop head) | |
1917 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { | |
1918 Node* head = loop->_head; | |
1919 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
1920 Node* use = n->fast_out(j); | |
1921 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { | |
1922 return true; | |
1923 } | |
1924 } | |
1925 return false; | |
1926 } | |
1927 | |
1928 | |
1929 //------------------------------ clone_for_use_outside_loop ------------------------------------- | |
1930 // clone "n" for uses that are outside of loop | |
1931 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { | |
1932 | |
1933 assert(worklist.size() == 0, "should be empty"); | |
1934 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
1935 Node* use = n->fast_out(j); | |
1936 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { | |
1937 worklist.push(use); | |
1938 } | |
1939 } | |
1940 while( worklist.size() ) { | |
1941 Node *use = worklist.pop(); | |
1942 if (!has_node(use) || use->in(0) == C->top()) continue; | |
1943 uint j; | |
1944 for (j = 0; j < use->req(); j++) { | |
1945 if (use->in(j) == n) break; | |
1946 } | |
1947 assert(j < use->req(), "must be there"); | |
1948 | |
1949 // clone "n" and insert it between the inputs of "n" and the use outside the loop | |
1950 Node* n_clone = n->clone(); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
1951 _igvn.replace_input_of(use, j, n_clone); |
251 | 1952 Node* use_c; |
0 | 1953 if (!use->is_Phi()) { |
251 | 1954 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); |
0 | 1955 } else { |
1956 // Use in a phi is considered a use in the associated predecessor block | |
251 | 1957 use_c = use->in(0)->in(j); |
0 | 1958 } |
251 | 1959 set_ctrl(n_clone, use_c); |
1960 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); | |
1961 get_loop(use_c)->_body.push(n_clone); | |
0 | 1962 _igvn.register_new_node_with_optimizer(n_clone); |
1963 #if !defined(PRODUCT) | |
1964 if (TracePartialPeeling) { | |
1965 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); | |
1966 } | |
1967 #endif | |
1968 } | |
1969 } | |
1970 | |
1971 | |
1972 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- | |
1973 // clone "n" for special uses that are in the not_peeled region. | |
1974 // If these def-uses occur in separate blocks, the code generator | |
1975 // marks the method as not compilable. For example, if a "BoolNode" | |
1976 // is in a different basic block than the "IfNode" that uses it, then | |
1977 // the compilation is aborted in the code generator. | |
1978 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, | |
1979 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { | |
1980 if (n->is_Phi() || n->is_Load()) { | |
1981 return; | |
1982 } | |
1983 assert(worklist.size() == 0, "should be empty"); | |
1984 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
1985 Node* use = n->fast_out(j); | |
1986 if ( not_peel.test(use->_idx) && | |
1987 (use->is_If() || use->is_CMove() || use->is_Bool()) && | |
1988 use->in(1) == n) { | |
1989 worklist.push(use); | |
1990 } | |
1991 } | |
1992 if (worklist.size() > 0) { | |
1993 // clone "n" and insert it between inputs of "n" and the use | |
1994 Node* n_clone = n->clone(); | |
1995 loop->_body.push(n_clone); | |
1996 _igvn.register_new_node_with_optimizer(n_clone); | |
1997 set_ctrl(n_clone, get_ctrl(n)); | |
1998 sink_list.push(n_clone); | |
1999 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. | |
2000 #if !defined(PRODUCT) | |
2001 if (TracePartialPeeling) { | |
2002 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); | |
2003 } | |
2004 #endif | |
2005 while( worklist.size() ) { | |
2006 Node *use = worklist.pop(); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2007 _igvn.rehash_node_delayed(use); |
0 | 2008 for (uint j = 1; j < use->req(); j++) { |
2009 if (use->in(j) == n) { | |
2010 use->set_req(j, n_clone); | |
2011 } | |
2012 } | |
2013 } | |
2014 } | |
2015 } | |
2016 | |
2017 | |
2018 //------------------------------ insert_phi_for_loop ------------------------------------- | |
2019 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist | |
2020 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { | |
2021 Node *phi = PhiNode::make(lp, back_edge_val); | |
2022 phi->set_req(LoopNode::EntryControl, lp_entry_val); | |
2023 // Use existing phi if it already exists | |
2024 Node *hit = _igvn.hash_find_insert(phi); | |
2025 if( hit == NULL ) { | |
2026 _igvn.register_new_node_with_optimizer(phi); | |
2027 set_ctrl(phi, lp); | |
2028 } else { | |
2029 // Remove the new phi from the graph and use the hit | |
2030 _igvn.remove_dead_node(phi); | |
2031 phi = hit; | |
2032 } | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2033 _igvn.replace_input_of(use, idx, phi); |
0 | 2034 } |
2035 | |
2036 #ifdef ASSERT | |
2037 //------------------------------ is_valid_loop_partition ------------------------------------- | |
2038 // Validate the loop partition sets: peel and not_peel | |
2039 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, | |
2040 VectorSet& not_peel ) { | |
2041 uint i; | |
2042 // Check that peel_list entries are in the peel set | |
2043 for (i = 0; i < peel_list.size(); i++) { | |
2044 if (!peel.test(peel_list.at(i)->_idx)) { | |
2045 return false; | |
2046 } | |
2047 } | |
2048 // Check at loop members are in one of peel set or not_peel set | |
2049 for (i = 0; i < loop->_body.size(); i++ ) { | |
2050 Node *def = loop->_body.at(i); | |
2051 uint di = def->_idx; | |
2052 // Check that peel set elements are in peel_list | |
2053 if (peel.test(di)) { | |
2054 if (not_peel.test(di)) { | |
2055 return false; | |
2056 } | |
2057 // Must be in peel_list also | |
2058 bool found = false; | |
2059 for (uint j = 0; j < peel_list.size(); j++) { | |
2060 if (peel_list.at(j)->_idx == di) { | |
2061 found = true; | |
2062 break; | |
2063 } | |
2064 } | |
2065 if (!found) { | |
2066 return false; | |
2067 } | |
2068 } else if (not_peel.test(di)) { | |
2069 if (peel.test(di)) { | |
2070 return false; | |
2071 } | |
2072 } else { | |
2073 return false; | |
2074 } | |
2075 } | |
2076 return true; | |
2077 } | |
2078 | |
2079 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- | |
2080 // Ensure a use outside of loop is of the right form | |
2081 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { | |
2082 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; | |
2083 return (use->is_Phi() && | |
2084 use_c->is_Region() && use_c->req() == 3 && | |
2085 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || | |
2086 use_c->in(exit_idx)->Opcode() == Op_IfFalse || | |
2087 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && | |
2088 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); | |
2089 } | |
2090 | |
2091 //------------------------------ is_valid_clone_loop_form ------------------------------------- | |
2092 // Ensure that all uses outside of loop are of the right form | |
2093 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, | |
2094 uint orig_exit_idx, uint clone_exit_idx) { | |
2095 uint len = peel_list.size(); | |
2096 for (uint i = 0; i < len; i++) { | |
2097 Node *def = peel_list.at(i); | |
2098 | |
2099 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { | |
2100 Node *use = def->fast_out(j); | |
2101 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; | |
2102 if (!loop->is_member(get_loop(use_c))) { | |
2103 // use is not in the loop, check for correct structure | |
2104 if (use->in(0) == def) { | |
2105 // Okay | |
2106 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { | |
2107 return false; | |
2108 } | |
2109 } | |
2110 } | |
2111 } | |
2112 return true; | |
2113 } | |
2114 #endif | |
2115 | |
2116 //------------------------------ partial_peel ------------------------------------- | |
2117 // Partially peel (aka loop rotation) the top portion of a loop (called | |
2118 // the peel section below) by cloning it and placing one copy just before | |
2119 // the new loop head and the other copy at the bottom of the new loop. | |
2120 // | |
2121 // before after where it came from | |
2122 // | |
2123 // stmt1 stmt1 | |
2124 // loop: stmt2 clone | |
2125 // stmt2 if condA goto exitA clone | |
2126 // if condA goto exitA new_loop: new | |
2127 // stmt3 stmt3 clone | |
2128 // if !condB goto loop if condB goto exitB clone | |
2129 // exitB: stmt2 orig | |
2130 // stmt4 if !condA goto new_loop orig | |
2131 // exitA: goto exitA | |
2132 // exitB: | |
2133 // stmt4 | |
2134 // exitA: | |
2135 // | |
2136 // Step 1: find the cut point: an exit test on probable | |
2137 // induction variable. | |
2138 // Step 2: schedule (with cloning) operations in the peel | |
2139 // section that can be executed after the cut into | |
2140 // the section that is not peeled. This may need | |
2141 // to clone operations into exit blocks. For | |
2142 // instance, a reference to A[i] in the not-peel | |
2143 // section and a reference to B[i] in an exit block | |
2144 // may cause a left-shift of i by 2 to be placed | |
2145 // in the peel block. This step will clone the left | |
2146 // shift into the exit block and sink the left shift | |
2147 // from the peel to the not-peel section. | |
2148 // Step 3: clone the loop, retarget the control, and insert | |
2149 // phis for values that are live across the new loop | |
2150 // head. This is very dependent on the graph structure | |
2151 // from clone_loop. It creates region nodes for | |
2152 // exit control and associated phi nodes for values | |
2153 // flow out of the loop through that exit. The region | |
2154 // node is dominated by the clone's control projection. | |
2155 // So the clone's peel section is placed before the | |
2156 // new loop head, and the clone's not-peel section is | |
2157 // forms the top part of the new loop. The original | |
2158 // peel section forms the tail of the new loop. | |
2159 // Step 4: update the dominator tree and recompute the | |
2160 // dominator depth. | |
2161 // | |
2162 // orig | |
2163 // | |
2445 | 2164 // stmt1 |
2165 // | | |
2166 // v | |
2167 // loop predicate | |
2168 // | | |
2169 // v | |
0 | 2170 // loop<----+ |
2171 // | | | |
2172 // stmt2 | | |
2173 // | | | |
2174 // v | | |
2175 // ifA | | |
2176 // / | | | |
2177 // v v | | |
2178 // false true ^ <-- last_peel | |
2179 // / | | | |
2180 // / ===|==cut | | |
2181 // / stmt3 | <-- first_not_peel | |
2182 // / | | | |
2183 // | v | | |
2184 // v ifB | | |
2185 // exitA: / \ | | |
2186 // / \ | | |
2187 // v v | | |
2188 // false true | | |
2189 // / \ | | |
2190 // / ----+ | |
2191 // | | |
2192 // v | |
2193 // exitB: | |
2194 // stmt4 | |
2195 // | |
2196 // | |
2197 // after clone loop | |
2198 // | |
2199 // stmt1 | |
2445 | 2200 // | |
2201 // v | |
2202 // loop predicate | |
0 | 2203 // / \ |
2204 // clone / \ orig | |
2205 // / \ | |
2206 // / \ | |
2207 // v v | |
2208 // +---->loop loop<----+ | |
2209 // | | | | | |
2210 // | stmt2 stmt2 | | |
2211 // | | | | | |
2212 // | v v | | |
2213 // | ifA ifA | | |
2214 // | | \ / | | | |
2215 // | v v v v | | |
2216 // ^ true false false true ^ <-- last_peel | |
2217 // | | ^ \ / | | | |
2218 // | cut==|== \ \ / ===|==cut | | |
2219 // | stmt3 \ \ / stmt3 | <-- first_not_peel | |
2220 // | | dom | | | | | |
2221 // | v \ 1v v2 v | | |
2222 // | ifB regionA ifB | | |
2223 // | / \ | / \ | | |
2224 // | / \ v / \ | | |
2225 // | v v exitA: v v | | |
2226 // | true false false true | | |
2227 // | / ^ \ / \ | | |
2228 // +---- \ \ / ----+ | |
2229 // dom \ / | |
2230 // \ 1v v2 | |
2231 // regionB | |
2232 // | | |
2233 // v | |
2234 // exitB: | |
2235 // stmt4 | |
2236 // | |
2237 // | |
2238 // after partial peel | |
2239 // | |
2240 // stmt1 | |
2445 | 2241 // | |
2242 // v | |
2243 // loop predicate | |
0 | 2244 // / |
2245 // clone / orig | |
2246 // / TOP | |
2247 // / \ | |
2248 // v v | |
2445 | 2249 // TOP->loop loop----+ |
0 | 2250 // | | | |
2251 // stmt2 stmt2 | | |
2252 // | | | | |
2253 // v v | | |
2254 // ifA ifA | | |
2255 // | \ / | | | |
2256 // v v v v | | |
2257 // true false false true | <-- last_peel | |
2258 // | ^ \ / +------|---+ | |
2259 // +->newloop \ \ / === ==cut | | | |
2260 // | stmt3 \ \ / TOP | | | |
2261 // | | dom | | stmt3 | | <-- first_not_peel | |
2262 // | v \ 1v v2 v | | | |
2263 // | ifB regionA ifB ^ v | |
2264 // | / \ | / \ | | | |
2265 // | / \ v / \ | | | |
2266 // | v v exitA: v v | | | |
2267 // | true false false true | | | |
2268 // | / ^ \ / \ | | | |
2269 // | | \ \ / v | | | |
2270 // | | dom \ / TOP | | | |
2271 // | | \ 1v v2 | | | |
2272 // ^ v regionB | | | |
2273 // | | | | | | |
2274 // | | v ^ v | |
2275 // | | exitB: | | | |
2276 // | | stmt4 | | | |
2277 // | +------------>-----------------+ | | |
2278 // | | | |
2279 // +-----------------<---------------------+ | |
2280 // | |
2281 // | |
2282 // final graph | |
2283 // | |
2284 // stmt1 | |
2285 // | | |
2286 // v | |
3281
273b56978029
7039586: test/java/util/Collections/Rotate.java failing with hs21-b09
kvn
parents:
2468
diff
changeset
|
2287 // loop predicate |
273b56978029
7039586: test/java/util/Collections/Rotate.java failing with hs21-b09
kvn
parents:
2468
diff
changeset
|
2288 // | |
273b56978029
7039586: test/java/util/Collections/Rotate.java failing with hs21-b09
kvn
parents:
2468
diff
changeset
|
2289 // v |
2445 | 2290 // stmt2 clone |
2291 // | | |
2292 // v | |
0 | 2293 // ........> ifA clone |
2294 // : / | | |
2295 // dom / | | |
2296 // : v v | |
2297 // : false true | |
2298 // : | | | |
2445 | 2299 // : | v |
0 | 2300 // : | newloop<-----+ |
2301 // : | | | | |
2302 // : | stmt3 clone | | |
2303 // : | | | | |
2304 // : | v | | |
2305 // : | ifB | | |
2306 // : | / \ | | |
2307 // : | v v | | |
2308 // : | false true | | |
2309 // : | | | | | |
2310 // : | v stmt2 | | |
2311 // : | exitB: | | | |
2312 // : | stmt4 v | | |
2313 // : | ifA orig | | |
2314 // : | / \ | | |
2315 // : | / \ | | |
2316 // : | v v | | |
2317 // : | false true | | |
2318 // : | / \ | | |
2319 // : v v -----+ | |
2320 // RegionA | |
2321 // | | |
2322 // v | |
2323 // exitA | |
2324 // | |
2325 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { | |
2326 | |
2445 | 2327 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); |
108 | 2328 if (!loop->_head->is_Loop()) { |
2329 return false; } | |
2330 | |
0 | 2331 LoopNode *head = loop->_head->as_Loop(); |
2332 | |
2333 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { | |
2334 return false; | |
2335 } | |
2336 | |
2337 // Check for complex exit control | |
2338 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { | |
2339 Node *n = loop->_body.at(ii); | |
2340 int opc = n->Opcode(); | |
2341 if (n->is_Call() || | |
2342 opc == Op_Catch || | |
2343 opc == Op_CatchProj || | |
2344 opc == Op_Jump || | |
2345 opc == Op_JumpProj) { | |
2346 #if !defined(PRODUCT) | |
2347 if (TracePartialPeeling) { | |
2348 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); | |
2349 } | |
2350 #endif | |
2351 return false; | |
2352 } | |
2353 } | |
2354 | |
2355 int dd = dom_depth(head); | |
2356 | |
2357 // Step 1: find cut point | |
2358 | |
2359 // Walk up dominators to loop head looking for first loop exit | |
2360 // which is executed on every path thru loop. | |
2361 IfNode *peel_if = NULL; | |
2362 IfNode *peel_if_cmpu = NULL; | |
2363 | |
2364 Node *iff = loop->tail(); | |
2365 while( iff != head ) { | |
2366 if( iff->is_If() ) { | |
2367 Node *ctrl = get_ctrl(iff->in(1)); | |
2368 if (ctrl->is_top()) return false; // Dead test on live IF. | |
2369 // If loop-varying exit-test, check for induction variable | |
2370 if( loop->is_member(get_loop(ctrl)) && | |
2371 loop->is_loop_exit(iff) && | |
2372 is_possible_iv_test(iff)) { | |
2373 Node* cmp = iff->in(1)->in(1); | |
2374 if (cmp->Opcode() == Op_CmpI) { | |
2375 peel_if = iff->as_If(); | |
2376 } else { | |
2377 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); | |
2378 peel_if_cmpu = iff->as_If(); | |
2379 } | |
2380 } | |
2381 } | |
2382 iff = idom(iff); | |
2383 } | |
2384 // Prefer signed compare over unsigned compare. | |
2385 IfNode* new_peel_if = NULL; | |
2386 if (peel_if == NULL) { | |
2387 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { | |
2388 return false; // No peel point found | |
2389 } | |
2390 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); | |
2391 if (new_peel_if == NULL) { | |
2392 return false; // No peel point found | |
2393 } | |
2394 peel_if = new_peel_if; | |
2395 } | |
2396 Node* last_peel = stay_in_loop(peel_if, loop); | |
2397 Node* first_not_peeled = stay_in_loop(last_peel, loop); | |
2398 if (first_not_peeled == NULL || first_not_peeled == head) { | |
2399 return false; | |
2400 } | |
2401 | |
2402 #if !defined(PRODUCT) | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2403 if (TraceLoopOpts) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2404 tty->print("PartialPeel "); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2405 loop->dump_head(); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2406 } |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2407 |
0 | 2408 if (TracePartialPeeling) { |
2409 tty->print_cr("before partial peel one iteration"); | |
2410 Node_List wl; | |
2411 Node* t = head->in(2); | |
2412 while (true) { | |
2413 wl.push(t); | |
2414 if (t == head) break; | |
2415 t = idom(t); | |
2416 } | |
2417 while (wl.size() > 0) { | |
2418 Node* tt = wl.pop(); | |
2419 tt->dump(); | |
2420 if (tt == last_peel) tty->print_cr("-- cut --"); | |
2421 } | |
2422 } | |
2423 #endif | |
2424 ResourceArea *area = Thread::current()->resource_area(); | |
2425 VectorSet peel(area); | |
2426 VectorSet not_peel(area); | |
2427 Node_List peel_list(area); | |
2428 Node_List worklist(area); | |
2429 Node_List sink_list(area); | |
2430 | |
2431 // Set of cfg nodes to peel are those that are executable from | |
2432 // the head through last_peel. | |
2433 assert(worklist.size() == 0, "should be empty"); | |
2434 worklist.push(head); | |
2435 peel.set(head->_idx); | |
2436 while (worklist.size() > 0) { | |
2437 Node *n = worklist.pop(); | |
2438 if (n != last_peel) { | |
2439 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
2440 Node* use = n->fast_out(j); | |
2441 if (use->is_CFG() && | |
2442 loop->is_member(get_loop(use)) && | |
2443 !peel.test_set(use->_idx)) { | |
2444 worklist.push(use); | |
2445 } | |
2446 } | |
2447 } | |
2448 } | |
2449 | |
2450 // Set of non-cfg nodes to peel are those that are control | |
2451 // dependent on the cfg nodes. | |
2452 uint i; | |
2453 for(i = 0; i < loop->_body.size(); i++ ) { | |
2454 Node *n = loop->_body.at(i); | |
2455 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; | |
2456 if (peel.test(n_c->_idx)) { | |
2457 peel.set(n->_idx); | |
2458 } else { | |
2459 not_peel.set(n->_idx); | |
2460 } | |
2461 } | |
2462 | |
2463 // Step 2: move operations from the peeled section down into the | |
2464 // not-peeled section | |
2465 | |
2466 // Get a post order schedule of nodes in the peel region | |
2467 // Result in right-most operand. | |
2468 scheduled_nodelist(loop, peel, peel_list ); | |
2469 | |
2470 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); | |
2471 | |
2472 // For future check for too many new phis | |
2473 uint old_phi_cnt = 0; | |
2474 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { | |
2475 Node* use = head->fast_out(j); | |
2476 if (use->is_Phi()) old_phi_cnt++; | |
2477 } | |
2478 | |
2479 #if !defined(PRODUCT) | |
2480 if (TracePartialPeeling) { | |
2481 tty->print_cr("\npeeled list"); | |
2482 } | |
2483 #endif | |
2484 | |
2485 // Evacuate nodes in peel region into the not_peeled region if possible | |
2486 uint new_phi_cnt = 0; | |
2487 for (i = 0; i < peel_list.size();) { | |
2488 Node* n = peel_list.at(i); | |
2489 #if !defined(PRODUCT) | |
2490 if (TracePartialPeeling) n->dump(); | |
2491 #endif | |
2492 bool incr = true; | |
2493 if ( !n->is_CFG() ) { | |
2494 | |
2495 if ( has_use_in_set(n, not_peel) ) { | |
2496 | |
2497 // If not used internal to the peeled region, | |
2498 // move "n" from peeled to not_peeled region. | |
2499 | |
2500 if ( !has_use_internal_to_set(n, peel, loop) ) { | |
2501 | |
2502 // if not pinned and not a load (which maybe anti-dependent on a store) | |
2503 // and not a CMove (Matcher expects only bool->cmove). | |
2504 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { | |
2505 clone_for_use_outside_loop( loop, n, worklist ); | |
2506 | |
2507 sink_list.push(n); | |
2508 peel >>= n->_idx; // delete n from peel set. | |
2509 not_peel <<= n->_idx; // add n to not_peel set. | |
2510 peel_list.remove(i); | |
2511 incr = false; | |
2512 #if !defined(PRODUCT) | |
2513 if (TracePartialPeeling) { | |
2514 tty->print_cr("sink to not_peeled region: %d newbb: %d", | |
2515 n->_idx, get_ctrl(n)->_idx); | |
2516 } | |
2517 #endif | |
2518 } | |
2519 } else { | |
2520 // Otherwise check for special def-use cases that span | |
2521 // the peel/not_peel boundary such as bool->if | |
2522 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); | |
2523 new_phi_cnt++; | |
2524 } | |
2525 } | |
2526 } | |
2527 if (incr) i++; | |
2528 } | |
2529 | |
2530 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { | |
2531 #if !defined(PRODUCT) | |
2532 if (TracePartialPeeling) { | |
2533 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", | |
2534 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); | |
2535 } | |
2536 #endif | |
2537 if (new_peel_if != NULL) { | |
2538 remove_cmpi_loop_exit(new_peel_if, loop); | |
2539 } | |
2540 // Inhibit more partial peeling on this loop | |
2541 assert(!head->is_partial_peel_loop(), "not partial peeled"); | |
2542 head->mark_partial_peel_failed(); | |
2543 return false; | |
2544 } | |
2545 | |
2546 // Step 3: clone loop, retarget control, and insert new phis | |
2547 | |
2548 // Create new loop head for new phis and to hang | |
2549 // the nodes being moved (sinked) from the peel region. | |
2550 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2551 new_head->set_unswitch_count(head->unswitch_count()); // Preserve |
0 | 2552 _igvn.register_new_node_with_optimizer(new_head); |
2553 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); | |
2554 first_not_peeled->set_req(0, new_head); | |
2555 set_loop(new_head, loop); | |
2556 loop->_body.push(new_head); | |
2557 not_peel.set(new_head->_idx); | |
2558 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); | |
2559 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); | |
2560 | |
2561 while (sink_list.size() > 0) { | |
2562 Node* n = sink_list.pop(); | |
2563 set_ctrl(n, new_head); | |
2564 } | |
2565 | |
2566 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); | |
2567 | |
2568 clone_loop( loop, old_new, dd ); | |
2569 | |
2570 const uint clone_exit_idx = 1; | |
2571 const uint orig_exit_idx = 2; | |
2572 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); | |
2573 | |
2574 Node* head_clone = old_new[head->_idx]; | |
2575 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); | |
2576 Node* orig_tail_clone = head_clone->in(2); | |
2577 | |
2578 // Add phi if "def" node is in peel set and "use" is not | |
2579 | |
2580 for(i = 0; i < peel_list.size(); i++ ) { | |
2581 Node *def = peel_list.at(i); | |
2582 if (!def->is_CFG()) { | |
2583 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { | |
2584 Node *use = def->fast_out(j); | |
2585 if (has_node(use) && use->in(0) != C->top() && | |
2586 (!peel.test(use->_idx) || | |
2587 (use->is_Phi() && use->in(0) == head)) ) { | |
2588 worklist.push(use); | |
2589 } | |
2590 } | |
2591 while( worklist.size() ) { | |
2592 Node *use = worklist.pop(); | |
2593 for (uint j = 1; j < use->req(); j++) { | |
2594 Node* n = use->in(j); | |
2595 if (n == def) { | |
2596 | |
2597 // "def" is in peel set, "use" is not in peel set | |
2598 // or "use" is in the entry boundary (a phi) of the peel set | |
2599 | |
2600 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; | |
2601 | |
2602 if ( loop->is_member(get_loop( use_c )) ) { | |
2603 // use is in loop | |
2604 if (old_new[use->_idx] != NULL) { // null for dead code | |
2605 Node* use_clone = old_new[use->_idx]; | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2606 _igvn.replace_input_of(use, j, C->top()); |
0 | 2607 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); |
2608 } | |
2609 } else { | |
2610 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); | |
2611 // use is not in the loop, check if the live range includes the cut | |
2612 Node* lp_if = use_c->in(orig_exit_idx)->in(0); | |
2613 if (not_peel.test(lp_if->_idx)) { | |
2614 assert(j == orig_exit_idx, "use from original loop"); | |
2615 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); | |
2616 } | |
2617 } | |
2618 } | |
2619 } | |
2620 } | |
2621 } | |
2622 } | |
2623 | |
2624 // Step 3b: retarget control | |
2625 | |
2626 // Redirect control to the new loop head if a cloned node in | |
2627 // the not_peeled region has control that points into the peeled region. | |
2628 // This necessary because the cloned peeled region will be outside | |
2629 // the loop. | |
2630 // from to | |
2631 // cloned-peeled <---+ | |
2632 // new_head_clone: | <--+ | |
2633 // cloned-not_peeled in(0) in(0) | |
2634 // orig-peeled | |
2635 | |
2636 for(i = 0; i < loop->_body.size(); i++ ) { | |
2637 Node *n = loop->_body.at(i); | |
2638 if (!n->is_CFG() && n->in(0) != NULL && | |
2639 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { | |
2640 Node* n_clone = old_new[n->_idx]; | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2641 _igvn.replace_input_of(n_clone, 0, new_head_clone); |
0 | 2642 } |
2643 } | |
2644 | |
2645 // Backedge of the surviving new_head (the clone) is original last_peel | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2646 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); |
0 | 2647 |
2648 // Cut first node in original not_peel set | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2649 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: |
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2650 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of |
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2651 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls |
0 | 2652 |
2653 // Copy head_clone back-branch info to original head | |
2654 // and remove original head's loop entry and | |
2655 // clone head's back-branch | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2656 _igvn.rehash_node_delayed(head); // Multiple edge updates |
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2657 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); |
0 | 2658 head->set_req(LoopNode::LoopBackControl, C->top()); |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2659 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); |
0 | 2660 |
2661 // Similarly modify the phis | |
2662 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { | |
2663 Node* use = head->fast_out(k); | |
2664 if (use->is_Phi() && use->outcnt() > 0) { | |
2665 Node* use_clone = old_new[use->_idx]; | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2666 _igvn.rehash_node_delayed(use); // Multiple edge updates |
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2667 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); |
0 | 2668 use->set_req(LoopNode::LoopBackControl, C->top()); |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2669 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); |
0 | 2670 } |
2671 } | |
2672 | |
2673 // Step 4: update dominator tree and dominator depth | |
2674 | |
2675 set_idom(head, orig_tail_clone, dd); | |
2676 recompute_dom_depth(); | |
2677 | |
2678 // Inhibit more partial peeling on this loop | |
2679 new_head_clone->set_partial_peel_loop(); | |
2680 C->set_major_progress(); | |
2681 | |
2682 #if !defined(PRODUCT) | |
2683 if (TracePartialPeeling) { | |
2684 tty->print_cr("\nafter partial peel one iteration"); | |
2685 Node_List wl(area); | |
2686 Node* t = last_peel; | |
2687 while (true) { | |
2688 wl.push(t); | |
2689 if (t == head_clone) break; | |
2690 t = idom(t); | |
2691 } | |
2692 while (wl.size() > 0) { | |
2693 Node* tt = wl.pop(); | |
2694 if (tt == head) tty->print_cr("orig head"); | |
2695 else if (tt == new_head_clone) tty->print_cr("new head"); | |
2696 else if (tt == head_clone) tty->print_cr("clone head"); | |
2697 tt->dump(); | |
2698 } | |
2699 } | |
2700 #endif | |
2701 return true; | |
2702 } | |
2703 | |
2704 //------------------------------reorg_offsets---------------------------------- | |
2705 // Reorganize offset computations to lower register pressure. Mostly | |
2706 // prevent loop-fallout uses of the pre-incremented trip counter (which are | |
2707 // then alive with the post-incremented trip counter forcing an extra | |
2708 // register move) | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2709 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2710 // Perform it only for canonical counted loops. |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2711 // Loop's shape could be messed up by iteration_split_impl. |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2712 if (!loop->_head->is_CountedLoop()) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2713 return; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2714 if (!loop->_head->as_Loop()->is_valid_counted_loop()) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2715 return; |
0 | 2716 |
2717 CountedLoopNode *cl = loop->_head->as_CountedLoop(); | |
2718 CountedLoopEndNode *cle = cl->loopexit(); | |
2719 Node *exit = cle->proj_out(false); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2720 Node *phi = cl->phi(); |
0 | 2721 |
2722 // Check for the special case of folks using the pre-incremented | |
2723 // trip-counter on the fall-out path (forces the pre-incremented | |
2724 // and post-incremented trip counter to be live at the same time). | |
2725 // Fix this by adjusting to use the post-increment trip counter. | |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
318
diff
changeset
|
2726 |
0 | 2727 bool progress = true; |
2728 while (progress) { | |
2729 progress = false; | |
2730 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { | |
2731 Node* use = phi->fast_out(i); // User of trip-counter | |
2732 if (!has_ctrl(use)) continue; | |
2733 Node *u_ctrl = get_ctrl(use); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2734 if (use->is_Phi()) { |
0 | 2735 u_ctrl = NULL; |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2736 for (uint j = 1; j < use->req(); j++) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2737 if (use->in(j) == phi) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2738 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); |
0 | 2739 } |
2740 IdealLoopTree *u_loop = get_loop(u_ctrl); | |
2741 // Look for loop-invariant use | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2742 if (u_loop == loop) continue; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2743 if (loop->is_member(u_loop)) continue; |
0 | 2744 // Check that use is live out the bottom. Assuming the trip-counter |
2745 // update is right at the bottom, uses of of the loop middle are ok. | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2746 if (dom_lca(exit, u_ctrl) != exit) continue; |
0 | 2747 // Hit! Refactor use to use the post-incremented tripcounter. |
2748 // Compute a post-increment tripcounter. | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
164
diff
changeset
|
2749 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() ); |
0 | 2750 register_new_node( opaq, u_ctrl ); |
2751 Node *neg_stride = _igvn.intcon(-cle->stride_con()); | |
2752 set_ctrl(neg_stride, C->root()); | |
2753 Node *post = new (C, 3) AddINode( opaq, neg_stride); | |
2754 register_new_node( post, u_ctrl ); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
4779
diff
changeset
|
2755 _igvn.rehash_node_delayed(use); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2756 for (uint j = 1; j < use->req(); j++) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2757 if (use->in(j) == phi) |
0 | 2758 use->set_req(j, post); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2759 } |
0 | 2760 // Since DU info changed, rerun loop |
2761 progress = true; | |
2762 break; | |
2763 } | |
2764 } | |
2765 | |
2766 } |