Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
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 */
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"
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 */
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 }
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();
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 }
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 }
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 }
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();
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");
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 }
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);
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 }
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.rehash_node_delayed(use);
216 use->add_req(use->in(proj_index));
217 has_phi = true;
218 }
219 }
220 assert(!has_phi || rgn->req() > 3, "no phis when region is created");
222 if (new_entry == NULL) {
223 // Attach if_cont to iff
224 _igvn.hash_delete(iff);
225 iff->set_req(0, if_cont);
226 if (_idom != NULL) {
227 set_idom(iff, if_cont, dom_depth(iff));
228 }
229 }
230 return if_cont->as_Proj();
231 }
233 //------------------------------create_new_if_for_predicate------------------------
234 // Create a new if below new_entry for the predicate to be cloned (IGVN optimization)
235 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
236 Deoptimization::DeoptReason reason) {
237 assert(new_entry != 0, "only used for clone predicate");
238 assert(PhaseIdealLoop::is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
239 IfNode* iff = cont_proj->in(0)->as_If();
241 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
242 Node *rgn = uncommon_proj->unique_ctrl_out();
243 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
245 uint proj_index = 1; // region's edge corresponding to uncommon_proj
246 if (!rgn->is_Region()) { // create a region to guard the call
247 assert(rgn->is_Call(), "must be call uct");
248 CallNode* call = rgn->as_Call();
249 rgn = new (C, 1) RegionNode(1);
250 register_new_node_with_optimizer(rgn);
251 rgn->add_req(uncommon_proj);
252 hash_delete(call);
253 call->set_req(0, rgn);
254 } else {
255 // Find region's edge corresponding to uncommon_proj
256 for (; proj_index < rgn->req(); proj_index++)
257 if (rgn->in(proj_index) == uncommon_proj) break;
258 assert(proj_index < rgn->req(), "sanity");
259 }
261 // Create new_iff in new location.
262 IfNode *new_iff = iff->clone()->as_If();
263 new_iff->set_req(0, new_entry);
265 register_new_node_with_optimizer(new_iff);
266 Node *if_cont = new (C, 1) IfTrueNode(new_iff);
267 Node *if_uct = new (C, 1) IfFalseNode(new_iff);
268 if (cont_proj->is_IfFalse()) {
269 // Swap
270 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
271 }
272 register_new_node_with_optimizer(if_cont);
273 register_new_node_with_optimizer(if_uct);
275 // if_uct to rgn
276 hash_delete(rgn);
277 rgn->add_req(if_uct);
279 // If rgn has phis add corresponding new edges which has the same
280 // value as on original uncommon_proj pass.
281 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
282 bool has_phi = false;
283 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
284 Node* use = rgn->fast_out(i);
285 if (use->is_Phi() && use->outcnt() > 0) {
286 rehash_node_delayed(use);
287 use->add_req(use->in(proj_index));
288 has_phi = true;
289 }
290 }
291 assert(!has_phi || rgn->req() > 3, "no phis when region is created");
293 return if_cont->as_Proj();
294 }
296 //--------------------------clone_predicate-----------------------
297 ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry,
298 Deoptimization::DeoptReason reason,
299 PhaseIdealLoop* loop_phase,
300 PhaseIterGVN* igvn) {
301 ProjNode* new_predicate_proj;
302 if (loop_phase != NULL) {
303 new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason);
304 } else {
305 new_predicate_proj = igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason);
306 }
307 IfNode* iff = new_predicate_proj->in(0)->as_If();
308 Node* ctrl = iff->in(0);
310 // Match original condition since predicate's projections could be swapped.
311 assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
312 Node* opq = new (igvn->C, 2) Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1));
313 igvn->C->add_predicate_opaq(opq);
315 Node* bol = new (igvn->C, 2) Conv2BNode(opq);
316 if (loop_phase != NULL) {
317 loop_phase->register_new_node(opq, ctrl);
318 loop_phase->register_new_node(bol, ctrl);
319 } else {
320 igvn->register_new_node_with_optimizer(opq);
321 igvn->register_new_node_with_optimizer(bol);
322 }
323 igvn->hash_delete(iff);
324 iff->set_req(1, bol);
325 return new_predicate_proj;
326 }
329 //--------------------------clone_loop_predicates-----------------------
330 // Interface from IGVN
331 Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
332 return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, clone_limit_check, NULL, this);
333 }
335 // Interface from PhaseIdealLoop
336 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
337 return clone_loop_predicates(old_entry, new_entry, clone_limit_check, this, &this->_igvn);
338 }
340 // Clone loop predicates to cloned loops (peeled, unswitched, split_if).
341 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry,
342 bool clone_limit_check,
343 PhaseIdealLoop* loop_phase,
344 PhaseIterGVN* igvn) {
345 #ifdef ASSERT
346 if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) {
347 if (new_entry != NULL)
348 new_entry->dump();
349 assert(false, "not IfTrue, IfFalse, Region or SafePoint");
350 }
351 #endif
352 // Search original predicates
353 Node* entry = old_entry;
354 ProjNode* limit_check_proj = NULL;
355 if (LoopLimitCheck) {
356 limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
357 if (limit_check_proj != NULL) {
358 entry = entry->in(0)->in(0);
359 }
360 }
361 if (UseLoopPredicate) {
362 ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
363 if (predicate_proj != NULL) { // right pattern that can be used by loop predication
364 // clone predicate
365 new_entry = clone_predicate(predicate_proj, new_entry,
366 Deoptimization::Reason_predicate,
367 loop_phase, igvn);
368 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate");
369 if (TraceLoopPredicate) {
370 tty->print("Loop Predicate cloned: ");
371 debug_only( new_entry->in(0)->dump(); )
372 }
373 }
374 }
375 if (limit_check_proj != NULL && clone_limit_check) {
376 // Clone loop limit check last to insert it before loop.
377 // Don't clone a limit check which was already finalized
378 // for this counted loop (only one limit check is needed).
379 new_entry = clone_predicate(limit_check_proj, new_entry,
380 Deoptimization::Reason_loop_limit_check,
381 loop_phase, igvn);
382 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check");
383 if (TraceLoopLimitCheck) {
384 tty->print("Loop Limit Check cloned: ");
385 debug_only( new_entry->in(0)->dump(); )
386 }
387 }
388 return new_entry;
389 }
391 //--------------------------skip_loop_predicates------------------------------
392 // Skip related predicates.
393 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
394 Node* predicate = NULL;
395 if (LoopLimitCheck) {
396 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
397 if (predicate != NULL) {
398 entry = entry->in(0)->in(0);
399 }
400 }
401 if (UseLoopPredicate) {
402 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
403 if (predicate != NULL) { // right pattern that can be used by loop predication
404 IfNode* iff = entry->in(0)->as_If();
405 ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
406 Node* rgn = uncommon_proj->unique_ctrl_out();
407 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
408 entry = entry->in(0)->in(0);
409 while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
410 uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
411 if (uncommon_proj->unique_ctrl_out() != rgn)
412 break;
413 entry = entry->in(0)->in(0);
414 }
415 }
416 }
417 return entry;
418 }
420 //--------------------------find_predicate_insertion_point-------------------
421 // Find a good location to insert a predicate
422 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
423 if (start_c == NULL || !start_c->is_Proj())
424 return NULL;
425 if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) {
426 return start_c->as_Proj();
427 }
428 return NULL;
429 }
431 //--------------------------find_predicate------------------------------------
432 // Find a predicate
433 Node* PhaseIdealLoop::find_predicate(Node* entry) {
434 Node* predicate = NULL;
435 if (LoopLimitCheck) {
436 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
437 if (predicate != NULL) { // right pattern that can be used by loop predication
438 return entry;
439 }
440 }
441 if (UseLoopPredicate) {
442 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
443 if (predicate != NULL) { // right pattern that can be used by loop predication
444 return entry;
445 }
446 }
447 return NULL;
448 }
450 //------------------------------Invariance-----------------------------------
451 // Helper class for loop_predication_impl to compute invariance on the fly and
452 // clone invariants.
453 class Invariance : public StackObj {
454 VectorSet _visited, _invariant;
455 Node_Stack _stack;
456 VectorSet _clone_visited;
457 Node_List _old_new; // map of old to new (clone)
458 IdealLoopTree* _lpt;
459 PhaseIdealLoop* _phase;
461 // Helper function to set up the invariance for invariance computation
462 // If n is a known invariant, set up directly. Otherwise, look up the
463 // the possibility to push n onto the stack for further processing.
464 void visit(Node* use, Node* n) {
465 if (_lpt->is_invariant(n)) { // known invariant
466 _invariant.set(n->_idx);
467 } else if (!n->is_CFG()) {
468 Node *n_ctrl = _phase->ctrl_or_self(n);
469 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
470 if (_phase->is_dominator(n_ctrl, u_ctrl)) {
471 _stack.push(n, n->in(0) == NULL ? 1 : 0);
472 }
473 }
474 }
476 // Compute invariance for "the_node" and (possibly) all its inputs recursively
477 // on the fly
478 void compute_invariance(Node* n) {
479 assert(_visited.test(n->_idx), "must be");
480 visit(n, n);
481 while (_stack.is_nonempty()) {
482 Node* n = _stack.node();
483 uint idx = _stack.index();
484 if (idx == n->req()) { // all inputs are processed
485 _stack.pop();
486 // n is invariant if it's inputs are all invariant
487 bool all_inputs_invariant = true;
488 for (uint i = 0; i < n->req(); i++) {
489 Node* in = n->in(i);
490 if (in == NULL) continue;
491 assert(_visited.test(in->_idx), "must have visited input");
492 if (!_invariant.test(in->_idx)) { // bad guy
493 all_inputs_invariant = false;
494 break;
495 }
496 }
497 if (all_inputs_invariant) {
498 _invariant.set(n->_idx); // I am a invariant too
499 }
500 } else { // process next input
501 _stack.set_index(idx + 1);
502 Node* m = n->in(idx);
503 if (m != NULL && !_visited.test_set(m->_idx)) {
504 visit(n, m);
505 }
506 }
507 }
508 }
510 // Helper function to set up _old_new map for clone_nodes.
511 // If n is a known invariant, set up directly ("clone" of n == n).
512 // Otherwise, push n onto the stack for real cloning.
513 void clone_visit(Node* n) {
514 assert(_invariant.test(n->_idx), "must be invariant");
515 if (_lpt->is_invariant(n)) { // known invariant
516 _old_new.map(n->_idx, n);
517 } else { // to be cloned
518 assert(!n->is_CFG(), "should not see CFG here");
519 _stack.push(n, n->in(0) == NULL ? 1 : 0);
520 }
521 }
523 // Clone "n" and (possibly) all its inputs recursively
524 void clone_nodes(Node* n, Node* ctrl) {
525 clone_visit(n);
526 while (_stack.is_nonempty()) {
527 Node* n = _stack.node();
528 uint idx = _stack.index();
529 if (idx == n->req()) { // all inputs processed, clone n!
530 _stack.pop();
531 // clone invariant node
532 Node* n_cl = n->clone();
533 _old_new.map(n->_idx, n_cl);
534 _phase->register_new_node(n_cl, ctrl);
535 for (uint i = 0; i < n->req(); i++) {
536 Node* in = n_cl->in(i);
537 if (in == NULL) continue;
538 n_cl->set_req(i, _old_new[in->_idx]);
539 }
540 } else { // process next input
541 _stack.set_index(idx + 1);
542 Node* m = n->in(idx);
543 if (m != NULL && !_clone_visited.test_set(m->_idx)) {
544 clone_visit(m); // visit the input
545 }
546 }
547 }
548 }
550 public:
551 Invariance(Arena* area, IdealLoopTree* lpt) :
552 _lpt(lpt), _phase(lpt->_phase),
553 _visited(area), _invariant(area), _stack(area, 10 /* guess */),
554 _clone_visited(area), _old_new(area)
555 {}
557 // Map old to n for invariance computation and clone
558 void map_ctrl(Node* old, Node* n) {
559 assert(old->is_CFG() && n->is_CFG(), "must be");
560 _old_new.map(old->_idx, n); // "clone" of old is n
561 _invariant.set(old->_idx); // old is invariant
562 _clone_visited.set(old->_idx);
563 }
565 // Driver function to compute invariance
566 bool is_invariant(Node* n) {
567 if (!_visited.test_set(n->_idx))
568 compute_invariance(n);
569 return (_invariant.test(n->_idx) != 0);
570 }
572 // Driver function to clone invariant
573 Node* clone(Node* n, Node* ctrl) {
574 assert(ctrl->is_CFG(), "must be");
575 assert(_invariant.test(n->_idx), "must be an invariant");
576 if (!_clone_visited.test(n->_idx))
577 clone_nodes(n, ctrl);
578 return _old_new[n->_idx];
579 }
580 };
582 //------------------------------is_range_check_if -----------------------------------
583 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
584 // Note: this function is particularly designed for loop predication. We require load_range
585 // and offset to be loop invariant computed on the fly by "invar"
586 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
587 if (!is_loop_exit(iff)) {
588 return false;
589 }
590 if (!iff->in(1)->is_Bool()) {
591 return false;
592 }
593 const BoolNode *bol = iff->in(1)->as_Bool();
594 if (bol->_test._test != BoolTest::lt) {
595 return false;
596 }
597 if (!bol->in(1)->is_Cmp()) {
598 return false;
599 }
600 const CmpNode *cmp = bol->in(1)->as_Cmp();
601 if (cmp->Opcode() != Op_CmpU) {
602 return false;
603 }
604 Node* range = cmp->in(2);
605 if (range->Opcode() != Op_LoadRange) {
606 const TypeInt* tint = phase->_igvn.type(range)->isa_int();
607 if (tint == NULL || tint->empty() || tint->_lo < 0) {
608 // Allow predication on positive values that aren't LoadRanges.
609 // This allows optimization of loops where the length of the
610 // array is a known value and doesn't need to be loaded back
611 // from the array.
612 return false;
613 }
614 }
615 if (!invar.is_invariant(range)) {
616 return false;
617 }
618 Node *iv = _head->as_CountedLoop()->phi();
619 int scale = 0;
620 Node *offset = NULL;
621 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
622 return false;
623 }
624 if (offset && !invar.is_invariant(offset)) { // offset must be invariant
625 return false;
626 }
627 return true;
628 }
630 //------------------------------rc_predicate-----------------------------------
631 // Create a range check predicate
632 //
633 // for (i = init; i < limit; i += stride) {
634 // a[scale*i+offset]
635 // }
636 //
637 // Compute max(scale*i + offset) for init <= i < limit and build the predicate
638 // as "max(scale*i + offset) u< a.length".
639 //
640 // There are two cases for max(scale*i + offset):
641 // (1) stride*scale > 0
642 // max(scale*i + offset) = scale*(limit-stride) + offset
643 // (2) stride*scale < 0
644 // max(scale*i + offset) = scale*init + offset
645 BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl,
646 int scale, Node* offset,
647 Node* init, Node* limit, Node* stride,
648 Node* range, bool upper) {
649 stringStream* predString = NULL;
650 if (TraceLoopPredicate) {
651 predString = new stringStream();
652 predString->print("rc_predicate ");
653 }
655 Node* max_idx_expr = init;
656 int stride_con = stride->get_int();
657 if ((stride_con > 0) == (scale > 0) == upper) {
658 if (LoopLimitCheck) {
659 // With LoopLimitCheck limit is not exact.
660 // Calculate exact limit here.
661 // Note, counted loop's test is '<' or '>'.
662 limit = exact_limit(loop);
663 max_idx_expr = new (C, 3) SubINode(limit, stride);
664 register_new_node(max_idx_expr, ctrl);
665 if (TraceLoopPredicate) predString->print("(limit - stride) ");
666 } else {
667 max_idx_expr = new (C, 3) SubINode(limit, stride);
668 register_new_node(max_idx_expr, ctrl);
669 if (TraceLoopPredicate) predString->print("(limit - stride) ");
670 }
671 } else {
672 if (TraceLoopPredicate) predString->print("init ");
673 }
675 if (scale != 1) {
676 ConNode* con_scale = _igvn.intcon(scale);
677 max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale);
678 register_new_node(max_idx_expr, ctrl);
679 if (TraceLoopPredicate) predString->print("* %d ", scale);
680 }
682 if (offset && (!offset->is_Con() || offset->get_int() != 0)){
683 max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset);
684 register_new_node(max_idx_expr, ctrl);
685 if (TraceLoopPredicate)
686 if (offset->is_Con()) predString->print("+ %d ", offset->get_int());
687 else predString->print("+ offset ");
688 }
690 CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range);
691 register_new_node(cmp, ctrl);
692 BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt);
693 register_new_node(bol, ctrl);
695 if (TraceLoopPredicate) {
696 predString->print_cr("<u range");
697 tty->print(predString->as_string());
698 }
699 return bol;
700 }
702 //------------------------------ loop_predication_impl--------------------------
703 // Insert loop predicates for null checks and range checks
704 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
705 if (!UseLoopPredicate) return false;
707 if (!loop->_head->is_Loop()) {
708 // Could be a simple region when irreducible loops are present.
709 return false;
710 }
711 LoopNode* head = loop->_head->as_Loop();
713 if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
714 // do nothing for infinite loops
715 return false;
716 }
718 CountedLoopNode *cl = NULL;
719 if (head->is_valid_counted_loop()) {
720 cl = head->as_CountedLoop();
721 // do nothing for iteration-splitted loops
722 if (!cl->is_normal_loop()) return false;
723 // Avoid RCE if Counted loop's test is '!='.
724 BoolTest::mask bt = cl->loopexit()->test_trip();
725 if (bt != BoolTest::lt && bt != BoolTest::gt)
726 cl = NULL;
727 }
729 Node* entry = head->in(LoopNode::EntryControl);
730 ProjNode *predicate_proj = NULL;
731 // Loop limit check predicate should be near the loop.
732 if (LoopLimitCheck) {
733 predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
734 if (predicate_proj != NULL)
735 entry = predicate_proj->in(0)->in(0);
736 }
738 predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
739 if (!predicate_proj) {
740 #ifndef PRODUCT
741 if (TraceLoopPredicate) {
742 tty->print("missing predicate:");
743 loop->dump_head();
744 head->dump(1);
745 }
746 #endif
747 return false;
748 }
749 ConNode* zero = _igvn.intcon(0);
750 set_ctrl(zero, C->root());
752 ResourceArea *area = Thread::current()->resource_area();
753 Invariance invar(area, loop);
755 // Create list of if-projs such that a newer proj dominates all older
756 // projs in the list, and they all dominate loop->tail()
757 Node_List if_proj_list(area);
758 Node *current_proj = loop->tail(); //start from tail
759 while (current_proj != head) {
760 if (loop == get_loop(current_proj) && // still in the loop ?
761 current_proj->is_Proj() && // is a projection ?
762 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
763 if_proj_list.push(current_proj);
764 }
765 current_proj = idom(current_proj);
766 }
768 bool hoisted = false; // true if at least one proj is promoted
769 while (if_proj_list.size() > 0) {
770 // Following are changed to nonnull when a predicate can be hoisted
771 ProjNode* new_predicate_proj = NULL;
773 ProjNode* proj = if_proj_list.pop()->as_Proj();
774 IfNode* iff = proj->in(0)->as_If();
776 if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) {
777 if (loop->is_loop_exit(iff)) {
778 // stop processing the remaining projs in the list because the execution of them
779 // depends on the condition of "iff" (iff->in(1)).
780 break;
781 } else {
782 // Both arms are inside the loop. There are two cases:
783 // (1) there is one backward branch. In this case, any remaining proj
784 // in the if_proj list post-dominates "iff". So, the condition of "iff"
785 // does not determine the execution the remining projs directly, and we
786 // can safely continue.
787 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
788 // does not dominate loop->tail(), so it can not be in the if_proj list.
789 continue;
790 }
791 }
793 Node* test = iff->in(1);
794 if (!test->is_Bool()){ //Conv2B, ...
795 continue;
796 }
797 BoolNode* bol = test->as_Bool();
798 if (invar.is_invariant(bol)) {
799 // Invariant test
800 new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
801 Deoptimization::Reason_predicate);
802 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
803 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
805 // Negate test if necessary
806 bool negated = false;
807 if (proj->_con != predicate_proj->_con) {
808 new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
809 register_new_node(new_predicate_bol, ctrl);
810 negated = true;
811 }
812 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
813 _igvn.hash_delete(new_predicate_iff);
814 new_predicate_iff->set_req(1, new_predicate_bol);
815 #ifndef PRODUCT
816 if (TraceLoopPredicate) {
817 tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
818 loop->dump_head();
819 } else if (TraceLoopOpts) {
820 tty->print("Predicate IC ");
821 loop->dump_head();
822 }
823 #endif
824 } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
825 assert(proj->_con == predicate_proj->_con, "must match");
827 // Range check for counted loops
828 const Node* cmp = bol->in(1)->as_Cmp();
829 Node* idx = cmp->in(1);
830 assert(!invar.is_invariant(idx), "index is variant");
831 Node* rng = cmp->in(2);
832 assert(rng->Opcode() == Op_LoadRange || _igvn.type(rng)->is_int() >= 0, "must be");
833 assert(invar.is_invariant(rng), "range must be invariant");
834 int scale = 1;
835 Node* offset = zero;
836 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
837 assert(ok, "must be index expression");
839 Node* init = cl->init_trip();
840 Node* limit = cl->limit();
841 Node* stride = cl->stride();
843 // Build if's for the upper and lower bound tests. The
844 // lower_bound test will dominate the upper bound test and all
845 // cloned or created nodes will use the lower bound test as
846 // their declared control.
847 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
848 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
849 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
850 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0);
852 // Perform cloning to keep Invariance state correct since the
853 // late schedule will place invariant things in the loop.
854 rng = invar.clone(rng, ctrl);
855 if (offset && offset != zero) {
856 assert(invar.is_invariant(offset), "offset must be loop invariant");
857 offset = invar.clone(offset, ctrl);
858 }
860 // Test the lower bound
861 Node* lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false);
862 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
863 _igvn.hash_delete(lower_bound_iff);
864 lower_bound_iff->set_req(1, lower_bound_bol);
865 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx);
867 // Test the upper bound
868 Node* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true);
869 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
870 _igvn.hash_delete(upper_bound_iff);
871 upper_bound_iff->set_req(1, upper_bound_bol);
872 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx);
874 // Fall through into rest of the clean up code which will move
875 // any dependent nodes onto the upper bound test.
876 new_predicate_proj = upper_bound_proj;
878 #ifndef PRODUCT
879 if (TraceLoopOpts && !TraceLoopPredicate) {
880 tty->print("Predicate RC ");
881 loop->dump_head();
882 }
883 #endif
884 } else {
885 // Loop variant check (for example, range check in non-counted loop)
886 // with uncommon trap.
887 continue;
888 }
889 assert(new_predicate_proj != NULL, "sanity");
890 // Success - attach condition (new_predicate_bol) to predicate if
891 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
893 // Eliminate the old If in the loop body
894 dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
896 hoisted = true;
897 C->set_major_progress();
898 } // end while
900 #ifndef PRODUCT
901 // report that the loop predication has been actually performed
902 // for this loop
903 if (TraceLoopPredicate && hoisted) {
904 tty->print("Loop Predication Performed:");
905 loop->dump_head();
906 }
907 #endif
909 return hoisted;
910 }
912 //------------------------------loop_predication--------------------------------
913 // driver routine for loop predication optimization
914 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
915 bool hoisted = false;
916 // Recursively promote predicates
917 if (_child) {
918 hoisted = _child->loop_predication( phase);
919 }
921 // self
922 if (!_irreducible && !tail()->is_top()) {
923 hoisted |= phase->loop_predication_impl(this);
924 }
926 if (_next) { //sibling
927 hoisted |= _next->loop_predication( phase);
928 }
930 return hoisted;
931 }