1 /*
2 * Copyright (c) 2011, 2013, 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/constnode.hpp"
30 #include "opto/convertnode.hpp"
31 #include "opto/loopnode.hpp"
32 #include "opto/mulnode.hpp"
33 #include "opto/optonode.hpp"
34 #include "opto/rootnode.hpp"
35 #include "opto/subnode.hpp"
36
37 /*
38 * The general idea of Loop Predication is to insert a predicate on the entry
39 * path to a loop, and raise a uncommon trap if the check of the condition fails.
40 * The condition checks are promoted from inside the loop body, and thus
41 * the checks inside the loop could be eliminated. Currently, loop predication
42 * optimization has been applied to remove array range check and loop invariant
43 * checks (such as null checks).
44 */
45
46 //-------------------------------register_control-------------------------
47 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
48 assert(n->is_CFG(), "must be control node");
49 _igvn.register_new_node_with_optimizer(n);
50 loop->_body.push(n);
51 set_loop(n, loop);
52 // When called from beautify_loops() idom is not constructed yet.
53 if (_idom != NULL) {
54 set_idom(n, pred, dom_depth(pred));
55 }
56 }
57
58 //------------------------------create_new_if_for_predicate------------------------
59 // create a new if above the uct_if_pattern for the predicate to be promoted.
60 //
61 // before after
62 // ---------- ----------
63 // ctrl ctrl
64 // | |
65 // | |
66 // v v
67 // iff new_iff
68 // / \ / \
69 // / \ / \
70 // v v v v
71 // uncommon_proj cont_proj if_uct if_cont
72 // \ | | | |
73 // \ | | | |
74 // v v v | v
75 // rgn loop | iff
76 // | | / \
77 // | | / \
78 // v | v v
79 // uncommon_trap | uncommon_proj cont_proj
80 // \ \ | |
81 // \ \ | |
82 // v v v v
83 // rgn loop
84 // |
85 // |
86 // v
87 // uncommon_trap
88 //
89 //
90 // We will create a region to guard the uct call if there is no one there.
91 // The true projecttion (if_cont) of the new_iff is returned.
92 // This code is also used to clone predicates to clonned loops.
93 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
94 Deoptimization::DeoptReason reason) {
95 assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
96 IfNode* iff = cont_proj->in(0)->as_If();
97
98 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
99 Node *rgn = uncommon_proj->unique_ctrl_out();
100 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
101
102 uint proj_index = 1; // region's edge corresponding to uncommon_proj
103 if (!rgn->is_Region()) { // create a region to guard the call
104 assert(rgn->is_Call(), "must be call uct");
105 CallNode* call = rgn->as_Call();
106 IdealLoopTree* loop = get_loop(call);
107 rgn = new (C) RegionNode(1);
108 rgn->add_req(uncommon_proj);
109 register_control(rgn, loop, uncommon_proj);
110 _igvn.hash_delete(call);
111 call->set_req(0, rgn);
112 // When called from beautify_loops() idom is not constructed yet.
113 if (_idom != NULL) {
114 set_idom(call, rgn, dom_depth(rgn));
115 }
116 } else {
117 // Find region's edge corresponding to uncommon_proj
118 for (; proj_index < rgn->req(); proj_index++)
119 if (rgn->in(proj_index) == uncommon_proj) break;
120 assert(proj_index < rgn->req(), "sanity");
121 }
122
123 Node* entry = iff->in(0);
124 if (new_entry != NULL) {
125 // Clonning the predicate to new location.
126 entry = new_entry;
127 }
128 // Create new_iff
129 IdealLoopTree* lp = get_loop(entry);
130 IfNode *new_iff = iff->clone()->as_If();
131 new_iff->set_req(0, entry);
132 register_control(new_iff, lp, entry);
133 Node *if_cont = new (C) IfTrueNode(new_iff);
134 Node *if_uct = new (C) IfFalseNode(new_iff);
135 if (cont_proj->is_IfFalse()) {
136 // Swap
137 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
138 }
139 register_control(if_cont, lp, new_iff);
140 register_control(if_uct, get_loop(rgn), new_iff);
141
142 // if_uct to rgn
143 _igvn.hash_delete(rgn);
144 rgn->add_req(if_uct);
145 // When called from beautify_loops() idom is not constructed yet.
146 if (_idom != NULL) {
147 Node* ridom = idom(rgn);
148 Node* nrdom = dom_lca(ridom, new_iff);
149 set_idom(rgn, nrdom, dom_depth(rgn));
150 }
151
152 // If rgn has phis add new edges which has the same
153 // value as on original uncommon_proj pass.
154 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
155 bool has_phi = false;
156 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
157 Node* use = rgn->fast_out(i);
158 if (use->is_Phi() && use->outcnt() > 0) {
159 assert(use->in(0) == rgn, "");
160 _igvn.rehash_node_delayed(use);
161 use->add_req(use->in(proj_index));
162 has_phi = true;
163 }
164 }
165 assert(!has_phi || rgn->req() > 3, "no phis when region is created");
166
167 if (new_entry == NULL) {
168 // Attach if_cont to iff
169 _igvn.hash_delete(iff);
170 iff->set_req(0, if_cont);
171 if (_idom != NULL) {
172 set_idom(iff, if_cont, dom_depth(iff));
173 }
174 }
175 return if_cont->as_Proj();
176 }
177
178 //------------------------------create_new_if_for_predicate------------------------
179 // Create a new if below new_entry for the predicate to be cloned (IGVN optimization)
180 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
181 Deoptimization::DeoptReason reason) {
182 assert(new_entry != 0, "only used for clone predicate");
183 assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
184 IfNode* iff = cont_proj->in(0)->as_If();
185
186 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
187 Node *rgn = uncommon_proj->unique_ctrl_out();
188 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
189
190 uint proj_index = 1; // region's edge corresponding to uncommon_proj
191 if (!rgn->is_Region()) { // create a region to guard the call
192 assert(rgn->is_Call(), "must be call uct");
193 CallNode* call = rgn->as_Call();
194 rgn = new (C) RegionNode(1);
195 register_new_node_with_optimizer(rgn);
196 rgn->add_req(uncommon_proj);
197 hash_delete(call);
198 call->set_req(0, rgn);
199 } else {
200 // Find region's edge corresponding to uncommon_proj
201 for (; proj_index < rgn->req(); proj_index++)
202 if (rgn->in(proj_index) == uncommon_proj) break;
203 assert(proj_index < rgn->req(), "sanity");
204 }
205
206 // Create new_iff in new location.
207 IfNode *new_iff = iff->clone()->as_If();
208 new_iff->set_req(0, new_entry);
209
210 register_new_node_with_optimizer(new_iff);
211 Node *if_cont = new (C) IfTrueNode(new_iff);
212 Node *if_uct = new (C) IfFalseNode(new_iff);
213 if (cont_proj->is_IfFalse()) {
214 // Swap
215 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
216 }
217 register_new_node_with_optimizer(if_cont);
218 register_new_node_with_optimizer(if_uct);
219
220 // if_uct to rgn
221 hash_delete(rgn);
222 rgn->add_req(if_uct);
223
224 // If rgn has phis add corresponding new edges which has the same
225 // value as on original uncommon_proj pass.
226 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
227 bool has_phi = false;
228 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
229 Node* use = rgn->fast_out(i);
230 if (use->is_Phi() && use->outcnt() > 0) {
231 rehash_node_delayed(use);
232 use->add_req(use->in(proj_index));
233 has_phi = true;
234 }
235 }
236 assert(!has_phi || rgn->req() > 3, "no phis when region is created");
237
238 return if_cont->as_Proj();
239 }
240
241 //--------------------------clone_predicate-----------------------
242 ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry,
243 Deoptimization::DeoptReason reason,
244 PhaseIdealLoop* loop_phase,
245 PhaseIterGVN* igvn) {
246 ProjNode* new_predicate_proj;
247 if (loop_phase != NULL) {
248 new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason);
249 } else {
250 new_predicate_proj = igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason);
251 }
252 IfNode* iff = new_predicate_proj->in(0)->as_If();
253 Node* ctrl = iff->in(0);
254
255 // Match original condition since predicate's projections could be swapped.
256 assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
257 Node* opq = new (igvn->C) Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1));
258 igvn->C->add_predicate_opaq(opq);
259
260 Node* bol = new (igvn->C) Conv2BNode(opq);
261 if (loop_phase != NULL) {
262 loop_phase->register_new_node(opq, ctrl);
263 loop_phase->register_new_node(bol, ctrl);
264 } else {
265 igvn->register_new_node_with_optimizer(opq);
266 igvn->register_new_node_with_optimizer(bol);
267 }
268 igvn->hash_delete(iff);
269 iff->set_req(1, bol);
270 return new_predicate_proj;
271 }
272
273
274 //--------------------------clone_loop_predicates-----------------------
275 // Interface from IGVN
276 Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
277 return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, clone_limit_check, NULL, this);
278 }
279
280 // Interface from PhaseIdealLoop
281 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
282 return clone_loop_predicates(old_entry, new_entry, clone_limit_check, this, &this->_igvn);
283 }
284
285 // Clone loop predicates to cloned loops (peeled, unswitched, split_if).
286 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry,
287 bool clone_limit_check,
288 PhaseIdealLoop* loop_phase,
289 PhaseIterGVN* igvn) {
290 #ifdef ASSERT
291 if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) {
292 if (new_entry != NULL)
293 new_entry->dump();
294 assert(false, "not IfTrue, IfFalse, Region or SafePoint");
295 }
296 #endif
297 // Search original predicates
298 Node* entry = old_entry;
299 ProjNode* limit_check_proj = NULL;
300 if (LoopLimitCheck) {
301 limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
302 if (limit_check_proj != NULL) {
303 entry = entry->in(0)->in(0);
304 }
305 }
306 if (UseLoopPredicate) {
307 ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
308 if (predicate_proj != NULL) { // right pattern that can be used by loop predication
309 // clone predicate
310 new_entry = clone_predicate(predicate_proj, new_entry,
311 Deoptimization::Reason_predicate,
312 loop_phase, igvn);
313 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate");
314 if (TraceLoopPredicate) {
315 tty->print("Loop Predicate cloned: ");
316 debug_only( new_entry->in(0)->dump(); )
317 }
318 }
319 }
320 if (limit_check_proj != NULL && clone_limit_check) {
321 // Clone loop limit check last to insert it before loop.
322 // Don't clone a limit check which was already finalized
323 // for this counted loop (only one limit check is needed).
324 new_entry = clone_predicate(limit_check_proj, new_entry,
325 Deoptimization::Reason_loop_limit_check,
326 loop_phase, igvn);
327 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check");
328 if (TraceLoopLimitCheck) {
329 tty->print("Loop Limit Check cloned: ");
330 debug_only( new_entry->in(0)->dump(); )
331 }
332 }
333 return new_entry;
334 }
335
336 //--------------------------skip_loop_predicates------------------------------
337 // Skip related predicates.
338 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
339 Node* predicate = NULL;
340 if (LoopLimitCheck) {
341 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
342 if (predicate != NULL) {
343 entry = entry->in(0)->in(0);
344 }
345 }
346 if (UseLoopPredicate) {
347 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
348 if (predicate != NULL) { // right pattern that can be used by loop predication
349 IfNode* iff = entry->in(0)->as_If();
350 ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
351 Node* rgn = uncommon_proj->unique_ctrl_out();
352 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
353 entry = entry->in(0)->in(0);
354 while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
355 uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
356 if (uncommon_proj->unique_ctrl_out() != rgn)
357 break;
358 entry = entry->in(0)->in(0);
359 }
360 }
361 }
362 return entry;
363 }
364
365 //--------------------------find_predicate_insertion_point-------------------
366 // Find a good location to insert a predicate
367 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
368 if (start_c == NULL || !start_c->is_Proj())
369 return NULL;
370 if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) {
371 return start_c->as_Proj();
372 }
373 return NULL;
374 }
375
376 //--------------------------find_predicate------------------------------------
377 // Find a predicate
378 Node* PhaseIdealLoop::find_predicate(Node* entry) {
379 Node* predicate = NULL;
380 if (LoopLimitCheck) {
381 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
382 if (predicate != NULL) { // right pattern that can be used by loop predication
383 return entry;
384 }
385 }
386 if (UseLoopPredicate) {
387 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
388 if (predicate != NULL) { // right pattern that can be used by loop predication
389 return entry;
390 }
391 }
392 return NULL;
393 }
394
395 //------------------------------Invariance-----------------------------------
396 // Helper class for loop_predication_impl to compute invariance on the fly and
397 // clone invariants.
398 class Invariance : public StackObj {
399 VectorSet _visited, _invariant;
400 Node_Stack _stack;
401 VectorSet _clone_visited;
402 Node_List _old_new; // map of old to new (clone)
403 IdealLoopTree* _lpt;
404 PhaseIdealLoop* _phase;
405
406 // Helper function to set up the invariance for invariance computation
407 // If n is a known invariant, set up directly. Otherwise, look up the
408 // the possibility to push n onto the stack for further processing.
409 void visit(Node* use, Node* n) {
410 if (_lpt->is_invariant(n)) { // known invariant
411 _invariant.set(n->_idx);
412 } else if (!n->is_CFG()) {
413 Node *n_ctrl = _phase->ctrl_or_self(n);
414 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
415 if (_phase->is_dominator(n_ctrl, u_ctrl)) {
416 _stack.push(n, n->in(0) == NULL ? 1 : 0);
417 }
418 }
419 }
420
421 // Compute invariance for "the_node" and (possibly) all its inputs recursively
422 // on the fly
423 void compute_invariance(Node* n) {
424 assert(_visited.test(n->_idx), "must be");
425 visit(n, n);
426 while (_stack.is_nonempty()) {
427 Node* n = _stack.node();
428 uint idx = _stack.index();
429 if (idx == n->req()) { // all inputs are processed
430 _stack.pop();
431 // n is invariant if it's inputs are all invariant
432 bool all_inputs_invariant = true;
433 for (uint i = 0; i < n->req(); i++) {
434 Node* in = n->in(i);
435 if (in == NULL) continue;
436 assert(_visited.test(in->_idx), "must have visited input");
437 if (!_invariant.test(in->_idx)) { // bad guy
438 all_inputs_invariant = false;
439 break;
440 }
441 }
442 if (all_inputs_invariant) {
443 _invariant.set(n->_idx); // I am a invariant too
444 }
445 } else { // process next input
446 _stack.set_index(idx + 1);
447 Node* m = n->in(idx);
448 if (m != NULL && !_visited.test_set(m->_idx)) {
449 visit(n, m);
450 }
451 }
452 }
453 }
454
455 // Helper function to set up _old_new map for clone_nodes.
456 // If n is a known invariant, set up directly ("clone" of n == n).
457 // Otherwise, push n onto the stack for real cloning.
458 void clone_visit(Node* n) {
459 assert(_invariant.test(n->_idx), "must be invariant");
460 if (_lpt->is_invariant(n)) { // known invariant
461 _old_new.map(n->_idx, n);
462 } else { // to be cloned
463 assert(!n->is_CFG(), "should not see CFG here");
464 _stack.push(n, n->in(0) == NULL ? 1 : 0);
465 }
466 }
467
468 // Clone "n" and (possibly) all its inputs recursively
469 void clone_nodes(Node* n, Node* ctrl) {
470 clone_visit(n);
471 while (_stack.is_nonempty()) {
472 Node* n = _stack.node();
473 uint idx = _stack.index();
474 if (idx == n->req()) { // all inputs processed, clone n!
475 _stack.pop();
476 // clone invariant node
477 Node* n_cl = n->clone();
478 _old_new.map(n->_idx, n_cl);
479 _phase->register_new_node(n_cl, ctrl);
480 for (uint i = 0; i < n->req(); i++) {
481 Node* in = n_cl->in(i);
482 if (in == NULL) continue;
483 n_cl->set_req(i, _old_new[in->_idx]);
484 }
485 } else { // process next input
486 _stack.set_index(idx + 1);
487 Node* m = n->in(idx);
488 if (m != NULL && !_clone_visited.test_set(m->_idx)) {
489 clone_visit(m); // visit the input
490 }
491 }
492 }
493 }
494
495 public:
496 Invariance(Arena* area, IdealLoopTree* lpt) :
497 _lpt(lpt), _phase(lpt->_phase),
498 _visited(area), _invariant(area), _stack(area, 10 /* guess */),
499 _clone_visited(area), _old_new(area)
500 {}
501
502 // Map old to n for invariance computation and clone
503 void map_ctrl(Node* old, Node* n) {
504 assert(old->is_CFG() && n->is_CFG(), "must be");
505 _old_new.map(old->_idx, n); // "clone" of old is n
506 _invariant.set(old->_idx); // old is invariant
507 _clone_visited.set(old->_idx);
508 }
509
510 // Driver function to compute invariance
511 bool is_invariant(Node* n) {
512 if (!_visited.test_set(n->_idx))
513 compute_invariance(n);
514 return (_invariant.test(n->_idx) != 0);
515 }
516
517 // Driver function to clone invariant
518 Node* clone(Node* n, Node* ctrl) {
519 assert(ctrl->is_CFG(), "must be");
520 assert(_invariant.test(n->_idx), "must be an invariant");
521 if (!_clone_visited.test(n->_idx))
522 clone_nodes(n, ctrl);
523 return _old_new[n->_idx];
524 }
525 };
526
527 //------------------------------is_range_check_if -----------------------------------
528 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
529 // Note: this function is particularly designed for loop predication. We require load_range
530 // and offset to be loop invariant computed on the fly by "invar"
531 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
532 if (!is_loop_exit(iff)) {
533 return false;
534 }
535 if (!iff->in(1)->is_Bool()) {
536 return false;
537 }
538 const BoolNode *bol = iff->in(1)->as_Bool();
539 if (bol->_test._test != BoolTest::lt) {
540 return false;
541 }
542 if (!bol->in(1)->is_Cmp()) {
543 return false;
544 }
545 const CmpNode *cmp = bol->in(1)->as_Cmp();
546 if (cmp->Opcode() != Op_CmpU) {
547 return false;
548 }
549 Node* range = cmp->in(2);
550 if (range->Opcode() != Op_LoadRange) {
551 const TypeInt* tint = phase->_igvn.type(range)->isa_int();
552 if (tint == NULL || tint->empty() || tint->_lo < 0) {
553 // Allow predication on positive values that aren't LoadRanges.
554 // This allows optimization of loops where the length of the
555 // array is a known value and doesn't need to be loaded back
556 // from the array.
557 return false;
558 }
559 }
560 if (!invar.is_invariant(range)) {
561 return false;
562 }
563 Node *iv = _head->as_CountedLoop()->phi();
564 int scale = 0;
565 Node *offset = NULL;
566 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
567 return false;
568 }
569 if (offset && !invar.is_invariant(offset)) { // offset must be invariant
570 return false;
571 }
572 return true;
573 }
574
575 //------------------------------rc_predicate-----------------------------------
576 // Create a range check predicate
577 //
578 // for (i = init; i < limit; i += stride) {
579 // a[scale*i+offset]
580 // }
581 //
582 // Compute max(scale*i + offset) for init <= i < limit and build the predicate
583 // as "max(scale*i + offset) u< a.length".
584 //
585 // There are two cases for max(scale*i + offset):
586 // (1) stride*scale > 0
587 // max(scale*i + offset) = scale*(limit-stride) + offset
588 // (2) stride*scale < 0
589 // max(scale*i + offset) = scale*init + offset
590 BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl,
591 int scale, Node* offset,
592 Node* init, Node* limit, Node* stride,
593 Node* range, bool upper) {
594 stringStream* predString = NULL;
595 if (TraceLoopPredicate) {
596 predString = new stringStream();
597 predString->print("rc_predicate ");
598 }
599
600 Node* max_idx_expr = init;
601 int stride_con = stride->get_int();
602 if ((stride_con > 0) == (scale > 0) == upper) {
603 if (LoopLimitCheck) {
604 // With LoopLimitCheck limit is not exact.
605 // Calculate exact limit here.
606 // Note, counted loop's test is '<' or '>'.
607 limit = exact_limit(loop);
608 max_idx_expr = new (C) SubINode(limit, stride);
609 register_new_node(max_idx_expr, ctrl);
610 if (TraceLoopPredicate) predString->print("(limit - stride) ");
611 } else {
612 max_idx_expr = new (C) SubINode(limit, stride);
613 register_new_node(max_idx_expr, ctrl);
614 if (TraceLoopPredicate) predString->print("(limit - stride) ");
615 }
616 } else {
617 if (TraceLoopPredicate) predString->print("init ");
618 }
619
620 if (scale != 1) {
621 ConNode* con_scale = _igvn.intcon(scale);
622 max_idx_expr = new (C) MulINode(max_idx_expr, con_scale);
623 register_new_node(max_idx_expr, ctrl);
624 if (TraceLoopPredicate) predString->print("* %d ", scale);
625 }
626
627 if (offset && (!offset->is_Con() || offset->get_int() != 0)){
628 max_idx_expr = new (C) AddINode(max_idx_expr, offset);
629 register_new_node(max_idx_expr, ctrl);
630 if (TraceLoopPredicate)
631 if (offset->is_Con()) predString->print("+ %d ", offset->get_int());
632 else predString->print("+ offset ");
633 }
634
635 CmpUNode* cmp = new (C) CmpUNode(max_idx_expr, range);
636 register_new_node(cmp, ctrl);
637 BoolNode* bol = new (C) BoolNode(cmp, BoolTest::lt);
638 register_new_node(bol, ctrl);
639
640 if (TraceLoopPredicate) {
641 predString->print_cr("<u range");
642 tty->print(predString->as_string());
643 }
644 return bol;
645 }
646
647 //------------------------------ loop_predication_impl--------------------------
648 // Insert loop predicates for null checks and range checks
649 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
650 if (!UseLoopPredicate) return false;
651
652 if (!loop->_head->is_Loop()) {
653 // Could be a simple region when irreducible loops are present.
654 return false;
655 }
656 LoopNode* head = loop->_head->as_Loop();
657
658 if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
659 // do nothing for infinite loops
660 return false;
661 }
662
663 CountedLoopNode *cl = NULL;
664 if (head->is_valid_counted_loop()) {
665 cl = head->as_CountedLoop();
666 // do nothing for iteration-splitted loops
667 if (!cl->is_normal_loop()) return false;
668 // Avoid RCE if Counted loop's test is '!='.
669 BoolTest::mask bt = cl->loopexit()->test_trip();
670 if (bt != BoolTest::lt && bt != BoolTest::gt)
671 cl = NULL;
672 }
673
674 Node* entry = head->in(LoopNode::EntryControl);
675 ProjNode *predicate_proj = NULL;
676 // Loop limit check predicate should be near the loop.
677 if (LoopLimitCheck) {
678 predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
679 if (predicate_proj != NULL)
680 entry = predicate_proj->in(0)->in(0);
681 }
682
683 predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
684 if (!predicate_proj) {
685 #ifndef PRODUCT
686 if (TraceLoopPredicate) {
687 tty->print("missing predicate:");
688 loop->dump_head();
689 head->dump(1);
690 }
691 #endif
692 return false;
693 }
694 ConNode* zero = _igvn.intcon(0);
695 set_ctrl(zero, C->root());
696
697 ResourceArea *area = Thread::current()->resource_area();
698 Invariance invar(area, loop);
699
700 // Create list of if-projs such that a newer proj dominates all older
701 // projs in the list, and they all dominate loop->tail()
702 Node_List if_proj_list(area);
703 Node *current_proj = loop->tail(); //start from tail
704 while (current_proj != head) {
705 if (loop == get_loop(current_proj) && // still in the loop ?
706 current_proj->is_Proj() && // is a projection ?
707 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
708 if_proj_list.push(current_proj);
709 }
710 current_proj = idom(current_proj);
711 }
712
713 bool hoisted = false; // true if at least one proj is promoted
714 while (if_proj_list.size() > 0) {
715 // Following are changed to nonnull when a predicate can be hoisted
716 ProjNode* new_predicate_proj = NULL;
717
718 ProjNode* proj = if_proj_list.pop()->as_Proj();
719 IfNode* iff = proj->in(0)->as_If();
720
721 if (!proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
722 if (loop->is_loop_exit(iff)) {
723 // stop processing the remaining projs in the list because the execution of them
724 // depends on the condition of "iff" (iff->in(1)).
725 break;
726 } else {
727 // Both arms are inside the loop. There are two cases:
728 // (1) there is one backward branch. In this case, any remaining proj
729 // in the if_proj list post-dominates "iff". So, the condition of "iff"
730 // does not determine the execution the remining projs directly, and we
731 // can safely continue.
732 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
733 // does not dominate loop->tail(), so it can not be in the if_proj list.
734 continue;
735 }
736 }
737
738 Node* test = iff->in(1);
739 if (!test->is_Bool()){ //Conv2B, ...
740 continue;
741 }
742 BoolNode* bol = test->as_Bool();
743 if (invar.is_invariant(bol)) {
744 // Invariant test
745 new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
746 Deoptimization::Reason_predicate);
747 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
748 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
749
750 // Negate test if necessary
751 bool negated = false;
752 if (proj->_con != predicate_proj->_con) {
753 new_predicate_bol = new (C) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
754 register_new_node(new_predicate_bol, ctrl);
755 negated = true;
756 }
757 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
758 _igvn.hash_delete(new_predicate_iff);
759 new_predicate_iff->set_req(1, new_predicate_bol);
760 #ifndef PRODUCT
761 if (TraceLoopPredicate) {
762 tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
763 loop->dump_head();
764 } else if (TraceLoopOpts) {
765 tty->print("Predicate IC ");
766 loop->dump_head();
767 }
768 #endif
769 } else if ((cl != NULL) && (proj->_con == predicate_proj->_con) &&
770 loop->is_range_check_if(iff, this, invar)) {
771
772 // Range check for counted loops
773 const Node* cmp = bol->in(1)->as_Cmp();
774 Node* idx = cmp->in(1);
775 assert(!invar.is_invariant(idx), "index is variant");
776 Node* rng = cmp->in(2);
777 assert(rng->Opcode() == Op_LoadRange || _igvn.type(rng)->is_int() >= 0, "must be");
778 assert(invar.is_invariant(rng), "range must be invariant");
779 int scale = 1;
780 Node* offset = zero;
781 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
782 assert(ok, "must be index expression");
783
784 Node* init = cl->init_trip();
785 Node* limit = cl->limit();
786 Node* stride = cl->stride();
787
788 // Build if's for the upper and lower bound tests. The
789 // lower_bound test will dominate the upper bound test and all
790 // cloned or created nodes will use the lower bound test as
791 // their declared control.
792 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
793 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
794 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
795 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0);
796
797 // Perform cloning to keep Invariance state correct since the
798 // late schedule will place invariant things in the loop.
799 rng = invar.clone(rng, ctrl);
800 if (offset && offset != zero) {
801 assert(invar.is_invariant(offset), "offset must be loop invariant");
802 offset = invar.clone(offset, ctrl);
803 }
804
805 // Test the lower bound
806 Node* lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false);
807 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
808 _igvn.hash_delete(lower_bound_iff);
809 lower_bound_iff->set_req(1, lower_bound_bol);
810 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx);
811
812 // Test the upper bound
813 Node* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true);
814 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
815 _igvn.hash_delete(upper_bound_iff);
816 upper_bound_iff->set_req(1, upper_bound_bol);
817 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx);
818
819 // Fall through into rest of the clean up code which will move
820 // any dependent nodes onto the upper bound test.
821 new_predicate_proj = upper_bound_proj;
822
823 #ifndef PRODUCT
824 if (TraceLoopOpts && !TraceLoopPredicate) {
825 tty->print("Predicate RC ");
826 loop->dump_head();
827 }
828 #endif
829 } else {
830 // Loop variant check (for example, range check in non-counted loop)
831 // with uncommon trap.
832 continue;
833 }
834 assert(new_predicate_proj != NULL, "sanity");
835 // Success - attach condition (new_predicate_bol) to predicate if
836 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
837
838 // Eliminate the old If in the loop body
839 dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
840
841 hoisted = true;
842 C->set_major_progress();
843 } // end while
844
845 #ifndef PRODUCT
846 // report that the loop predication has been actually performed
847 // for this loop
848 if (TraceLoopPredicate && hoisted) {
849 tty->print("Loop Predication Performed:");
850 loop->dump_head();
851 }
852 #endif
853
854 return hoisted;
855 }
856
857 //------------------------------loop_predication--------------------------------
858 // driver routine for loop predication optimization
859 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
860 bool hoisted = false;
861 // Recursively promote predicates
862 if (_child) {
863 hoisted = _child->loop_predication( phase);
864 }
865
866 // self
867 if (!_irreducible && !tail()->is_top()) {
868 hoisted |= phase->loop_predication_impl(this);
869 }
870
871 if (_next) { //sibling
872 hoisted |= _next->loop_predication( phase);
873 }
874
875 return hoisted;
876 }