Mercurial > hg > graal-jvmci-8
comparison src/share/vm/opto/loopPredicate.cpp @ 2445:08eb13460b3a
7004535: Clone loop predicate during loop unswitch
Summary: Clone loop predicate for clonned loops
Reviewed-by: never
author | kvn |
---|---|
date | Sat, 02 Apr 2011 10:54:15 -0700 |
parents | |
children | 2e038ad0c1d0 |
comparison
equal
deleted
inserted
replaced
2444:07acc51c1d2a | 2445:08eb13460b3a |
---|---|
1 /* | |
2 * Copyright (c) 2011, Oracle and/or its affiliates. All rights reserved. | |
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA | |
20 * or visit www.oracle.com if you need additional information or have any | |
21 * questions. | |
22 * | |
23 */ | |
24 | |
25 #include "precompiled.hpp" | |
26 #include "opto/loopnode.hpp" | |
27 #include "opto/addnode.hpp" | |
28 #include "opto/callnode.hpp" | |
29 #include "opto/connode.hpp" | |
30 #include "opto/loopnode.hpp" | |
31 #include "opto/mulnode.hpp" | |
32 #include "opto/rootnode.hpp" | |
33 #include "opto/subnode.hpp" | |
34 | |
35 /* | |
36 * The general idea of Loop Predication is to insert a predicate on the entry | |
37 * path to a loop, and raise a uncommon trap if the check of the condition fails. | |
38 * The condition checks are promoted from inside the loop body, and thus | |
39 * the checks inside the loop could be eliminated. Currently, loop predication | |
40 * optimization has been applied to remove array range check and loop invariant | |
41 * checks (such as null checks). | |
42 */ | |
43 | |
44 //-------------------------------is_uncommon_trap_proj---------------------------- | |
45 // Return true if proj is the form of "proj->[region->..]call_uct" | |
46 bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason) { | |
47 int path_limit = 10; | |
48 assert(proj, "invalid argument"); | |
49 Node* out = proj; | |
50 for (int ct = 0; ct < path_limit; ct++) { | |
51 out = out->unique_ctrl_out(); | |
52 if (out == NULL) | |
53 return false; | |
54 if (out->is_CallStaticJava()) { | |
55 int req = out->as_CallStaticJava()->uncommon_trap_request(); | |
56 if (req != 0) { | |
57 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req); | |
58 if (trap_reason == reason || reason == Deoptimization::Reason_none) { | |
59 return true; | |
60 } | |
61 } | |
62 return false; // don't do further after call | |
63 } | |
64 if (out->Opcode() != Op_Region) | |
65 return false; | |
66 } | |
67 return false; | |
68 } | |
69 | |
70 //-------------------------------is_uncommon_trap_if_pattern------------------------- | |
71 // Return true for "if(test)-> proj -> ... | |
72 // | | |
73 // V | |
74 // other_proj->[region->..]call_uct" | |
75 // | |
76 // "must_reason_predicate" means the uct reason must be Reason_predicate | |
77 bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, Deoptimization::DeoptReason reason) { | |
78 Node *in0 = proj->in(0); | |
79 if (!in0->is_If()) return false; | |
80 // Variation of a dead If node. | |
81 if (in0->outcnt() < 2) return false; | |
82 IfNode* iff = in0->as_If(); | |
83 | |
84 // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate | |
85 if (reason != Deoptimization::Reason_none) { | |
86 if (iff->in(1)->Opcode() != Op_Conv2B || | |
87 iff->in(1)->in(1)->Opcode() != Op_Opaque1) { | |
88 return false; | |
89 } | |
90 } | |
91 | |
92 ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj(); | |
93 if (is_uncommon_trap_proj(other_proj, reason)) { | |
94 assert(reason == Deoptimization::Reason_none || | |
95 Compile::current()->is_predicate_opaq(iff->in(1)->in(1)), "should be on the list"); | |
96 return true; | |
97 } | |
98 return false; | |
99 } | |
100 | |
101 //-------------------------------register_control------------------------- | |
102 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) { | |
103 assert(n->is_CFG(), "must be control node"); | |
104 _igvn.register_new_node_with_optimizer(n); | |
105 loop->_body.push(n); | |
106 set_loop(n, loop); | |
107 // When called from beautify_loops() idom is not constructed yet. | |
108 if (_idom != NULL) { | |
109 set_idom(n, pred, dom_depth(pred)); | |
110 } | |
111 } | |
112 | |
113 //------------------------------create_new_if_for_predicate------------------------ | |
114 // create a new if above the uct_if_pattern for the predicate to be promoted. | |
115 // | |
116 // before after | |
117 // ---------- ---------- | |
118 // ctrl ctrl | |
119 // | | | |
120 // | | | |
121 // v v | |
122 // iff new_iff | |
123 // / \ / \ | |
124 // / \ / \ | |
125 // v v v v | |
126 // uncommon_proj cont_proj if_uct if_cont | |
127 // \ | | | | | |
128 // \ | | | | | |
129 // v v v | v | |
130 // rgn loop | iff | |
131 // | | / \ | |
132 // | | / \ | |
133 // v | v v | |
134 // uncommon_trap | uncommon_proj cont_proj | |
135 // \ \ | | | |
136 // \ \ | | | |
137 // v v v v | |
138 // rgn loop | |
139 // | | |
140 // | | |
141 // v | |
142 // uncommon_trap | |
143 // | |
144 // | |
145 // We will create a region to guard the uct call if there is no one there. | |
146 // The true projecttion (if_cont) of the new_iff is returned. | |
147 // This code is also used to clone predicates to clonned loops. | |
148 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, | |
149 Deoptimization::DeoptReason reason) { | |
150 assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!"); | |
151 IfNode* iff = cont_proj->in(0)->as_If(); | |
152 | |
153 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); | |
154 Node *rgn = uncommon_proj->unique_ctrl_out(); | |
155 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); | |
156 | |
157 uint proj_index = 1; // region's edge corresponding to uncommon_proj | |
158 if (!rgn->is_Region()) { // create a region to guard the call | |
159 assert(rgn->is_Call(), "must be call uct"); | |
160 CallNode* call = rgn->as_Call(); | |
161 IdealLoopTree* loop = get_loop(call); | |
162 rgn = new (C, 1) RegionNode(1); | |
163 rgn->add_req(uncommon_proj); | |
164 register_control(rgn, loop, uncommon_proj); | |
165 _igvn.hash_delete(call); | |
166 call->set_req(0, rgn); | |
167 // When called from beautify_loops() idom is not constructed yet. | |
168 if (_idom != NULL) { | |
169 set_idom(call, rgn, dom_depth(rgn)); | |
170 } | |
171 } else { | |
172 // Find region's edge corresponding to uncommon_proj | |
173 for (; proj_index < rgn->req(); proj_index++) | |
174 if (rgn->in(proj_index) == uncommon_proj) break; | |
175 assert(proj_index < rgn->req(), "sanity"); | |
176 } | |
177 | |
178 Node* entry = iff->in(0); | |
179 if (new_entry != NULL) { | |
180 // Clonning the predicate to new location. | |
181 entry = new_entry; | |
182 } | |
183 // Create new_iff | |
184 IdealLoopTree* lp = get_loop(entry); | |
185 IfNode *new_iff = iff->clone()->as_If(); | |
186 new_iff->set_req(0, entry); | |
187 register_control(new_iff, lp, entry); | |
188 Node *if_cont = new (C, 1) IfTrueNode(new_iff); | |
189 Node *if_uct = new (C, 1) IfFalseNode(new_iff); | |
190 if (cont_proj->is_IfFalse()) { | |
191 // Swap | |
192 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; | |
193 } | |
194 register_control(if_cont, lp, new_iff); | |
195 register_control(if_uct, get_loop(rgn), new_iff); | |
196 | |
197 // if_uct to rgn | |
198 _igvn.hash_delete(rgn); | |
199 rgn->add_req(if_uct); | |
200 // When called from beautify_loops() idom is not constructed yet. | |
201 if (_idom != NULL) { | |
202 Node* ridom = idom(rgn); | |
203 Node* nrdom = dom_lca(ridom, new_iff); | |
204 set_idom(rgn, nrdom, dom_depth(rgn)); | |
205 } | |
206 | |
207 // If rgn has phis add new edges which has the same | |
208 // value as on original uncommon_proj pass. | |
209 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last"); | |
210 bool has_phi = false; | |
211 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) { | |
212 Node* use = rgn->fast_out(i); | |
213 if (use->is_Phi() && use->outcnt() > 0) { | |
214 assert(use->in(0) == rgn, ""); | |
215 _igvn.hash_delete(use); | |
216 use->add_req(use->in(proj_index)); | |
217 _igvn._worklist.push(use); | |
218 has_phi = true; | |
219 } | |
220 } | |
221 assert(!has_phi || rgn->req() > 3, "no phis when region is created"); | |
222 | |
223 if (new_entry == NULL) { | |
224 // Attach if_cont to iff | |
225 _igvn.hash_delete(iff); | |
226 iff->set_req(0, if_cont); | |
227 if (_idom != NULL) { | |
228 set_idom(iff, if_cont, dom_depth(iff)); | |
229 } | |
230 } | |
231 return if_cont->as_Proj(); | |
232 } | |
233 | |
234 //------------------------------create_new_if_for_predicate------------------------ | |
235 // Create a new if below new_entry for the predicate to be cloned (IGVN optimization) | |
236 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, | |
237 Deoptimization::DeoptReason reason) { | |
238 assert(new_entry != 0, "only used for clone predicate"); | |
239 assert(PhaseIdealLoop::is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!"); | |
240 IfNode* iff = cont_proj->in(0)->as_If(); | |
241 | |
242 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); | |
243 Node *rgn = uncommon_proj->unique_ctrl_out(); | |
244 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); | |
245 | |
246 uint proj_index = 1; // region's edge corresponding to uncommon_proj | |
247 if (!rgn->is_Region()) { // create a region to guard the call | |
248 assert(rgn->is_Call(), "must be call uct"); | |
249 CallNode* call = rgn->as_Call(); | |
250 rgn = new (C, 1) RegionNode(1); | |
251 register_new_node_with_optimizer(rgn); | |
252 rgn->add_req(uncommon_proj); | |
253 hash_delete(call); | |
254 call->set_req(0, rgn); | |
255 } else { | |
256 // Find region's edge corresponding to uncommon_proj | |
257 for (; proj_index < rgn->req(); proj_index++) | |
258 if (rgn->in(proj_index) == uncommon_proj) break; | |
259 assert(proj_index < rgn->req(), "sanity"); | |
260 } | |
261 | |
262 // Create new_iff in new location. | |
263 IfNode *new_iff = iff->clone()->as_If(); | |
264 new_iff->set_req(0, new_entry); | |
265 | |
266 register_new_node_with_optimizer(new_iff); | |
267 Node *if_cont = new (C, 1) IfTrueNode(new_iff); | |
268 Node *if_uct = new (C, 1) IfFalseNode(new_iff); | |
269 if (cont_proj->is_IfFalse()) { | |
270 // Swap | |
271 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; | |
272 } | |
273 register_new_node_with_optimizer(if_cont); | |
274 register_new_node_with_optimizer(if_uct); | |
275 | |
276 // if_uct to rgn | |
277 hash_delete(rgn); | |
278 rgn->add_req(if_uct); | |
279 | |
280 // If rgn has phis add corresponding new edges which has the same | |
281 // value as on original uncommon_proj pass. | |
282 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last"); | |
283 bool has_phi = false; | |
284 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) { | |
285 Node* use = rgn->fast_out(i); | |
286 if (use->is_Phi() && use->outcnt() > 0) { | |
287 hash_delete(use); | |
288 use->add_req(use->in(proj_index)); | |
289 _worklist.push(use); | |
290 has_phi = true; | |
291 } | |
292 } | |
293 assert(!has_phi || rgn->req() > 3, "no phis when region is created"); | |
294 | |
295 return if_cont->as_Proj(); | |
296 } | |
297 | |
298 //--------------------------clone_predicate----------------------- | |
299 ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry, | |
300 Deoptimization::DeoptReason reason, | |
301 PhaseIdealLoop* loop_phase, | |
302 PhaseIterGVN* igvn) { | |
303 ProjNode* new_predicate_proj; | |
304 if (loop_phase != NULL) { | |
305 new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason); | |
306 } else { | |
307 new_predicate_proj = igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason); | |
308 } | |
309 IfNode* iff = new_predicate_proj->in(0)->as_If(); | |
310 Node* ctrl = iff->in(0); | |
311 | |
312 // Match original condition since predicate's projections could be swapped. | |
313 assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); | |
314 Node* opq = new (igvn->C, 2) Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1)); | |
315 igvn->C->add_predicate_opaq(opq); | |
316 | |
317 Node* bol = new (igvn->C, 2) Conv2BNode(opq); | |
318 if (loop_phase != NULL) { | |
319 loop_phase->register_new_node(opq, ctrl); | |
320 loop_phase->register_new_node(bol, ctrl); | |
321 } else { | |
322 igvn->register_new_node_with_optimizer(opq); | |
323 igvn->register_new_node_with_optimizer(bol); | |
324 } | |
325 igvn->hash_delete(iff); | |
326 iff->set_req(1, bol); | |
327 return new_predicate_proj; | |
328 } | |
329 | |
330 //--------------------------move_predicate----------------------- | |
331 // Cut predicate from old place and move it to new. | |
332 ProjNode* PhaseIdealLoop::move_predicate(ProjNode* predicate_proj, Node* new_entry, | |
333 Deoptimization::DeoptReason reason, | |
334 PhaseIdealLoop* loop_phase, | |
335 PhaseIterGVN* igvn) { | |
336 assert(new_entry != NULL, "must be"); | |
337 assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); | |
338 IfNode* iff = predicate_proj->in(0)->as_If(); | |
339 Node* old_entry = iff->in(0); | |
340 | |
341 // Cut predicate from old place. | |
342 Node* old = predicate_proj; | |
343 igvn->_worklist.push(old); | |
344 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { | |
345 Node* use = old->last_out(i); // for each use... | |
346 igvn->hash_delete(use); | |
347 igvn->_worklist.push(use); | |
348 // Update use-def info | |
349 uint uses_found = 0; | |
350 for (uint j = 0; j < use->req(); j++) { | |
351 if (use->in(j) == old) { | |
352 use->set_req(j, old_entry); | |
353 uses_found++; | |
354 if (loop_phase != NULL) { | |
355 if (use->is_CFG()) { | |
356 // When called from beautify_loops() idom is not constructed yet. | |
357 if (loop_phase->_idom != NULL) | |
358 loop_phase->set_idom(use, old_entry, loop_phase->dom_depth(use)); | |
359 } else { | |
360 loop_phase->set_ctrl(use, old_entry); | |
361 } | |
362 } | |
363 } | |
364 } | |
365 i -= uses_found; // we deleted 1 or more copies of this edge | |
366 } | |
367 | |
368 // Move predicate. | |
369 igvn->hash_delete(iff); | |
370 iff->set_req(0, new_entry); | |
371 igvn->_worklist.push(iff); | |
372 | |
373 if (loop_phase != NULL) { | |
374 // Fix up idom and ctrl. | |
375 loop_phase->set_ctrl(iff->in(1), new_entry); | |
376 loop_phase->set_ctrl(iff->in(1)->in(1), new_entry); | |
377 // When called from beautify_loops() idom is not constructed yet. | |
378 if (loop_phase->_idom != NULL) | |
379 loop_phase->set_idom(iff, new_entry, loop_phase->dom_depth(iff)); | |
380 } | |
381 | |
382 return predicate_proj; | |
383 } | |
384 | |
385 //--------------------------clone_loop_predicates----------------------- | |
386 // Interface from IGVN | |
387 Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry) { | |
388 return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, false, NULL, this); | |
389 } | |
390 Node* PhaseIterGVN::move_loop_predicates(Node* old_entry, Node* new_entry) { | |
391 return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, true, NULL, this); | |
392 } | |
393 | |
394 // Interface from PhaseIdealLoop | |
395 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry) { | |
396 return clone_loop_predicates(old_entry, new_entry, false, this, &this->_igvn); | |
397 } | |
398 Node* PhaseIdealLoop::move_loop_predicates(Node* old_entry, Node* new_entry) { | |
399 return clone_loop_predicates(old_entry, new_entry, true, this, &this->_igvn); | |
400 } | |
401 | |
402 // Clone loop predicates to cloned loops (peeled, unswitched, split_if). | |
403 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, | |
404 bool move_predicates, | |
405 PhaseIdealLoop* loop_phase, | |
406 PhaseIterGVN* igvn) { | |
407 #ifdef ASSERT | |
408 if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) { | |
409 if (new_entry != NULL) | |
410 new_entry->dump(); | |
411 assert(false, "not IfTrue, IfFalse, Region or SafePoint"); | |
412 } | |
413 #endif | |
414 // Search original predicates | |
415 Node* entry = old_entry; | |
416 if (UseLoopPredicate) { | |
417 ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); | |
418 if (predicate_proj != NULL) { // right pattern that can be used by loop predication | |
419 assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); | |
420 if (move_predicates) { | |
421 new_entry = move_predicate(predicate_proj, new_entry, | |
422 Deoptimization::Reason_predicate, | |
423 loop_phase, igvn); | |
424 assert(new_entry == predicate_proj, "old predicate fall through projection"); | |
425 } else { | |
426 // clone predicate | |
427 new_entry = clone_predicate(predicate_proj, new_entry, | |
428 Deoptimization::Reason_predicate, | |
429 loop_phase, igvn); | |
430 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate"); | |
431 } | |
432 if (TraceLoopPredicate) { | |
433 tty->print_cr("Loop Predicate %s: ", move_predicates ? "moved" : "cloned"); | |
434 debug_only( new_entry->in(0)->dump(); ) | |
435 } | |
436 } | |
437 } | |
438 return new_entry; | |
439 } | |
440 | |
441 //--------------------------eliminate_loop_predicates----------------------- | |
442 void PhaseIdealLoop::eliminate_loop_predicates(Node* entry) { | |
443 if (UseLoopPredicate) { | |
444 ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); | |
445 if (predicate_proj != NULL) { // right pattern that can be used by loop predication | |
446 Node* n = entry->in(0)->in(1)->in(1); | |
447 assert(n->Opcode()==Op_Opaque1, "must be"); | |
448 // Remove Opaque1 node from predicates list. | |
449 // IGVN will remove this predicate check. | |
450 _igvn.replace_node(n, n->in(1)); | |
451 } | |
452 } | |
453 } | |
454 | |
455 //--------------------------skip_loop_predicates------------------------------ | |
456 // Skip related predicates. | |
457 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) { | |
458 Node* predicate = NULL; | |
459 if (UseLoopPredicate) { | |
460 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); | |
461 if (predicate != NULL) { // right pattern that can be used by loop predication | |
462 assert(entry->is_Proj() && entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); | |
463 IfNode* iff = entry->in(0)->as_If(); | |
464 ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con); | |
465 Node* rgn = uncommon_proj->unique_ctrl_out(); | |
466 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); | |
467 entry = entry->in(0)->in(0); | |
468 while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) { | |
469 uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con); | |
470 if (uncommon_proj->unique_ctrl_out() != rgn) | |
471 break; | |
472 entry = entry->in(0)->in(0); | |
473 } | |
474 } | |
475 } | |
476 return entry; | |
477 } | |
478 | |
479 //--------------------------find_predicate_insertion_point------------------- | |
480 // Find a good location to insert a predicate | |
481 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) { | |
482 if (start_c == NULL || !start_c->is_Proj()) | |
483 return NULL; | |
484 if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) { | |
485 return start_c->as_Proj(); | |
486 } | |
487 return NULL; | |
488 } | |
489 | |
490 //--------------------------find_predicate------------------------------------ | |
491 // Find a predicate | |
492 Node* PhaseIdealLoop::find_predicate(Node* entry) { | |
493 Node* predicate = NULL; | |
494 if (UseLoopPredicate) { | |
495 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); | |
496 if (predicate != NULL) { // right pattern that can be used by loop predication | |
497 assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); | |
498 return entry; | |
499 } | |
500 } | |
501 return NULL; | |
502 } | |
503 | |
504 //------------------------------Invariance----------------------------------- | |
505 // Helper class for loop_predication_impl to compute invariance on the fly and | |
506 // clone invariants. | |
507 class Invariance : public StackObj { | |
508 VectorSet _visited, _invariant; | |
509 Node_Stack _stack; | |
510 VectorSet _clone_visited; | |
511 Node_List _old_new; // map of old to new (clone) | |
512 IdealLoopTree* _lpt; | |
513 PhaseIdealLoop* _phase; | |
514 | |
515 // Helper function to set up the invariance for invariance computation | |
516 // If n is a known invariant, set up directly. Otherwise, look up the | |
517 // the possibility to push n onto the stack for further processing. | |
518 void visit(Node* use, Node* n) { | |
519 if (_lpt->is_invariant(n)) { // known invariant | |
520 _invariant.set(n->_idx); | |
521 } else if (!n->is_CFG()) { | |
522 Node *n_ctrl = _phase->ctrl_or_self(n); | |
523 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG | |
524 if (_phase->is_dominator(n_ctrl, u_ctrl)) { | |
525 _stack.push(n, n->in(0) == NULL ? 1 : 0); | |
526 } | |
527 } | |
528 } | |
529 | |
530 // Compute invariance for "the_node" and (possibly) all its inputs recursively | |
531 // on the fly | |
532 void compute_invariance(Node* n) { | |
533 assert(_visited.test(n->_idx), "must be"); | |
534 visit(n, n); | |
535 while (_stack.is_nonempty()) { | |
536 Node* n = _stack.node(); | |
537 uint idx = _stack.index(); | |
538 if (idx == n->req()) { // all inputs are processed | |
539 _stack.pop(); | |
540 // n is invariant if it's inputs are all invariant | |
541 bool all_inputs_invariant = true; | |
542 for (uint i = 0; i < n->req(); i++) { | |
543 Node* in = n->in(i); | |
544 if (in == NULL) continue; | |
545 assert(_visited.test(in->_idx), "must have visited input"); | |
546 if (!_invariant.test(in->_idx)) { // bad guy | |
547 all_inputs_invariant = false; | |
548 break; | |
549 } | |
550 } | |
551 if (all_inputs_invariant) { | |
552 _invariant.set(n->_idx); // I am a invariant too | |
553 } | |
554 } else { // process next input | |
555 _stack.set_index(idx + 1); | |
556 Node* m = n->in(idx); | |
557 if (m != NULL && !_visited.test_set(m->_idx)) { | |
558 visit(n, m); | |
559 } | |
560 } | |
561 } | |
562 } | |
563 | |
564 // Helper function to set up _old_new map for clone_nodes. | |
565 // If n is a known invariant, set up directly ("clone" of n == n). | |
566 // Otherwise, push n onto the stack for real cloning. | |
567 void clone_visit(Node* n) { | |
568 assert(_invariant.test(n->_idx), "must be invariant"); | |
569 if (_lpt->is_invariant(n)) { // known invariant | |
570 _old_new.map(n->_idx, n); | |
571 } else { // to be cloned | |
572 assert(!n->is_CFG(), "should not see CFG here"); | |
573 _stack.push(n, n->in(0) == NULL ? 1 : 0); | |
574 } | |
575 } | |
576 | |
577 // Clone "n" and (possibly) all its inputs recursively | |
578 void clone_nodes(Node* n, Node* ctrl) { | |
579 clone_visit(n); | |
580 while (_stack.is_nonempty()) { | |
581 Node* n = _stack.node(); | |
582 uint idx = _stack.index(); | |
583 if (idx == n->req()) { // all inputs processed, clone n! | |
584 _stack.pop(); | |
585 // clone invariant node | |
586 Node* n_cl = n->clone(); | |
587 _old_new.map(n->_idx, n_cl); | |
588 _phase->register_new_node(n_cl, ctrl); | |
589 for (uint i = 0; i < n->req(); i++) { | |
590 Node* in = n_cl->in(i); | |
591 if (in == NULL) continue; | |
592 n_cl->set_req(i, _old_new[in->_idx]); | |
593 } | |
594 } else { // process next input | |
595 _stack.set_index(idx + 1); | |
596 Node* m = n->in(idx); | |
597 if (m != NULL && !_clone_visited.test_set(m->_idx)) { | |
598 clone_visit(m); // visit the input | |
599 } | |
600 } | |
601 } | |
602 } | |
603 | |
604 public: | |
605 Invariance(Arena* area, IdealLoopTree* lpt) : | |
606 _lpt(lpt), _phase(lpt->_phase), | |
607 _visited(area), _invariant(area), _stack(area, 10 /* guess */), | |
608 _clone_visited(area), _old_new(area) | |
609 {} | |
610 | |
611 // Map old to n for invariance computation and clone | |
612 void map_ctrl(Node* old, Node* n) { | |
613 assert(old->is_CFG() && n->is_CFG(), "must be"); | |
614 _old_new.map(old->_idx, n); // "clone" of old is n | |
615 _invariant.set(old->_idx); // old is invariant | |
616 _clone_visited.set(old->_idx); | |
617 } | |
618 | |
619 // Driver function to compute invariance | |
620 bool is_invariant(Node* n) { | |
621 if (!_visited.test_set(n->_idx)) | |
622 compute_invariance(n); | |
623 return (_invariant.test(n->_idx) != 0); | |
624 } | |
625 | |
626 // Driver function to clone invariant | |
627 Node* clone(Node* n, Node* ctrl) { | |
628 assert(ctrl->is_CFG(), "must be"); | |
629 assert(_invariant.test(n->_idx), "must be an invariant"); | |
630 if (!_clone_visited.test(n->_idx)) | |
631 clone_nodes(n, ctrl); | |
632 return _old_new[n->_idx]; | |
633 } | |
634 }; | |
635 | |
636 //------------------------------is_range_check_if ----------------------------------- | |
637 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format | |
638 // Note: this function is particularly designed for loop predication. We require load_range | |
639 // and offset to be loop invariant computed on the fly by "invar" | |
640 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const { | |
641 if (!is_loop_exit(iff)) { | |
642 return false; | |
643 } | |
644 if (!iff->in(1)->is_Bool()) { | |
645 return false; | |
646 } | |
647 const BoolNode *bol = iff->in(1)->as_Bool(); | |
648 if (bol->_test._test != BoolTest::lt) { | |
649 return false; | |
650 } | |
651 if (!bol->in(1)->is_Cmp()) { | |
652 return false; | |
653 } | |
654 const CmpNode *cmp = bol->in(1)->as_Cmp(); | |
655 if (cmp->Opcode() != Op_CmpU) { | |
656 return false; | |
657 } | |
658 Node* range = cmp->in(2); | |
659 if (range->Opcode() != Op_LoadRange) { | |
660 const TypeInt* tint = phase->_igvn.type(range)->isa_int(); | |
661 if (!OptimizeFill || tint == NULL || tint->empty() || tint->_lo < 0) { | |
662 // Allow predication on positive values that aren't LoadRanges. | |
663 // This allows optimization of loops where the length of the | |
664 // array is a known value and doesn't need to be loaded back | |
665 // from the array. | |
666 return false; | |
667 } | |
668 } | |
669 if (!invar.is_invariant(range)) { | |
670 return false; | |
671 } | |
672 Node *iv = _head->as_CountedLoop()->phi(); | |
673 int scale = 0; | |
674 Node *offset = NULL; | |
675 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) { | |
676 return false; | |
677 } | |
678 if (offset && !invar.is_invariant(offset)) { // offset must be invariant | |
679 return false; | |
680 } | |
681 return true; | |
682 } | |
683 | |
684 //------------------------------rc_predicate----------------------------------- | |
685 // Create a range check predicate | |
686 // | |
687 // for (i = init; i < limit; i += stride) { | |
688 // a[scale*i+offset] | |
689 // } | |
690 // | |
691 // Compute max(scale*i + offset) for init <= i < limit and build the predicate | |
692 // as "max(scale*i + offset) u< a.length". | |
693 // | |
694 // There are two cases for max(scale*i + offset): | |
695 // (1) stride*scale > 0 | |
696 // max(scale*i + offset) = scale*(limit-stride) + offset | |
697 // (2) stride*scale < 0 | |
698 // max(scale*i + offset) = scale*init + offset | |
699 BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl, | |
700 int scale, Node* offset, | |
701 Node* init, Node* limit, Node* stride, | |
702 Node* range, bool upper) { | |
703 DEBUG_ONLY(ttyLocker ttyl); | |
704 if (TraceLoopPredicate) tty->print("rc_predicate "); | |
705 | |
706 Node* max_idx_expr = init; | |
707 int stride_con = stride->get_int(); | |
708 if ((stride_con > 0) == (scale > 0) == upper) { | |
709 max_idx_expr = new (C, 3) SubINode(limit, stride); | |
710 register_new_node(max_idx_expr, ctrl); | |
711 if (TraceLoopPredicate) tty->print("(limit - stride) "); | |
712 } else { | |
713 if (TraceLoopPredicate) tty->print("init "); | |
714 } | |
715 | |
716 if (scale != 1) { | |
717 ConNode* con_scale = _igvn.intcon(scale); | |
718 max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale); | |
719 register_new_node(max_idx_expr, ctrl); | |
720 if (TraceLoopPredicate) tty->print("* %d ", scale); | |
721 } | |
722 | |
723 if (offset && (!offset->is_Con() || offset->get_int() != 0)){ | |
724 max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset); | |
725 register_new_node(max_idx_expr, ctrl); | |
726 if (TraceLoopPredicate) | |
727 if (offset->is_Con()) tty->print("+ %d ", offset->get_int()); | |
728 else tty->print("+ offset "); | |
729 } | |
730 | |
731 CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range); | |
732 register_new_node(cmp, ctrl); | |
733 BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt); | |
734 register_new_node(bol, ctrl); | |
735 | |
736 if (TraceLoopPredicate) tty->print_cr("<u range"); | |
737 return bol; | |
738 } | |
739 | |
740 //------------------------------ loop_predication_impl-------------------------- | |
741 // Insert loop predicates for null checks and range checks | |
742 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) { | |
743 if (!UseLoopPredicate) return false; | |
744 | |
745 if (!loop->_head->is_Loop()) { | |
746 // Could be a simple region when irreducible loops are present. | |
747 return false; | |
748 } | |
749 | |
750 if (loop->_head->unique_ctrl_out()->Opcode() == Op_NeverBranch) { | |
751 // do nothing for infinite loops | |
752 return false; | |
753 } | |
754 | |
755 CountedLoopNode *cl = NULL; | |
756 if (loop->_head->is_CountedLoop()) { | |
757 cl = loop->_head->as_CountedLoop(); | |
758 // do nothing for iteration-splitted loops | |
759 if (!cl->is_normal_loop()) return false; | |
760 } | |
761 | |
762 LoopNode *lpn = loop->_head->as_Loop(); | |
763 Node* entry = lpn->in(LoopNode::EntryControl); | |
764 | |
765 ProjNode *predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); | |
766 if (!predicate_proj) { | |
767 #ifndef PRODUCT | |
768 if (TraceLoopPredicate) { | |
769 tty->print("missing predicate:"); | |
770 loop->dump_head(); | |
771 lpn->dump(1); | |
772 } | |
773 #endif | |
774 return false; | |
775 } | |
776 ConNode* zero = _igvn.intcon(0); | |
777 set_ctrl(zero, C->root()); | |
778 | |
779 ResourceArea *area = Thread::current()->resource_area(); | |
780 Invariance invar(area, loop); | |
781 | |
782 // Create list of if-projs such that a newer proj dominates all older | |
783 // projs in the list, and they all dominate loop->tail() | |
784 Node_List if_proj_list(area); | |
785 LoopNode *head = loop->_head->as_Loop(); | |
786 Node *current_proj = loop->tail(); //start from tail | |
787 while (current_proj != head) { | |
788 if (loop == get_loop(current_proj) && // still in the loop ? | |
789 current_proj->is_Proj() && // is a projection ? | |
790 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ? | |
791 if_proj_list.push(current_proj); | |
792 } | |
793 current_proj = idom(current_proj); | |
794 } | |
795 | |
796 bool hoisted = false; // true if at least one proj is promoted | |
797 while (if_proj_list.size() > 0) { | |
798 // Following are changed to nonnull when a predicate can be hoisted | |
799 ProjNode* new_predicate_proj = NULL; | |
800 | |
801 ProjNode* proj = if_proj_list.pop()->as_Proj(); | |
802 IfNode* iff = proj->in(0)->as_If(); | |
803 | |
804 if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) { | |
805 if (loop->is_loop_exit(iff)) { | |
806 // stop processing the remaining projs in the list because the execution of them | |
807 // depends on the condition of "iff" (iff->in(1)). | |
808 break; | |
809 } else { | |
810 // Both arms are inside the loop. There are two cases: | |
811 // (1) there is one backward branch. In this case, any remaining proj | |
812 // in the if_proj list post-dominates "iff". So, the condition of "iff" | |
813 // does not determine the execution the remining projs directly, and we | |
814 // can safely continue. | |
815 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj" | |
816 // does not dominate loop->tail(), so it can not be in the if_proj list. | |
817 continue; | |
818 } | |
819 } | |
820 | |
821 Node* test = iff->in(1); | |
822 if (!test->is_Bool()){ //Conv2B, ... | |
823 continue; | |
824 } | |
825 BoolNode* bol = test->as_Bool(); | |
826 if (invar.is_invariant(bol)) { | |
827 // Invariant test | |
828 new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL, | |
829 Deoptimization::Reason_predicate); | |
830 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0); | |
831 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool(); | |
832 | |
833 // Negate test if necessary | |
834 bool negated = false; | |
835 if (proj->_con != predicate_proj->_con) { | |
836 new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate()); | |
837 register_new_node(new_predicate_bol, ctrl); | |
838 negated = true; | |
839 } | |
840 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If(); | |
841 _igvn.hash_delete(new_predicate_iff); | |
842 new_predicate_iff->set_req(1, new_predicate_bol); | |
843 #ifndef PRODUCT | |
844 if (TraceLoopPredicate) { | |
845 tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx); | |
846 loop->dump_head(); | |
847 } else if (TraceLoopOpts) { | |
848 tty->print("Predicate IC "); | |
849 loop->dump_head(); | |
850 } | |
851 #endif | |
852 } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) { | |
853 assert(proj->_con == predicate_proj->_con, "must match"); | |
854 | |
855 // Range check for counted loops | |
856 const Node* cmp = bol->in(1)->as_Cmp(); | |
857 Node* idx = cmp->in(1); | |
858 assert(!invar.is_invariant(idx), "index is variant"); | |
859 assert(cmp->in(2)->Opcode() == Op_LoadRange || OptimizeFill, "must be"); | |
860 Node* rng = cmp->in(2); | |
861 assert(invar.is_invariant(rng), "range must be invariant"); | |
862 int scale = 1; | |
863 Node* offset = zero; | |
864 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset); | |
865 assert(ok, "must be index expression"); | |
866 | |
867 Node* init = cl->init_trip(); | |
868 Node* limit = cl->limit(); | |
869 Node* stride = cl->stride(); | |
870 | |
871 // Build if's for the upper and lower bound tests. The | |
872 // lower_bound test will dominate the upper bound test and all | |
873 // cloned or created nodes will use the lower bound test as | |
874 // their declared control. | |
875 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate); | |
876 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate); | |
877 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate"); | |
878 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0); | |
879 | |
880 // Perform cloning to keep Invariance state correct since the | |
881 // late schedule will place invariant things in the loop. | |
882 rng = invar.clone(rng, ctrl); | |
883 if (offset && offset != zero) { | |
884 assert(invar.is_invariant(offset), "offset must be loop invariant"); | |
885 offset = invar.clone(offset, ctrl); | |
886 } | |
887 | |
888 // Test the lower bound | |
889 Node* lower_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, false); | |
890 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If(); | |
891 _igvn.hash_delete(lower_bound_iff); | |
892 lower_bound_iff->set_req(1, lower_bound_bol); | |
893 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx); | |
894 | |
895 // Test the upper bound | |
896 Node* upper_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, true); | |
897 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If(); | |
898 _igvn.hash_delete(upper_bound_iff); | |
899 upper_bound_iff->set_req(1, upper_bound_bol); | |
900 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx); | |
901 | |
902 // Fall through into rest of the clean up code which will move | |
903 // any dependent nodes onto the upper bound test. | |
904 new_predicate_proj = upper_bound_proj; | |
905 | |
906 #ifndef PRODUCT | |
907 if (TraceLoopOpts && !TraceLoopPredicate) { | |
908 tty->print("Predicate RC "); | |
909 loop->dump_head(); | |
910 } | |
911 #endif | |
912 } else { | |
913 // Loop variant check (for example, range check in non-counted loop) | |
914 // with uncommon trap. | |
915 continue; | |
916 } | |
917 assert(new_predicate_proj != NULL, "sanity"); | |
918 // Success - attach condition (new_predicate_bol) to predicate if | |
919 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate | |
920 | |
921 // Eliminate the old If in the loop body | |
922 dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con ); | |
923 | |
924 hoisted = true; | |
925 C->set_major_progress(); | |
926 } // end while | |
927 | |
928 #ifndef PRODUCT | |
929 // report that the loop predication has been actually performed | |
930 // for this loop | |
931 if (TraceLoopPredicate && hoisted) { | |
932 tty->print("Loop Predication Performed:"); | |
933 loop->dump_head(); | |
934 } | |
935 #endif | |
936 | |
937 return hoisted; | |
938 } | |
939 | |
940 //------------------------------loop_predication-------------------------------- | |
941 // driver routine for loop predication optimization | |
942 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) { | |
943 bool hoisted = false; | |
944 // Recursively promote predicates | |
945 if (_child) { | |
946 hoisted = _child->loop_predication( phase); | |
947 } | |
948 | |
949 // self | |
950 if (!_irreducible && !tail()->is_top()) { | |
951 hoisted |= phase->loop_predication_impl(this); | |
952 } | |
953 | |
954 if (_next) { //sibling | |
955 hoisted |= _next->loop_predication( phase); | |
956 } | |
957 | |
958 return hoisted; | |
959 } | |
960 |