|
1 /* |
|
2 * Copyright (c) 2011, 2014, 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 //-------------------------------register_control------------------------- |
|
45 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) { |
|
46 assert(n->is_CFG(), "must be control node"); |
|
47 _igvn.register_new_node_with_optimizer(n); |
|
48 loop->_body.push(n); |
|
49 set_loop(n, loop); |
|
50 // When called from beautify_loops() idom is not constructed yet. |
|
51 if (_idom != NULL) { |
|
52 set_idom(n, pred, dom_depth(pred)); |
|
53 } |
|
54 } |
|
55 |
|
56 //------------------------------create_new_if_for_predicate------------------------ |
|
57 // create a new if above the uct_if_pattern for the predicate to be promoted. |
|
58 // |
|
59 // before after |
|
60 // ---------- ---------- |
|
61 // ctrl ctrl |
|
62 // | | |
|
63 // | | |
|
64 // v v |
|
65 // iff new_iff |
|
66 // / \ / \ |
|
67 // / \ / \ |
|
68 // v v v v |
|
69 // uncommon_proj cont_proj if_uct if_cont |
|
70 // \ | | | | |
|
71 // \ | | | | |
|
72 // v v v | v |
|
73 // rgn loop | iff |
|
74 // | | / \ |
|
75 // | | / \ |
|
76 // v | v v |
|
77 // uncommon_trap | uncommon_proj cont_proj |
|
78 // \ \ | | |
|
79 // \ \ | | |
|
80 // v v v v |
|
81 // rgn loop |
|
82 // | |
|
83 // | |
|
84 // v |
|
85 // uncommon_trap |
|
86 // |
|
87 // |
|
88 // We will create a region to guard the uct call if there is no one there. |
|
89 // The true projecttion (if_cont) of the new_iff is returned. |
|
90 // This code is also used to clone predicates to clonned loops. |
|
91 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, |
|
92 Deoptimization::DeoptReason reason) { |
|
93 assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!"); |
|
94 IfNode* iff = cont_proj->in(0)->as_If(); |
|
95 |
|
96 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); |
|
97 Node *rgn = uncommon_proj->unique_ctrl_out(); |
|
98 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); |
|
99 |
|
100 uint proj_index = 1; // region's edge corresponding to uncommon_proj |
|
101 if (!rgn->is_Region()) { // create a region to guard the call |
|
102 assert(rgn->is_Call(), "must be call uct"); |
|
103 CallNode* call = rgn->as_Call(); |
|
104 IdealLoopTree* loop = get_loop(call); |
|
105 rgn = new (C) RegionNode(1); |
|
106 rgn->add_req(uncommon_proj); |
|
107 register_control(rgn, loop, uncommon_proj); |
|
108 _igvn.hash_delete(call); |
|
109 call->set_req(0, rgn); |
|
110 // When called from beautify_loops() idom is not constructed yet. |
|
111 if (_idom != NULL) { |
|
112 set_idom(call, rgn, dom_depth(rgn)); |
|
113 } |
|
114 } else { |
|
115 // Find region's edge corresponding to uncommon_proj |
|
116 for (; proj_index < rgn->req(); proj_index++) |
|
117 if (rgn->in(proj_index) == uncommon_proj) break; |
|
118 assert(proj_index < rgn->req(), "sanity"); |
|
119 } |
|
120 |
|
121 Node* entry = iff->in(0); |
|
122 if (new_entry != NULL) { |
|
123 // Clonning the predicate to new location. |
|
124 entry = new_entry; |
|
125 } |
|
126 // Create new_iff |
|
127 IdealLoopTree* lp = get_loop(entry); |
|
128 IfNode *new_iff = iff->clone()->as_If(); |
|
129 new_iff->set_req(0, entry); |
|
130 register_control(new_iff, lp, entry); |
|
131 Node *if_cont = new (C) IfTrueNode(new_iff); |
|
132 Node *if_uct = new (C) IfFalseNode(new_iff); |
|
133 if (cont_proj->is_IfFalse()) { |
|
134 // Swap |
|
135 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; |
|
136 } |
|
137 register_control(if_cont, lp, new_iff); |
|
138 register_control(if_uct, get_loop(rgn), new_iff); |
|
139 |
|
140 // if_uct to rgn |
|
141 _igvn.hash_delete(rgn); |
|
142 rgn->add_req(if_uct); |
|
143 // When called from beautify_loops() idom is not constructed yet. |
|
144 if (_idom != NULL) { |
|
145 Node* ridom = idom(rgn); |
|
146 Node* nrdom = dom_lca(ridom, new_iff); |
|
147 set_idom(rgn, nrdom, dom_depth(rgn)); |
|
148 } |
|
149 |
|
150 // If rgn has phis add new edges which has the same |
|
151 // value as on original uncommon_proj pass. |
|
152 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last"); |
|
153 bool has_phi = false; |
|
154 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) { |
|
155 Node* use = rgn->fast_out(i); |
|
156 if (use->is_Phi() && use->outcnt() > 0) { |
|
157 assert(use->in(0) == rgn, ""); |
|
158 _igvn.rehash_node_delayed(use); |
|
159 use->add_req(use->in(proj_index)); |
|
160 has_phi = true; |
|
161 } |
|
162 } |
|
163 assert(!has_phi || rgn->req() > 3, "no phis when region is created"); |
|
164 |
|
165 if (new_entry == NULL) { |
|
166 // Attach if_cont to iff |
|
167 _igvn.hash_delete(iff); |
|
168 iff->set_req(0, if_cont); |
|
169 if (_idom != NULL) { |
|
170 set_idom(iff, if_cont, dom_depth(iff)); |
|
171 } |
|
172 } |
|
173 return if_cont->as_Proj(); |
|
174 } |
|
175 |
|
176 //------------------------------create_new_if_for_predicate------------------------ |
|
177 // Create a new if below new_entry for the predicate to be cloned (IGVN optimization) |
|
178 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, |
|
179 Deoptimization::DeoptReason reason) { |
|
180 assert(new_entry != 0, "only used for clone predicate"); |
|
181 assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!"); |
|
182 IfNode* iff = cont_proj->in(0)->as_If(); |
|
183 |
|
184 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); |
|
185 Node *rgn = uncommon_proj->unique_ctrl_out(); |
|
186 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); |
|
187 |
|
188 uint proj_index = 1; // region's edge corresponding to uncommon_proj |
|
189 if (!rgn->is_Region()) { // create a region to guard the call |
|
190 assert(rgn->is_Call(), "must be call uct"); |
|
191 CallNode* call = rgn->as_Call(); |
|
192 rgn = new (C) RegionNode(1); |
|
193 register_new_node_with_optimizer(rgn); |
|
194 rgn->add_req(uncommon_proj); |
|
195 hash_delete(call); |
|
196 call->set_req(0, rgn); |
|
197 } else { |
|
198 // Find region's edge corresponding to uncommon_proj |
|
199 for (; proj_index < rgn->req(); proj_index++) |
|
200 if (rgn->in(proj_index) == uncommon_proj) break; |
|
201 assert(proj_index < rgn->req(), "sanity"); |
|
202 } |
|
203 |
|
204 // Create new_iff in new location. |
|
205 IfNode *new_iff = iff->clone()->as_If(); |
|
206 new_iff->set_req(0, new_entry); |
|
207 |
|
208 register_new_node_with_optimizer(new_iff); |
|
209 Node *if_cont = new (C) IfTrueNode(new_iff); |
|
210 Node *if_uct = new (C) IfFalseNode(new_iff); |
|
211 if (cont_proj->is_IfFalse()) { |
|
212 // Swap |
|
213 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; |
|
214 } |
|
215 register_new_node_with_optimizer(if_cont); |
|
216 register_new_node_with_optimizer(if_uct); |
|
217 |
|
218 // if_uct to rgn |
|
219 hash_delete(rgn); |
|
220 rgn->add_req(if_uct); |
|
221 |
|
222 // If rgn has phis add corresponding new edges which has the same |
|
223 // value as on original uncommon_proj pass. |
|
224 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last"); |
|
225 bool has_phi = false; |
|
226 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) { |
|
227 Node* use = rgn->fast_out(i); |
|
228 if (use->is_Phi() && use->outcnt() > 0) { |
|
229 rehash_node_delayed(use); |
|
230 use->add_req(use->in(proj_index)); |
|
231 has_phi = true; |
|
232 } |
|
233 } |
|
234 assert(!has_phi || rgn->req() > 3, "no phis when region is created"); |
|
235 |
|
236 return if_cont->as_Proj(); |
|
237 } |
|
238 |
|
239 //--------------------------clone_predicate----------------------- |
|
240 ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry, |
|
241 Deoptimization::DeoptReason reason, |
|
242 PhaseIdealLoop* loop_phase, |
|
243 PhaseIterGVN* igvn) { |
|
244 ProjNode* new_predicate_proj; |
|
245 if (loop_phase != NULL) { |
|
246 new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason); |
|
247 } else { |
|
248 new_predicate_proj = igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason); |
|
249 } |
|
250 IfNode* iff = new_predicate_proj->in(0)->as_If(); |
|
251 Node* ctrl = iff->in(0); |
|
252 |
|
253 // Match original condition since predicate's projections could be swapped. |
|
254 assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); |
|
255 Node* opq = new (igvn->C) Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1)); |
|
256 igvn->C->add_predicate_opaq(opq); |
|
257 |
|
258 Node* bol = new (igvn->C) Conv2BNode(opq); |
|
259 if (loop_phase != NULL) { |
|
260 loop_phase->register_new_node(opq, ctrl); |
|
261 loop_phase->register_new_node(bol, ctrl); |
|
262 } else { |
|
263 igvn->register_new_node_with_optimizer(opq); |
|
264 igvn->register_new_node_with_optimizer(bol); |
|
265 } |
|
266 igvn->hash_delete(iff); |
|
267 iff->set_req(1, bol); |
|
268 return new_predicate_proj; |
|
269 } |
|
270 |
|
271 |
|
272 //--------------------------clone_loop_predicates----------------------- |
|
273 // Interface from IGVN |
|
274 Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) { |
|
275 return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, clone_limit_check, NULL, this); |
|
276 } |
|
277 |
|
278 // Interface from PhaseIdealLoop |
|
279 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) { |
|
280 return clone_loop_predicates(old_entry, new_entry, clone_limit_check, this, &this->_igvn); |
|
281 } |
|
282 |
|
283 // Clone loop predicates to cloned loops (peeled, unswitched, split_if). |
|
284 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, |
|
285 bool clone_limit_check, |
|
286 PhaseIdealLoop* loop_phase, |
|
287 PhaseIterGVN* igvn) { |
|
288 #ifdef ASSERT |
|
289 if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) { |
|
290 if (new_entry != NULL) |
|
291 new_entry->dump(); |
|
292 assert(false, "not IfTrue, IfFalse, Region or SafePoint"); |
|
293 } |
|
294 #endif |
|
295 // Search original predicates |
|
296 Node* entry = old_entry; |
|
297 ProjNode* limit_check_proj = NULL; |
|
298 if (LoopLimitCheck) { |
|
299 limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
|
300 if (limit_check_proj != NULL) { |
|
301 entry = entry->in(0)->in(0); |
|
302 } |
|
303 } |
|
304 if (UseLoopPredicate) { |
|
305 ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
|
306 if (predicate_proj != NULL) { // right pattern that can be used by loop predication |
|
307 // clone predicate |
|
308 new_entry = clone_predicate(predicate_proj, new_entry, |
|
309 Deoptimization::Reason_predicate, |
|
310 loop_phase, igvn); |
|
311 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate"); |
|
312 if (TraceLoopPredicate) { |
|
313 tty->print("Loop Predicate cloned: "); |
|
314 debug_only( new_entry->in(0)->dump(); ) |
|
315 } |
|
316 } |
|
317 } |
|
318 if (limit_check_proj != NULL && clone_limit_check) { |
|
319 // Clone loop limit check last to insert it before loop. |
|
320 // Don't clone a limit check which was already finalized |
|
321 // for this counted loop (only one limit check is needed). |
|
322 new_entry = clone_predicate(limit_check_proj, new_entry, |
|
323 Deoptimization::Reason_loop_limit_check, |
|
324 loop_phase, igvn); |
|
325 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check"); |
|
326 if (TraceLoopLimitCheck) { |
|
327 tty->print("Loop Limit Check cloned: "); |
|
328 debug_only( new_entry->in(0)->dump(); ) |
|
329 } |
|
330 } |
|
331 return new_entry; |
|
332 } |
|
333 |
|
334 //--------------------------skip_loop_predicates------------------------------ |
|
335 // Skip related predicates. |
|
336 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) { |
|
337 Node* predicate = NULL; |
|
338 if (LoopLimitCheck) { |
|
339 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
|
340 if (predicate != NULL) { |
|
341 entry = entry->in(0)->in(0); |
|
342 } |
|
343 } |
|
344 if (UseLoopPredicate) { |
|
345 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
|
346 if (predicate != NULL) { // right pattern that can be used by loop predication |
|
347 IfNode* iff = entry->in(0)->as_If(); |
|
348 ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con); |
|
349 Node* rgn = uncommon_proj->unique_ctrl_out(); |
|
350 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); |
|
351 entry = entry->in(0)->in(0); |
|
352 while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) { |
|
353 uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con); |
|
354 if (uncommon_proj->unique_ctrl_out() != rgn) |
|
355 break; |
|
356 entry = entry->in(0)->in(0); |
|
357 } |
|
358 } |
|
359 } |
|
360 return entry; |
|
361 } |
|
362 |
|
363 //--------------------------find_predicate_insertion_point------------------- |
|
364 // Find a good location to insert a predicate |
|
365 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) { |
|
366 if (start_c == NULL || !start_c->is_Proj()) |
|
367 return NULL; |
|
368 if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) { |
|
369 return start_c->as_Proj(); |
|
370 } |
|
371 return NULL; |
|
372 } |
|
373 |
|
374 //--------------------------find_predicate------------------------------------ |
|
375 // Find a predicate |
|
376 Node* PhaseIdealLoop::find_predicate(Node* entry) { |
|
377 Node* predicate = NULL; |
|
378 if (LoopLimitCheck) { |
|
379 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
|
380 if (predicate != NULL) { // right pattern that can be used by loop predication |
|
381 return entry; |
|
382 } |
|
383 } |
|
384 if (UseLoopPredicate) { |
|
385 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
|
386 if (predicate != NULL) { // right pattern that can be used by loop predication |
|
387 return entry; |
|
388 } |
|
389 } |
|
390 return NULL; |
|
391 } |
|
392 |
|
393 //------------------------------Invariance----------------------------------- |
|
394 // Helper class for loop_predication_impl to compute invariance on the fly and |
|
395 // clone invariants. |
|
396 class Invariance : public StackObj { |
|
397 VectorSet _visited, _invariant; |
|
398 Node_Stack _stack; |
|
399 VectorSet _clone_visited; |
|
400 Node_List _old_new; // map of old to new (clone) |
|
401 IdealLoopTree* _lpt; |
|
402 PhaseIdealLoop* _phase; |
|
403 |
|
404 // Helper function to set up the invariance for invariance computation |
|
405 // If n is a known invariant, set up directly. Otherwise, look up the |
|
406 // the possibility to push n onto the stack for further processing. |
|
407 void visit(Node* use, Node* n) { |
|
408 if (_lpt->is_invariant(n)) { // known invariant |
|
409 _invariant.set(n->_idx); |
|
410 } else if (!n->is_CFG()) { |
|
411 Node *n_ctrl = _phase->ctrl_or_self(n); |
|
412 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG |
|
413 if (_phase->is_dominator(n_ctrl, u_ctrl)) { |
|
414 _stack.push(n, n->in(0) == NULL ? 1 : 0); |
|
415 } |
|
416 } |
|
417 } |
|
418 |
|
419 // Compute invariance for "the_node" and (possibly) all its inputs recursively |
|
420 // on the fly |
|
421 void compute_invariance(Node* n) { |
|
422 assert(_visited.test(n->_idx), "must be"); |
|
423 visit(n, n); |
|
424 while (_stack.is_nonempty()) { |
|
425 Node* n = _stack.node(); |
|
426 uint idx = _stack.index(); |
|
427 if (idx == n->req()) { // all inputs are processed |
|
428 _stack.pop(); |
|
429 // n is invariant if it's inputs are all invariant |
|
430 bool all_inputs_invariant = true; |
|
431 for (uint i = 0; i < n->req(); i++) { |
|
432 Node* in = n->in(i); |
|
433 if (in == NULL) continue; |
|
434 assert(_visited.test(in->_idx), "must have visited input"); |
|
435 if (!_invariant.test(in->_idx)) { // bad guy |
|
436 all_inputs_invariant = false; |
|
437 break; |
|
438 } |
|
439 } |
|
440 if (all_inputs_invariant) { |
|
441 _invariant.set(n->_idx); // I am a invariant too |
|
442 } |
|
443 } else { // process next input |
|
444 _stack.set_index(idx + 1); |
|
445 Node* m = n->in(idx); |
|
446 if (m != NULL && !_visited.test_set(m->_idx)) { |
|
447 visit(n, m); |
|
448 } |
|
449 } |
|
450 } |
|
451 } |
|
452 |
|
453 // Helper function to set up _old_new map for clone_nodes. |
|
454 // If n is a known invariant, set up directly ("clone" of n == n). |
|
455 // Otherwise, push n onto the stack for real cloning. |
|
456 void clone_visit(Node* n) { |
|
457 assert(_invariant.test(n->_idx), "must be invariant"); |
|
458 if (_lpt->is_invariant(n)) { // known invariant |
|
459 _old_new.map(n->_idx, n); |
|
460 } else { // to be cloned |
|
461 assert(!n->is_CFG(), "should not see CFG here"); |
|
462 _stack.push(n, n->in(0) == NULL ? 1 : 0); |
|
463 } |
|
464 } |
|
465 |
|
466 // Clone "n" and (possibly) all its inputs recursively |
|
467 void clone_nodes(Node* n, Node* ctrl) { |
|
468 clone_visit(n); |
|
469 while (_stack.is_nonempty()) { |
|
470 Node* n = _stack.node(); |
|
471 uint idx = _stack.index(); |
|
472 if (idx == n->req()) { // all inputs processed, clone n! |
|
473 _stack.pop(); |
|
474 // clone invariant node |
|
475 Node* n_cl = n->clone(); |
|
476 _old_new.map(n->_idx, n_cl); |
|
477 _phase->register_new_node(n_cl, ctrl); |
|
478 for (uint i = 0; i < n->req(); i++) { |
|
479 Node* in = n_cl->in(i); |
|
480 if (in == NULL) continue; |
|
481 n_cl->set_req(i, _old_new[in->_idx]); |
|
482 } |
|
483 } else { // process next input |
|
484 _stack.set_index(idx + 1); |
|
485 Node* m = n->in(idx); |
|
486 if (m != NULL && !_clone_visited.test_set(m->_idx)) { |
|
487 clone_visit(m); // visit the input |
|
488 } |
|
489 } |
|
490 } |
|
491 } |
|
492 |
|
493 public: |
|
494 Invariance(Arena* area, IdealLoopTree* lpt) : |
|
495 _lpt(lpt), _phase(lpt->_phase), |
|
496 _visited(area), _invariant(area), _stack(area, 10 /* guess */), |
|
497 _clone_visited(area), _old_new(area) |
|
498 {} |
|
499 |
|
500 // Map old to n for invariance computation and clone |
|
501 void map_ctrl(Node* old, Node* n) { |
|
502 assert(old->is_CFG() && n->is_CFG(), "must be"); |
|
503 _old_new.map(old->_idx, n); // "clone" of old is n |
|
504 _invariant.set(old->_idx); // old is invariant |
|
505 _clone_visited.set(old->_idx); |
|
506 } |
|
507 |
|
508 // Driver function to compute invariance |
|
509 bool is_invariant(Node* n) { |
|
510 if (!_visited.test_set(n->_idx)) |
|
511 compute_invariance(n); |
|
512 return (_invariant.test(n->_idx) != 0); |
|
513 } |
|
514 |
|
515 // Driver function to clone invariant |
|
516 Node* clone(Node* n, Node* ctrl) { |
|
517 assert(ctrl->is_CFG(), "must be"); |
|
518 assert(_invariant.test(n->_idx), "must be an invariant"); |
|
519 if (!_clone_visited.test(n->_idx)) |
|
520 clone_nodes(n, ctrl); |
|
521 return _old_new[n->_idx]; |
|
522 } |
|
523 }; |
|
524 |
|
525 //------------------------------is_range_check_if ----------------------------------- |
|
526 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format |
|
527 // Note: this function is particularly designed for loop predication. We require load_range |
|
528 // and offset to be loop invariant computed on the fly by "invar" |
|
529 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const { |
|
530 if (!is_loop_exit(iff)) { |
|
531 return false; |
|
532 } |
|
533 if (!iff->in(1)->is_Bool()) { |
|
534 return false; |
|
535 } |
|
536 const BoolNode *bol = iff->in(1)->as_Bool(); |
|
537 if (bol->_test._test != BoolTest::lt) { |
|
538 return false; |
|
539 } |
|
540 if (!bol->in(1)->is_Cmp()) { |
|
541 return false; |
|
542 } |
|
543 const CmpNode *cmp = bol->in(1)->as_Cmp(); |
|
544 if (cmp->Opcode() != Op_CmpU) { |
|
545 return false; |
|
546 } |
|
547 Node* range = cmp->in(2); |
|
548 if (range->Opcode() != Op_LoadRange) { |
|
549 const TypeInt* tint = phase->_igvn.type(range)->isa_int(); |
|
550 if (tint == NULL || tint->empty() || tint->_lo < 0) { |
|
551 // Allow predication on positive values that aren't LoadRanges. |
|
552 // This allows optimization of loops where the length of the |
|
553 // array is a known value and doesn't need to be loaded back |
|
554 // from the array. |
|
555 return false; |
|
556 } |
|
557 } |
|
558 if (!invar.is_invariant(range)) { |
|
559 return false; |
|
560 } |
|
561 Node *iv = _head->as_CountedLoop()->phi(); |
|
562 int scale = 0; |
|
563 Node *offset = NULL; |
|
564 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) { |
|
565 return false; |
|
566 } |
|
567 if (offset && !invar.is_invariant(offset)) { // offset must be invariant |
|
568 return false; |
|
569 } |
|
570 return true; |
|
571 } |
|
572 |
|
573 //------------------------------rc_predicate----------------------------------- |
|
574 // Create a range check predicate |
|
575 // |
|
576 // for (i = init; i < limit; i += stride) { |
|
577 // a[scale*i+offset] |
|
578 // } |
|
579 // |
|
580 // Compute max(scale*i + offset) for init <= i < limit and build the predicate |
|
581 // as "max(scale*i + offset) u< a.length". |
|
582 // |
|
583 // There are two cases for max(scale*i + offset): |
|
584 // (1) stride*scale > 0 |
|
585 // max(scale*i + offset) = scale*(limit-stride) + offset |
|
586 // (2) stride*scale < 0 |
|
587 // max(scale*i + offset) = scale*init + offset |
|
588 BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl, |
|
589 int scale, Node* offset, |
|
590 Node* init, Node* limit, Node* stride, |
|
591 Node* range, bool upper) { |
|
592 stringStream* predString = NULL; |
|
593 if (TraceLoopPredicate) { |
|
594 predString = new stringStream(); |
|
595 predString->print("rc_predicate "); |
|
596 } |
|
597 |
|
598 Node* max_idx_expr = init; |
|
599 int stride_con = stride->get_int(); |
|
600 if ((stride_con > 0) == (scale > 0) == upper) { |
|
601 if (LoopLimitCheck) { |
|
602 // With LoopLimitCheck limit is not exact. |
|
603 // Calculate exact limit here. |
|
604 // Note, counted loop's test is '<' or '>'. |
|
605 limit = exact_limit(loop); |
|
606 max_idx_expr = new (C) SubINode(limit, stride); |
|
607 register_new_node(max_idx_expr, ctrl); |
|
608 if (TraceLoopPredicate) predString->print("(limit - stride) "); |
|
609 } else { |
|
610 max_idx_expr = new (C) SubINode(limit, stride); |
|
611 register_new_node(max_idx_expr, ctrl); |
|
612 if (TraceLoopPredicate) predString->print("(limit - stride) "); |
|
613 } |
|
614 } else { |
|
615 if (TraceLoopPredicate) predString->print("init "); |
|
616 } |
|
617 |
|
618 if (scale != 1) { |
|
619 ConNode* con_scale = _igvn.intcon(scale); |
|
620 max_idx_expr = new (C) MulINode(max_idx_expr, con_scale); |
|
621 register_new_node(max_idx_expr, ctrl); |
|
622 if (TraceLoopPredicate) predString->print("* %d ", scale); |
|
623 } |
|
624 |
|
625 if (offset && (!offset->is_Con() || offset->get_int() != 0)){ |
|
626 max_idx_expr = new (C) AddINode(max_idx_expr, offset); |
|
627 register_new_node(max_idx_expr, ctrl); |
|
628 if (TraceLoopPredicate) |
|
629 if (offset->is_Con()) predString->print("+ %d ", offset->get_int()); |
|
630 else predString->print("+ offset "); |
|
631 } |
|
632 |
|
633 CmpUNode* cmp = new (C) CmpUNode(max_idx_expr, range); |
|
634 register_new_node(cmp, ctrl); |
|
635 BoolNode* bol = new (C) BoolNode(cmp, BoolTest::lt); |
|
636 register_new_node(bol, ctrl); |
|
637 |
|
638 if (TraceLoopPredicate) { |
|
639 predString->print_cr("<u range"); |
|
640 tty->print("%s", predString->as_string()); |
|
641 } |
|
642 return bol; |
|
643 } |
|
644 |
|
645 //------------------------------ loop_predication_impl-------------------------- |
|
646 // Insert loop predicates for null checks and range checks |
|
647 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) { |
|
648 if (!UseLoopPredicate) return false; |
|
649 |
|
650 if (!loop->_head->is_Loop()) { |
|
651 // Could be a simple region when irreducible loops are present. |
|
652 return false; |
|
653 } |
|
654 LoopNode* head = loop->_head->as_Loop(); |
|
655 |
|
656 if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) { |
|
657 // do nothing for infinite loops |
|
658 return false; |
|
659 } |
|
660 |
|
661 CountedLoopNode *cl = NULL; |
|
662 if (head->is_valid_counted_loop()) { |
|
663 cl = head->as_CountedLoop(); |
|
664 // do nothing for iteration-splitted loops |
|
665 if (!cl->is_normal_loop()) return false; |
|
666 // Avoid RCE if Counted loop's test is '!='. |
|
667 BoolTest::mask bt = cl->loopexit()->test_trip(); |
|
668 if (bt != BoolTest::lt && bt != BoolTest::gt) |
|
669 cl = NULL; |
|
670 } |
|
671 |
|
672 Node* entry = head->in(LoopNode::EntryControl); |
|
673 ProjNode *predicate_proj = NULL; |
|
674 // Loop limit check predicate should be near the loop. |
|
675 if (LoopLimitCheck) { |
|
676 predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
|
677 if (predicate_proj != NULL) |
|
678 entry = predicate_proj->in(0)->in(0); |
|
679 } |
|
680 |
|
681 predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
|
682 if (!predicate_proj) { |
|
683 #ifndef PRODUCT |
|
684 if (TraceLoopPredicate) { |
|
685 tty->print("missing predicate:"); |
|
686 loop->dump_head(); |
|
687 head->dump(1); |
|
688 } |
|
689 #endif |
|
690 return false; |
|
691 } |
|
692 ConNode* zero = _igvn.intcon(0); |
|
693 set_ctrl(zero, C->root()); |
|
694 |
|
695 ResourceArea *area = Thread::current()->resource_area(); |
|
696 Invariance invar(area, loop); |
|
697 |
|
698 // Create list of if-projs such that a newer proj dominates all older |
|
699 // projs in the list, and they all dominate loop->tail() |
|
700 Node_List if_proj_list(area); |
|
701 Node *current_proj = loop->tail(); //start from tail |
|
702 while (current_proj != head) { |
|
703 if (loop == get_loop(current_proj) && // still in the loop ? |
|
704 current_proj->is_Proj() && // is a projection ? |
|
705 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ? |
|
706 if_proj_list.push(current_proj); |
|
707 } |
|
708 current_proj = idom(current_proj); |
|
709 } |
|
710 |
|
711 bool hoisted = false; // true if at least one proj is promoted |
|
712 while (if_proj_list.size() > 0) { |
|
713 // Following are changed to nonnull when a predicate can be hoisted |
|
714 ProjNode* new_predicate_proj = NULL; |
|
715 |
|
716 ProjNode* proj = if_proj_list.pop()->as_Proj(); |
|
717 IfNode* iff = proj->in(0)->as_If(); |
|
718 |
|
719 if (!proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { |
|
720 if (loop->is_loop_exit(iff)) { |
|
721 // stop processing the remaining projs in the list because the execution of them |
|
722 // depends on the condition of "iff" (iff->in(1)). |
|
723 break; |
|
724 } else { |
|
725 // Both arms are inside the loop. There are two cases: |
|
726 // (1) there is one backward branch. In this case, any remaining proj |
|
727 // in the if_proj list post-dominates "iff". So, the condition of "iff" |
|
728 // does not determine the execution the remining projs directly, and we |
|
729 // can safely continue. |
|
730 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj" |
|
731 // does not dominate loop->tail(), so it can not be in the if_proj list. |
|
732 continue; |
|
733 } |
|
734 } |
|
735 |
|
736 Node* test = iff->in(1); |
|
737 if (!test->is_Bool()){ //Conv2B, ... |
|
738 continue; |
|
739 } |
|
740 BoolNode* bol = test->as_Bool(); |
|
741 if (invar.is_invariant(bol)) { |
|
742 // Invariant test |
|
743 new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL, |
|
744 Deoptimization::Reason_predicate); |
|
745 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0); |
|
746 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool(); |
|
747 |
|
748 // Negate test if necessary |
|
749 bool negated = false; |
|
750 if (proj->_con != predicate_proj->_con) { |
|
751 new_predicate_bol = new (C) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate()); |
|
752 register_new_node(new_predicate_bol, ctrl); |
|
753 negated = true; |
|
754 } |
|
755 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If(); |
|
756 _igvn.hash_delete(new_predicate_iff); |
|
757 new_predicate_iff->set_req(1, new_predicate_bol); |
|
758 #ifndef PRODUCT |
|
759 if (TraceLoopPredicate) { |
|
760 tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx); |
|
761 loop->dump_head(); |
|
762 } else if (TraceLoopOpts) { |
|
763 tty->print("Predicate IC "); |
|
764 loop->dump_head(); |
|
765 } |
|
766 #endif |
|
767 } else if ((cl != NULL) && (proj->_con == predicate_proj->_con) && |
|
768 loop->is_range_check_if(iff, this, invar)) { |
|
769 |
|
770 // Range check for counted loops |
|
771 const Node* cmp = bol->in(1)->as_Cmp(); |
|
772 Node* idx = cmp->in(1); |
|
773 assert(!invar.is_invariant(idx), "index is variant"); |
|
774 Node* rng = cmp->in(2); |
|
775 assert(rng->Opcode() == Op_LoadRange || _igvn.type(rng)->is_int() >= 0, "must be"); |
|
776 assert(invar.is_invariant(rng), "range must be invariant"); |
|
777 int scale = 1; |
|
778 Node* offset = zero; |
|
779 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset); |
|
780 assert(ok, "must be index expression"); |
|
781 |
|
782 Node* init = cl->init_trip(); |
|
783 Node* limit = cl->limit(); |
|
784 Node* stride = cl->stride(); |
|
785 |
|
786 // Build if's for the upper and lower bound tests. The |
|
787 // lower_bound test will dominate the upper bound test and all |
|
788 // cloned or created nodes will use the lower bound test as |
|
789 // their declared control. |
|
790 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate); |
|
791 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate); |
|
792 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate"); |
|
793 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0); |
|
794 |
|
795 // Perform cloning to keep Invariance state correct since the |
|
796 // late schedule will place invariant things in the loop. |
|
797 rng = invar.clone(rng, ctrl); |
|
798 if (offset && offset != zero) { |
|
799 assert(invar.is_invariant(offset), "offset must be loop invariant"); |
|
800 offset = invar.clone(offset, ctrl); |
|
801 } |
|
802 |
|
803 // Test the lower bound |
|
804 Node* lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false); |
|
805 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If(); |
|
806 _igvn.hash_delete(lower_bound_iff); |
|
807 lower_bound_iff->set_req(1, lower_bound_bol); |
|
808 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx); |
|
809 |
|
810 // Test the upper bound |
|
811 Node* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true); |
|
812 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If(); |
|
813 _igvn.hash_delete(upper_bound_iff); |
|
814 upper_bound_iff->set_req(1, upper_bound_bol); |
|
815 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx); |
|
816 |
|
817 // Fall through into rest of the clean up code which will move |
|
818 // any dependent nodes onto the upper bound test. |
|
819 new_predicate_proj = upper_bound_proj; |
|
820 |
|
821 #ifndef PRODUCT |
|
822 if (TraceLoopOpts && !TraceLoopPredicate) { |
|
823 tty->print("Predicate RC "); |
|
824 loop->dump_head(); |
|
825 } |
|
826 #endif |
|
827 } else { |
|
828 // Loop variant check (for example, range check in non-counted loop) |
|
829 // with uncommon trap. |
|
830 continue; |
|
831 } |
|
832 assert(new_predicate_proj != NULL, "sanity"); |
|
833 // Success - attach condition (new_predicate_bol) to predicate if |
|
834 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate |
|
835 |
|
836 // Eliminate the old If in the loop body |
|
837 dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con ); |
|
838 |
|
839 hoisted = true; |
|
840 C->set_major_progress(); |
|
841 } // end while |
|
842 |
|
843 #ifndef PRODUCT |
|
844 // report that the loop predication has been actually performed |
|
845 // for this loop |
|
846 if (TraceLoopPredicate && hoisted) { |
|
847 tty->print("Loop Predication Performed:"); |
|
848 loop->dump_head(); |
|
849 } |
|
850 #endif |
|
851 |
|
852 return hoisted; |
|
853 } |
|
854 |
|
855 //------------------------------loop_predication-------------------------------- |
|
856 // driver routine for loop predication optimization |
|
857 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) { |
|
858 bool hoisted = false; |
|
859 // Recursively promote predicates |
|
860 if (_child) { |
|
861 hoisted = _child->loop_predication( phase); |
|
862 } |
|
863 |
|
864 // self |
|
865 if (!_irreducible && !tail()->is_top()) { |
|
866 hoisted |= phase->loop_predication_impl(this); |
|
867 } |
|
868 |
|
869 if (_next) { //sibling |
|
870 hoisted |= _next->loop_predication( phase); |
|
871 } |
|
872 |
|
873 return hoisted; |
|
874 } |