1 /*
2 * Copyright (c) 2000, 2017, 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 "compiler/compileLog.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/convertnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/opaquenode.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/superword.hpp"
42 #include "opto/vectornode.hpp"
43
44 //------------------------------is_loop_exit-----------------------------------
45 // Given an IfNode, return the loop-exiting projection or NULL if both
46 // arms remain in the loop.
47 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
48 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
49 PhaseIdealLoop *phase = _phase;
50 // Test is an IfNode, has 2 projections. If BOTH are in the loop
51 // we need loop unswitching instead of peeling.
52 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
53 return iff->raw_out(0);
54 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
55 return iff->raw_out(1);
56 return NULL;
57 }
58
59
60 //=============================================================================
61
62
63 //------------------------------record_for_igvn----------------------------
64 // Put loop body on igvn work list
65 void IdealLoopTree::record_for_igvn() {
66 for( uint i = 0; i < _body.size(); i++ ) {
67 Node *n = _body.at(i);
68 _phase->_igvn._worklist.push(n);
69 }
70 // put body of outer strip mined loop on igvn work list as well
71 if (_head->is_CountedLoop() && _head->as_Loop()->is_strip_mined()) {
72 CountedLoopNode* l = _head->as_CountedLoop();
73 _phase->_igvn._worklist.push(l->outer_loop());
74 _phase->_igvn._worklist.push(l->outer_loop_tail());
75 _phase->_igvn._worklist.push(l->outer_loop_end());
76 _phase->_igvn._worklist.push(l->outer_safepoint());
77 _phase->_igvn._worklist.push(l->outer_bol());
78 _phase->_igvn._worklist.push(l->outer_cmp());
79 _phase->_igvn._worklist.push(l->outer_opaq());
80 Node* cle_out = _head->as_CountedLoop()->loopexit()->proj_out(false);
81 _phase->_igvn._worklist.push(cle_out);
82 }
83 }
84
85 //------------------------------compute_exact_trip_count-----------------------
86 // Compute loop trip count if possible. Do not recalculate trip count for
87 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
88 void IdealLoopTree::compute_trip_count(PhaseIdealLoop* phase) {
89 if (!_head->as_Loop()->is_valid_counted_loop()) {
90 return;
91 }
92 CountedLoopNode* cl = _head->as_CountedLoop();
93 // Trip count may become nonexact for iteration split loops since
94 // RCE modifies limits. Note, _trip_count value is not reset since
95 // it is used to limit unrolling of main loop.
96 cl->set_nonexact_trip_count();
97
98 // Loop's test should be part of loop.
99 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
100 return; // Infinite loop
101
102 #ifdef ASSERT
103 BoolTest::mask bt = cl->loopexit()->test_trip();
104 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
105 bt == BoolTest::ne, "canonical test is expected");
106 #endif
107
108 Node* init_n = cl->init_trip();
109 Node* limit_n = cl->limit();
110 if (init_n != NULL && limit_n != NULL) {
111 // Use longs to avoid integer overflow.
112 int stride_con = cl->stride_con();
113 jlong init_con = phase->_igvn.type(init_n)->is_int()->_lo;
114 jlong limit_con = phase->_igvn.type(limit_n)->is_int()->_hi;
115 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
116 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
117 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
118 if (init_n->is_Con() && limit_n->is_Con()) {
119 // Set exact trip count.
120 cl->set_exact_trip_count((uint)trip_count);
121 } else if (cl->unrolled_count() == 1) {
122 // Set maximum trip count before unrolling.
123 cl->set_trip_count((uint)trip_count);
124 }
125 }
126 }
127 }
128
129 //------------------------------compute_profile_trip_cnt----------------------------
130 // Compute loop trip count from profile data as
131 // (backedge_count + loop_exit_count) / loop_exit_count
132 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
133 if (!_head->is_CountedLoop()) {
134 return;
135 }
136 CountedLoopNode* head = _head->as_CountedLoop();
137 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
138 return; // Already computed
139 }
140 float trip_cnt = (float)max_jint; // default is big
141
142 Node* back = head->in(LoopNode::LoopBackControl);
143 while (back != head) {
144 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
145 back->in(0) &&
146 back->in(0)->is_If() &&
147 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
148 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
149 break;
150 }
151 back = phase->idom(back);
152 }
153 if (back != head) {
154 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
155 back->in(0), "if-projection exists");
156 IfNode* back_if = back->in(0)->as_If();
157 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
158
159 // Now compute a loop exit count
160 float loop_exit_cnt = 0.0f;
161 for( uint i = 0; i < _body.size(); i++ ) {
162 Node *n = _body[i];
163 if( n->is_If() ) {
164 IfNode *iff = n->as_If();
165 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
166 Node *exit = is_loop_exit(iff);
167 if( exit ) {
168 float exit_prob = iff->_prob;
169 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
170 if (exit_prob > PROB_MIN) {
171 float exit_cnt = iff->_fcnt * exit_prob;
172 loop_exit_cnt += exit_cnt;
173 }
174 }
175 }
176 }
177 }
178 if (loop_exit_cnt > 0.0f) {
179 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
180 } else {
181 // No exit count so use
182 trip_cnt = loop_back_cnt;
183 }
184 }
185 #ifndef PRODUCT
186 if (TraceProfileTripCount) {
187 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
188 }
189 #endif
190 head->set_profile_trip_cnt(trip_cnt);
191 }
192
193 //---------------------is_invariant_addition-----------------------------
194 // Return nonzero index of invariant operand for an Add or Sub
195 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
196 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
197 int op = n->Opcode();
198 if (op == Op_AddI || op == Op_SubI) {
199 bool in1_invar = this->is_invariant(n->in(1));
200 bool in2_invar = this->is_invariant(n->in(2));
201 if (in1_invar && !in2_invar) return 1;
202 if (!in1_invar && in2_invar) return 2;
203 }
204 return 0;
205 }
206
207 //---------------------reassociate_add_sub-----------------------------
208 // Reassociate invariant add and subtract expressions:
209 //
210 // inv1 + (x + inv2) => ( inv1 + inv2) + x
211 // (x + inv2) + inv1 => ( inv1 + inv2) + x
212 // inv1 + (x - inv2) => ( inv1 - inv2) + x
213 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
214 // (x + inv2) - inv1 => (-inv1 + inv2) + x
215 // (x - inv2) + inv1 => ( inv1 - inv2) + x
216 // (x - inv2) - inv1 => (-inv1 - inv2) + x
217 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
218 // inv1 - (x - inv2) => ( inv1 + inv2) - x
219 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
220 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
221 // inv1 - (x + inv2) => ( inv1 - inv2) - x
222 //
223 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
224 if ((!n1->is_Add() && !n1->is_Sub()) || n1->outcnt() == 0) return NULL;
225 if (is_invariant(n1)) return NULL;
226 int inv1_idx = is_invariant_addition(n1, phase);
227 if (!inv1_idx) return NULL;
228 // Don't mess with add of constant (igvn moves them to expression tree root.)
229 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
230 Node* inv1 = n1->in(inv1_idx);
231 Node* n2 = n1->in(3 - inv1_idx);
232 int inv2_idx = is_invariant_addition(n2, phase);
233 if (!inv2_idx) return NULL;
234 Node* x = n2->in(3 - inv2_idx);
235 Node* inv2 = n2->in(inv2_idx);
236
237 bool neg_x = n2->is_Sub() && inv2_idx == 1;
238 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
239 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
240 if (n1->is_Sub() && inv1_idx == 1) {
241 neg_x = !neg_x;
242 neg_inv2 = !neg_inv2;
243 }
244 Node* inv1_c = phase->get_ctrl(inv1);
245 Node* inv2_c = phase->get_ctrl(inv2);
246 Node* n_inv1;
247 if (neg_inv1) {
248 Node *zero = phase->_igvn.intcon(0);
249 phase->set_ctrl(zero, phase->C->root());
250 n_inv1 = new SubINode(zero, inv1);
251 phase->register_new_node(n_inv1, inv1_c);
252 } else {
253 n_inv1 = inv1;
254 }
255 Node* inv;
256 if (neg_inv2) {
257 inv = new SubINode(n_inv1, inv2);
258 } else {
259 inv = new AddINode(n_inv1, inv2);
260 }
261 phase->register_new_node(inv, phase->get_early_ctrl(inv));
262
263 Node* addx;
264 if (neg_x) {
265 addx = new SubINode(inv, x);
266 } else {
267 addx = new AddINode(x, inv);
268 }
269 phase->register_new_node(addx, phase->get_ctrl(x));
270 phase->_igvn.replace_node(n1, addx);
271 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
272 _body.yank(n1);
273 return addx;
274 }
275
276 //---------------------reassociate_invariants-----------------------------
277 // Reassociate invariant expressions:
278 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
279 for (int i = _body.size() - 1; i >= 0; i--) {
280 Node *n = _body.at(i);
281 for (int j = 0; j < 5; j++) {
282 Node* nn = reassociate_add_sub(n, phase);
283 if (nn == NULL) break;
284 n = nn; // again
285 };
286 }
287 }
288
289 //------------------------------policy_peeling---------------------------------
290 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
291 // make some loop-invariant test (usually a null-check) happen before the loop.
292 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
293 Node *test = ((IdealLoopTree*)this)->tail();
294 int body_size = ((IdealLoopTree*)this)->_body.size();
295 // Peeling does loop cloning which can result in O(N^2) node construction
296 if( body_size > 255 /* Prevent overflow for large body_size */
297 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) {
298 return false; // too large to safely clone
299 }
300
301 // check for vectorized loops, any peeling done was already applied
302 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false;
303
304 while( test != _head ) { // Scan till run off top of loop
305 if( test->is_If() ) { // Test?
306 Node *ctrl = phase->get_ctrl(test->in(1));
307 if (ctrl->is_top())
308 return false; // Found dead test on live IF? No peeling!
309 // Standard IF only has one input value to check for loop invariance
310 assert(test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd || test->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
311 // Condition is not a member of this loop?
312 if( !is_member(phase->get_loop(ctrl)) &&
313 is_loop_exit(test) )
314 return true; // Found reason to peel!
315 }
316 // Walk up dominators to loop _head looking for test which is
317 // executed on every path thru loop.
318 test = phase->idom(test);
319 }
320 return false;
321 }
322
323 //------------------------------peeled_dom_test_elim---------------------------
324 // If we got the effect of peeling, either by actually peeling or by making
325 // a pre-loop which must execute at least once, we can remove all
326 // loop-invariant dominated tests in the main body.
327 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
328 bool progress = true;
329 while( progress ) {
330 progress = false; // Reset for next iteration
331 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
332 Node *test = prev->in(0);
333 while( test != loop->_head ) { // Scan till run off top of loop
334
335 int p_op = prev->Opcode();
336 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
337 test->is_If() && // Test?
338 !test->in(1)->is_Con() && // And not already obvious?
339 // Condition is not a member of this loop?
340 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
341 // Walk loop body looking for instances of this test
342 for( uint i = 0; i < loop->_body.size(); i++ ) {
343 Node *n = loop->_body.at(i);
344 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
345 // IfNode was dominated by version in peeled loop body
346 progress = true;
347 dominated_by( old_new[prev->_idx], n );
348 }
349 }
350 }
351 prev = test;
352 test = idom(test);
353 } // End of scan tests in loop
354
355 } // End of while( progress )
356 }
357
358 //------------------------------do_peeling-------------------------------------
359 // Peel the first iteration of the given loop.
360 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
361 // The pre-loop illegally has 2 control users (old & new loops).
362 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
363 // Do this by making the old-loop fall-in edges act as if they came
364 // around the loopback from the prior iteration (follow the old-loop
365 // backedges) and then map to the new peeled iteration. This leaves
366 // the pre-loop with only 1 user (the new peeled iteration), but the
367 // peeled-loop backedge has 2 users.
368 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
369 // extra backedge user.
370 //
371 // orig
372 //
373 // stmt1
374 // |
375 // v
376 // loop predicate
377 // |
378 // v
379 // loop<----+
380 // | |
381 // stmt2 |
382 // | |
383 // v |
384 // if ^
385 // / \ |
386 // / \ |
387 // v v |
388 // false true |
389 // / \ |
390 // / ----+
391 // |
392 // v
393 // exit
394 //
395 //
396 // after clone loop
397 //
398 // stmt1
399 // |
400 // v
401 // loop predicate
402 // / \
403 // clone / \ orig
404 // / \
405 // / \
406 // v v
407 // +---->loop clone loop<----+
408 // | | | |
409 // | stmt2 clone stmt2 |
410 // | | | |
411 // | v v |
412 // ^ if clone If ^
413 // | / \ / \ |
414 // | / \ / \ |
415 // | v v v v |
416 // | true false false true |
417 // | / \ / \ |
418 // +---- \ / ----+
419 // \ /
420 // 1v v2
421 // region
422 // |
423 // v
424 // exit
425 //
426 //
427 // after peel and predicate move
428 //
429 // stmt1
430 // /
431 // /
432 // clone / orig
433 // /
434 // / +----------+
435 // / | |
436 // / loop predicate |
437 // / | |
438 // v v |
439 // TOP-->loop clone loop<----+ |
440 // | | | |
441 // stmt2 clone stmt2 | |
442 // | | | ^
443 // v v | |
444 // if clone If ^ |
445 // / \ / \ | |
446 // / \ / \ | |
447 // v v v v | |
448 // true false false true | |
449 // | \ / \ | |
450 // | \ / ----+ ^
451 // | \ / |
452 // | 1v v2 |
453 // v region |
454 // | | |
455 // | v |
456 // | exit |
457 // | |
458 // +--------------->-----------------+
459 //
460 //
461 // final graph
462 //
463 // stmt1
464 // |
465 // v
466 // stmt2 clone
467 // |
468 // v
469 // if clone
470 // / |
471 // / |
472 // v v
473 // false true
474 // | |
475 // | v
476 // | loop predicate
477 // | |
478 // | v
479 // | loop<----+
480 // | | |
481 // | stmt2 |
482 // | | |
483 // | v |
484 // v if ^
485 // | / \ |
486 // | / \ |
487 // | v v |
488 // | false true |
489 // | | \ |
490 // v v --+
491 // region
492 // |
493 // v
494 // exit
495 //
496 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
497
498 C->set_major_progress();
499 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
500 // 'pre' loop from the main and the 'pre' can no longer have its
501 // iterations adjusted. Therefore, we need to declare this loop as
502 // no longer a 'main' loop; it will need new pre and post loops before
503 // we can do further RCE.
504 #ifndef PRODUCT
505 if (TraceLoopOpts) {
506 tty->print("Peel ");
507 loop->dump_head();
508 }
509 #endif
510 LoopNode* head = loop->_head->as_Loop();
511 bool counted_loop = head->is_CountedLoop();
512 if (counted_loop) {
513 CountedLoopNode *cl = head->as_CountedLoop();
514 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
515 cl->set_trip_count(cl->trip_count() - 1);
516 if (cl->is_main_loop()) {
517 cl->set_normal_loop();
518 #ifndef PRODUCT
519 if (PrintOpto && VerifyLoopOptimizations) {
520 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
521 loop->dump_head();
522 }
523 #endif
524 }
525 }
526 Node* entry = head->in(LoopNode::EntryControl);
527
528 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
529 // The pre-loop illegally has 2 control users (old & new loops).
530 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), ControlAroundStripMined);
531
532 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
533 // Do this by making the old-loop fall-in edges act as if they came
534 // around the loopback from the prior iteration (follow the old-loop
535 // backedges) and then map to the new peeled iteration. This leaves
536 // the pre-loop with only 1 user (the new peeled iteration), but the
537 // peeled-loop backedge has 2 users.
538 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
539 _igvn.hash_delete(head->skip_strip_mined());
540 head->skip_strip_mined()->set_req(LoopNode::EntryControl, new_entry);
541 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
542 Node* old = head->fast_out(j);
543 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
544 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
545 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
546 // Then loop body backedge value remains the same.
547 new_exit_value = old->in(LoopNode::LoopBackControl);
548 _igvn.hash_delete(old);
549 old->set_req(LoopNode::EntryControl, new_exit_value);
550 }
551 }
552
553
554 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
555 // extra backedge user.
556 Node* new_head = old_new[head->_idx];
557 _igvn.hash_delete(new_head);
558 new_head->set_req(LoopNode::LoopBackControl, C->top());
559 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
560 Node* use = new_head->fast_out(j2);
561 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
562 _igvn.hash_delete(use);
563 use->set_req(LoopNode::LoopBackControl, C->top());
564 }
565 }
566
567
568 // Step 4: Correct dom-depth info. Set to loop-head depth.
569 int dd = dom_depth(head);
570 set_idom(head, head->in(1), dd);
571 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
572 Node *old = loop->_body.at(j3);
573 Node *nnn = old_new[old->_idx];
574 if (!has_ctrl(nnn))
575 set_idom(nnn, idom(nnn), dd-1);
576 }
577
578 // Now force out all loop-invariant dominating tests. The optimizer
579 // finds some, but we _know_ they are all useless.
580 peeled_dom_test_elim(loop,old_new);
581
582 loop->record_for_igvn();
583 }
584
585 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
586
587 //------------------------------policy_maximally_unroll------------------------
588 // Calculate exact loop trip count and return true if loop can be maximally
589 // unrolled.
590 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
591 CountedLoopNode *cl = _head->as_CountedLoop();
592 assert(cl->is_normal_loop(), "");
593 if (!cl->is_valid_counted_loop())
594 return false; // Malformed counted loop
595
596 if (!cl->has_exact_trip_count()) {
597 // Trip count is not exact.
598 return false;
599 }
600
601 uint trip_count = cl->trip_count();
602 // Note, max_juint is used to indicate unknown trip count.
603 assert(trip_count > 1, "one iteration loop should be optimized out already");
604 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
605
606 // Real policy: if we maximally unroll, does it get too big?
607 // Allow the unrolled mess to get larger than standard loop
608 // size. After all, it will no longer be a loop.
609 uint body_size = _body.size();
610 uint unroll_limit = (uint)LoopUnrollLimit * 4;
611 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
612 if (trip_count > unroll_limit || body_size > unroll_limit) {
613 return false;
614 }
615
616 // Fully unroll a loop with few iterations regardless next
617 // conditions since following loop optimizations will split
618 // such loop anyway (pre-main-post).
619 if (trip_count <= 3)
620 return true;
621
622 // Take into account that after unroll conjoined heads and tails will fold,
623 // otherwise policy_unroll() may allow more unrolling than max unrolling.
624 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
625 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
626 if (body_size != tst_body_size) // Check for int overflow
627 return false;
628 if (new_body_size > unroll_limit ||
629 // Unrolling can result in a large amount of node construction
630 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) {
631 return false;
632 }
633
634 // Do not unroll a loop with String intrinsics code.
635 // String intrinsics are large and have loops.
636 for (uint k = 0; k < _body.size(); k++) {
637 Node* n = _body.at(k);
638 switch (n->Opcode()) {
639 case Op_StrComp:
640 case Op_StrEquals:
641 case Op_StrIndexOf:
642 case Op_StrIndexOfChar:
643 case Op_EncodeISOArray:
644 case Op_AryEq:
645 case Op_HasNegatives: {
646 return false;
647 }
648 #if INCLUDE_RTM_OPT
649 case Op_FastLock:
650 case Op_FastUnlock: {
651 // Don't unroll RTM locking code because it is large.
652 if (UseRTMLocking) {
653 return false;
654 }
655 }
656 #endif
657 } // switch
658 }
659
660 return true; // Do maximally unroll
661 }
662
663
664 //------------------------------policy_unroll----------------------------------
665 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
666 // the loop is a CountedLoop and the body is small enough.
667 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
668
669 CountedLoopNode *cl = _head->as_CountedLoop();
670 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
671
672 if (!cl->is_valid_counted_loop())
673 return false; // Malformed counted loop
674
675 // Protect against over-unrolling.
676 // After split at least one iteration will be executed in pre-loop.
677 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
678
679 _local_loop_unroll_limit = LoopUnrollLimit;
680 _local_loop_unroll_factor = 4;
681 int future_unroll_ct = cl->unrolled_count() * 2;
682 if (!cl->do_unroll_only()) {
683 if (future_unroll_ct > LoopMaxUnroll) return false;
684 } else {
685 // obey user constraints on vector mapped loops with additional unrolling applied
686 int unroll_constraint = (cl->slp_max_unroll()) ? cl->slp_max_unroll() : 1;
687 if ((future_unroll_ct / unroll_constraint) > LoopMaxUnroll) return false;
688 }
689
690 // Check for initial stride being a small enough constant
691 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
692
693 // Don't unroll if the next round of unrolling would push us
694 // over the expected trip count of the loop. One is subtracted
695 // from the expected trip count because the pre-loop normally
696 // executes 1 iteration.
697 if (UnrollLimitForProfileCheck > 0 &&
698 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
699 future_unroll_ct > UnrollLimitForProfileCheck &&
700 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
701 return false;
702 }
703
704 // When unroll count is greater than LoopUnrollMin, don't unroll if:
705 // the residual iterations are more than 10% of the trip count
706 // and rounds of "unroll,optimize" are not making significant progress
707 // Progress defined as current size less than 20% larger than previous size.
708 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
709 future_unroll_ct > LoopUnrollMin &&
710 (future_unroll_ct - 1) * (100 / LoopPercentProfileLimit) > cl->profile_trip_cnt() &&
711 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
712 return false;
713 }
714
715 Node *init_n = cl->init_trip();
716 Node *limit_n = cl->limit();
717 int stride_con = cl->stride_con();
718 // Non-constant bounds.
719 // Protect against over-unrolling when init or/and limit are not constant
720 // (so that trip_count's init value is maxint) but iv range is known.
721 if (init_n == NULL || !init_n->is_Con() ||
722 limit_n == NULL || !limit_n->is_Con()) {
723 Node* phi = cl->phi();
724 if (phi != NULL) {
725 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
726 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
727 int next_stride = stride_con * 2; // stride after this unroll
728 if (next_stride > 0) {
729 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
730 iv_type->_lo + next_stride > iv_type->_hi) {
731 return false; // over-unrolling
732 }
733 } else if (next_stride < 0) {
734 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
735 iv_type->_hi + next_stride < iv_type->_lo) {
736 return false; // over-unrolling
737 }
738 }
739 }
740 }
741
742 // After unroll limit will be adjusted: new_limit = limit-stride.
743 // Bailout if adjustment overflow.
744 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
745 if ((stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi)) ||
746 (stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)))
747 return false; // overflow
748
749 // Adjust body_size to determine if we unroll or not
750 uint body_size = _body.size();
751 // Key test to unroll loop in CRC32 java code
752 int xors_in_loop = 0;
753 // Also count ModL, DivL and MulL which expand mightly
754 for (uint k = 0; k < _body.size(); k++) {
755 Node* n = _body.at(k);
756 switch (n->Opcode()) {
757 case Op_XorI: xors_in_loop++; break; // CRC32 java code
758 case Op_ModL: body_size += 30; break;
759 case Op_DivL: body_size += 30; break;
760 case Op_MulL: body_size += 10; break;
761 case Op_StrComp:
762 case Op_StrEquals:
763 case Op_StrIndexOf:
764 case Op_StrIndexOfChar:
765 case Op_EncodeISOArray:
766 case Op_AryEq:
767 case Op_HasNegatives: {
768 // Do not unroll a loop with String intrinsics code.
769 // String intrinsics are large and have loops.
770 return false;
771 }
772 #if INCLUDE_RTM_OPT
773 case Op_FastLock:
774 case Op_FastUnlock: {
775 // Don't unroll RTM locking code because it is large.
776 if (UseRTMLocking) {
777 return false;
778 }
779 }
780 #endif
781 } // switch
782 }
783
784 if (UseSuperWord) {
785 if (!cl->is_reduction_loop()) {
786 phase->mark_reductions(this);
787 }
788
789 // Only attempt slp analysis when user controls do not prohibit it
790 if (LoopMaxUnroll > _local_loop_unroll_factor) {
791 // Once policy_slp_analysis succeeds, mark the loop with the
792 // maximal unroll factor so that we minimize analysis passes
793 if (future_unroll_ct >= _local_loop_unroll_factor) {
794 policy_unroll_slp_analysis(cl, phase, future_unroll_ct);
795 }
796 }
797 }
798
799 int slp_max_unroll_factor = cl->slp_max_unroll();
800 if ((LoopMaxUnroll < slp_max_unroll_factor) && FLAG_IS_DEFAULT(LoopMaxUnroll) && UseSubwordForMaxVector) {
801 LoopMaxUnroll = slp_max_unroll_factor;
802 }
803 if (cl->has_passed_slp()) {
804 if (slp_max_unroll_factor >= future_unroll_ct) return true;
805 // Normal case: loop too big
806 return false;
807 }
808
809 // Check for being too big
810 if (body_size > (uint)_local_loop_unroll_limit) {
811 if ((UseSubwordForMaxVector || xors_in_loop >= 4) && body_size < (uint)LoopUnrollLimit * 4) return true;
812 // Normal case: loop too big
813 return false;
814 }
815
816 if (cl->do_unroll_only()) {
817 if (TraceSuperWordLoopUnrollAnalysis) {
818 tty->print_cr("policy_unroll passed vector loop(vlen=%d,factor = %d)\n", slp_max_unroll_factor, future_unroll_ct);
819 }
820 }
821
822 // Unroll once! (Each trip will soon do double iterations)
823 return true;
824 }
825
826 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) {
827 // Enable this functionality target by target as needed
828 if (SuperWordLoopUnrollAnalysis) {
829 if (!cl->was_slp_analyzed()) {
830 SuperWord sw(phase);
831 sw.transform_loop(this, false);
832
833 // If the loop is slp canonical analyze it
834 if (sw.early_return() == false) {
835 sw.unrolling_analysis(_local_loop_unroll_factor);
836 }
837 }
838
839 if (cl->has_passed_slp()) {
840 int slp_max_unroll_factor = cl->slp_max_unroll();
841 if (slp_max_unroll_factor >= future_unroll_ct) {
842 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor;
843 if (new_limit > LoopUnrollLimit) {
844 if (TraceSuperWordLoopUnrollAnalysis) {
845 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit);
846 }
847 _local_loop_unroll_limit = new_limit;
848 }
849 }
850 }
851 }
852 }
853
854 //------------------------------policy_align-----------------------------------
855 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
856 // expression that does the alignment. Note that only one array base can be
857 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
858 // if we vectorize short memory ops into longer memory ops, we may want to
859 // increase alignment.
860 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
861 return false;
862 }
863
864 //------------------------------policy_range_check-----------------------------
865 // Return TRUE or FALSE if the loop should be range-check-eliminated.
866 // Actually we do iteration-splitting, a more powerful form of RCE.
867 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
868 if (!RangeCheckElimination) return false;
869
870 CountedLoopNode *cl = _head->as_CountedLoop();
871 // If we unrolled with no intention of doing RCE and we later
872 // changed our minds, we got no pre-loop. Either we need to
873 // make a new pre-loop, or we gotta disallow RCE.
874 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
875 Node *trip_counter = cl->phi();
876
877 // check for vectorized loops, some opts are no longer needed
878 if (cl->do_unroll_only()) return false;
879
880 // Check loop body for tests of trip-counter plus loop-invariant vs
881 // loop-invariant.
882 for (uint i = 0; i < _body.size(); i++) {
883 Node *iff = _body[i];
884 if (iff->Opcode() == Op_If ||
885 iff->Opcode() == Op_RangeCheck) { // Test?
886
887 // Comparing trip+off vs limit
888 Node *bol = iff->in(1);
889 if (bol->req() != 2) continue; // dead constant test
890 if (!bol->is_Bool()) {
891 assert(bol->Opcode() == Op_Conv2B, "predicate check only");
892 continue;
893 }
894 if (bol->as_Bool()->_test._test == BoolTest::ne)
895 continue; // not RC
896
897 Node *cmp = bol->in(1);
898 Node *rc_exp = cmp->in(1);
899 Node *limit = cmp->in(2);
900
901 Node *limit_c = phase->get_ctrl(limit);
902 if( limit_c == phase->C->top() )
903 return false; // Found dead test on live IF? No RCE!
904 if( is_member(phase->get_loop(limit_c) ) ) {
905 // Compare might have operands swapped; commute them
906 rc_exp = cmp->in(2);
907 limit = cmp->in(1);
908 limit_c = phase->get_ctrl(limit);
909 if( is_member(phase->get_loop(limit_c) ) )
910 continue; // Both inputs are loop varying; cannot RCE
911 }
912
913 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
914 continue;
915 }
916 // Yeah! Found a test like 'trip+off vs limit'
917 // Test is an IfNode, has 2 projections. If BOTH are in the loop
918 // we need loop unswitching instead of iteration splitting.
919 if( is_loop_exit(iff) )
920 return true; // Found reason to split iterations
921 } // End of is IF
922 }
923
924 return false;
925 }
926
927 //------------------------------policy_peel_only-------------------------------
928 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
929 // for unrolling loops with NO array accesses.
930 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
931 // check for vectorized loops, any peeling done was already applied
932 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false;
933
934 for( uint i = 0; i < _body.size(); i++ )
935 if( _body[i]->is_Mem() )
936 return false;
937
938 // No memory accesses at all!
939 return true;
940 }
941
942 //------------------------------clone_up_backedge_goo--------------------------
943 // If Node n lives in the back_ctrl block and cannot float, we clone a private
944 // version of n in preheader_ctrl block and return that, otherwise return n.
945 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
946 if( get_ctrl(n) != back_ctrl ) return n;
947
948 // Only visit once
949 if (visited.test_set(n->_idx)) {
950 Node *x = clones.find(n->_idx);
951 if (x != NULL)
952 return x;
953 return n;
954 }
955
956 Node *x = NULL; // If required, a clone of 'n'
957 // Check for 'n' being pinned in the backedge.
958 if( n->in(0) && n->in(0) == back_ctrl ) {
959 assert(clones.find(n->_idx) == NULL, "dead loop");
960 x = n->clone(); // Clone a copy of 'n' to preheader
961 clones.push(x, n->_idx);
962 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
963 }
964
965 // Recursive fixup any other input edges into x.
966 // If there are no changes we can just return 'n', otherwise
967 // we need to clone a private copy and change it.
968 for( uint i = 1; i < n->req(); i++ ) {
969 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
970 if( g != n->in(i) ) {
971 if( !x ) {
972 assert(clones.find(n->_idx) == NULL, "dead loop");
973 x = n->clone();
974 clones.push(x, n->_idx);
975 }
976 x->set_req(i, g);
977 }
978 }
979 if( x ) { // x can legally float to pre-header location
980 register_new_node( x, preheader_ctrl );
981 return x;
982 } else { // raise n to cover LCA of uses
983 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
984 }
985 return n;
986 }
987
988 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) {
989 Node* castii = new CastIINode(incr, TypeInt::INT, true);
990 castii->set_req(0, ctrl);
991 register_new_node(castii, ctrl);
992 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) {
993 Node* n = incr->fast_out(i);
994 if (n->is_Phi() && n->in(0) == loop) {
995 int nrep = n->replace_edge(incr, castii);
996 return true;
997 }
998 }
999 return false;
1000 }
1001
1002 //------------------------------insert_pre_post_loops--------------------------
1003 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
1004 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
1005 // alignment. Useful to unroll loops that do no array accesses.
1006 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
1007
1008 #ifndef PRODUCT
1009 if (TraceLoopOpts) {
1010 if (peel_only)
1011 tty->print("PeelMainPost ");
1012 else
1013 tty->print("PreMainPost ");
1014 loop->dump_head();
1015 }
1016 #endif
1017 C->set_major_progress();
1018
1019 // Find common pieces of the loop being guarded with pre & post loops
1020 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1021 assert( main_head->is_normal_loop(), "" );
1022 CountedLoopEndNode *main_end = main_head->loopexit();
1023 guarantee(main_end != NULL, "no loop exit node");
1024 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
1025
1026 Node *pre_header= main_head->in(LoopNode::EntryControl);
1027 Node *init = main_head->init_trip();
1028 Node *incr = main_end ->incr();
1029 Node *limit = main_end ->limit();
1030 Node *stride = main_end ->stride();
1031 Node *cmp = main_end ->cmp_node();
1032 BoolTest::mask b_test = main_end->test_trip();
1033
1034 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
1035 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
1036 if( bol->outcnt() != 1 ) {
1037 bol = bol->clone();
1038 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
1039 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol);
1040 }
1041 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
1042 if( cmp->outcnt() != 1 ) {
1043 cmp = cmp->clone();
1044 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
1045 _igvn.replace_input_of(bol, 1, cmp);
1046 }
1047
1048 // Add the post loop
1049 CountedLoopNode *post_head = NULL;
1050 Node *main_exit = insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1051
1052 //------------------------------
1053 // Step B: Create Pre-Loop.
1054
1055 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1056 // loop pre-header illegally has 2 control users (old & new loops).
1057 LoopNode* outer_main_head = main_head;
1058 IdealLoopTree* outer_loop = loop;
1059 if (main_head->is_strip_mined()) {
1060 main_head->verify_strip_mined(1);
1061 outer_main_head = main_head->outer_loop();
1062 outer_loop = loop->_parent;
1063 assert(outer_loop->_head == outer_main_head, "broken loop tree");
1064 }
1065 uint dd_main_head = dom_depth(outer_main_head);
1066 clone_loop(loop, old_new, dd_main_head, ControlAroundStripMined);
1067 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1068 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1069 pre_head->set_pre_loop(main_head);
1070 Node *pre_incr = old_new[incr->_idx];
1071
1072 // Reduce the pre-loop trip count.
1073 pre_end->_prob = PROB_FAIR;
1074
1075 // Find the pre-loop normal exit.
1076 Node* pre_exit = pre_end->proj_out(false);
1077 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1078 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end);
1079 _igvn.register_new_node_with_optimizer( new_pre_exit );
1080 set_idom(new_pre_exit, pre_end, dd_main_head);
1081 set_loop(new_pre_exit, outer_loop->_parent);
1082
1083 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1084 // pre-loop, the main-loop may not execute at all. Later in life this
1085 // zero-trip guard will become the minimum-trip guard when we unroll
1086 // the main-loop.
1087 Node *min_opaq = new Opaque1Node(C, limit);
1088 Node *min_cmp = new CmpINode( pre_incr, min_opaq );
1089 Node *min_bol = new BoolNode( min_cmp, b_test );
1090 register_new_node( min_opaq, new_pre_exit );
1091 register_new_node( min_cmp , new_pre_exit );
1092 register_new_node( min_bol , new_pre_exit );
1093
1094 // Build the IfNode (assume the main-loop is executed always).
1095 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1096 _igvn.register_new_node_with_optimizer( min_iff );
1097 set_idom(min_iff, new_pre_exit, dd_main_head);
1098 set_loop(min_iff, outer_loop->_parent);
1099
1100 // Plug in the false-path, taken if we need to skip main-loop
1101 _igvn.hash_delete( pre_exit );
1102 pre_exit->set_req(0, min_iff);
1103 set_idom(pre_exit, min_iff, dd_main_head);
1104 set_idom(pre_exit->unique_ctrl_out(), min_iff, dd_main_head);
1105 // Make the true-path, must enter the main loop
1106 Node *min_taken = new IfTrueNode( min_iff );
1107 _igvn.register_new_node_with_optimizer( min_taken );
1108 set_idom(min_taken, min_iff, dd_main_head);
1109 set_loop(min_taken, outer_loop->_parent);
1110 // Plug in the true path
1111 _igvn.hash_delete(outer_main_head);
1112 outer_main_head->set_req(LoopNode::EntryControl, min_taken);
1113 set_idom(outer_main_head, min_taken, dd_main_head);
1114
1115 Arena *a = Thread::current()->resource_area();
1116 VectorSet visited(a);
1117 Node_Stack clones(a, main_head->back_control()->outcnt());
1118 // Step B3: Make the fall-in values to the main-loop come from the
1119 // fall-out values of the pre-loop.
1120 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1121 Node* main_phi = main_head->fast_out(i2);
1122 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1123 Node *pre_phi = old_new[main_phi->_idx];
1124 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1125 main_head->skip_strip_mined()->in(LoopNode::EntryControl),
1126 pre_phi->in(LoopNode::LoopBackControl),
1127 visited, clones);
1128 _igvn.hash_delete(main_phi);
1129 main_phi->set_req( LoopNode::EntryControl, fallpre );
1130 }
1131 }
1132
1133 // Nodes inside the loop may be control dependent on a predicate
1134 // that was moved before the preloop. If the back branch of the main
1135 // or post loops becomes dead, those nodes won't be dependent on the
1136 // test that guards that loop nest anymore which could lead to an
1137 // incorrect array access because it executes independently of the
1138 // test that was guarding the loop nest. We add a special CastII on
1139 // the if branch that enters the loop, between the input induction
1140 // variable value and the induction variable Phi to preserve correct
1141 // dependencies.
1142
1143 // CastII for the main loop:
1144 bool inserted = cast_incr_before_loop( pre_incr, min_taken, main_head );
1145 assert(inserted, "no castII inserted");
1146
1147 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1148 // RCE and alignment may change this later.
1149 Node *cmp_end = pre_end->cmp_node();
1150 assert( cmp_end->in(2) == limit, "" );
1151 Node *pre_limit = new AddINode( init, stride );
1152
1153 // Save the original loop limit in this Opaque1 node for
1154 // use by range check elimination.
1155 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit);
1156
1157 register_new_node( pre_limit, pre_head->in(0) );
1158 register_new_node( pre_opaq , pre_head->in(0) );
1159
1160 // Since no other users of pre-loop compare, I can hack limit directly
1161 assert( cmp_end->outcnt() == 1, "no other users" );
1162 _igvn.hash_delete(cmp_end);
1163 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1164
1165 // Special case for not-equal loop bounds:
1166 // Change pre loop test, main loop test, and the
1167 // main loop guard test to use lt or gt depending on stride
1168 // direction:
1169 // positive stride use <
1170 // negative stride use >
1171 //
1172 // not-equal test is kept for post loop to handle case
1173 // when init > limit when stride > 0 (and reverse).
1174
1175 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1176
1177 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1178 // Modify pre loop end condition
1179 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1180 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test);
1181 register_new_node( new_bol0, pre_head->in(0) );
1182 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0);
1183 // Modify main loop guard condition
1184 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1185 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test);
1186 register_new_node( new_bol1, new_pre_exit );
1187 _igvn.hash_delete(min_iff);
1188 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1189 // Modify main loop end condition
1190 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1191 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test);
1192 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1193 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2);
1194 if (main_head->is_strip_mined()) {
1195 Node* le = outer_main_head->outer_loop_end();
1196 Node* bol = outer_main_head->outer_bol();
1197 Node* new_bol3 = new_bol2->clone();
1198 new_bol3->set_req(1, bol->in(1));
1199 register_new_node(new_bol3, le->in(0));
1200 _igvn.replace_input_of(le, 1, new_bol3);
1201 }
1202 }
1203
1204 // Flag main loop
1205 main_head->set_main_loop();
1206 if( peel_only ) main_head->set_main_no_pre_loop();
1207
1208 // Subtract a trip count for the pre-loop.
1209 main_head->set_trip_count(main_head->trip_count() - 1);
1210
1211 // It's difficult to be precise about the trip-counts
1212 // for the pre/post loops. They are usually very short,
1213 // so guess that 4 trips is a reasonable value.
1214 post_head->set_profile_trip_cnt(4.0);
1215 pre_head->set_profile_trip_cnt(4.0);
1216
1217 // Now force out all loop-invariant dominating tests. The optimizer
1218 // finds some, but we _know_ they are all useless.
1219 peeled_dom_test_elim(loop,old_new);
1220 loop->record_for_igvn();
1221 }
1222
1223 //------------------------------insert_vector_post_loop------------------------
1224 // Insert a copy of the atomic unrolled vectorized main loop as a post loop,
1225 // unroll_policy has already informed us that more unrolling is about to happen to
1226 // the main loop. The resultant post loop will serve as a vectorized drain loop.
1227 void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1228 if (!loop->_head->is_CountedLoop()) return;
1229
1230 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1231
1232 // only process vectorized main loops
1233 if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return;
1234
1235 int slp_max_unroll_factor = cl->slp_max_unroll();
1236 int cur_unroll = cl->unrolled_count();
1237
1238 if (slp_max_unroll_factor == 0) return;
1239
1240 // only process atomic unroll vector loops (not super unrolled after vectorization)
1241 if (cur_unroll != slp_max_unroll_factor) return;
1242
1243 // we only ever process this one time
1244 if (cl->has_atomic_post_loop()) return;
1245
1246 #ifndef PRODUCT
1247 if (TraceLoopOpts) {
1248 tty->print("PostVector ");
1249 loop->dump_head();
1250 }
1251 #endif
1252 C->set_major_progress();
1253
1254 // Find common pieces of the loop being guarded with pre & post loops
1255 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1256 CountedLoopEndNode *main_end = main_head->loopexit();
1257 guarantee(main_end != NULL, "no loop exit node");
1258 // diagnostic to show loop end is not properly formed
1259 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1260
1261 // mark this loop as processed
1262 main_head->mark_has_atomic_post_loop();
1263
1264 Node *incr = main_end->incr();
1265 Node *limit = main_end->limit();
1266
1267 // In this case we throw away the result as we are not using it to connect anything else.
1268 CountedLoopNode *post_head = NULL;
1269 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1270
1271 // It's difficult to be precise about the trip-counts
1272 // for post loops. They are usually very short,
1273 // so guess that unit vector trips is a reasonable value.
1274 post_head->set_profile_trip_cnt(cur_unroll);
1275
1276 // Now force out all loop-invariant dominating tests. The optimizer
1277 // finds some, but we _know_ they are all useless.
1278 peeled_dom_test_elim(loop, old_new);
1279 loop->record_for_igvn();
1280 }
1281
1282
1283 //-------------------------insert_scalar_rced_post_loop------------------------
1284 // Insert a copy of the rce'd main loop as a post loop,
1285 // We have not unrolled the main loop, so this is the right time to inject this.
1286 // Later we will examine the partner of this post loop pair which still has range checks
1287 // to see inject code which tests at runtime if the range checks are applicable.
1288 void PhaseIdealLoop::insert_scalar_rced_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1289 if (!loop->_head->is_CountedLoop()) return;
1290
1291 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1292
1293 // only process RCE'd main loops
1294 if (!cl->is_main_loop() || cl->range_checks_present()) return;
1295
1296 #ifndef PRODUCT
1297 if (TraceLoopOpts) {
1298 tty->print("PostScalarRce ");
1299 loop->dump_head();
1300 }
1301 #endif
1302 C->set_major_progress();
1303
1304 // Find common pieces of the loop being guarded with pre & post loops
1305 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1306 CountedLoopEndNode *main_end = main_head->loopexit();
1307 guarantee(main_end != NULL, "no loop exit node");
1308 // diagnostic to show loop end is not properly formed
1309 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1310
1311 Node *incr = main_end->incr();
1312 Node *limit = main_end->limit();
1313
1314 // In this case we throw away the result as we are not using it to connect anything else.
1315 CountedLoopNode *post_head = NULL;
1316 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1317
1318 // It's difficult to be precise about the trip-counts
1319 // for post loops. They are usually very short,
1320 // so guess that unit vector trips is a reasonable value.
1321 post_head->set_profile_trip_cnt(4.0);
1322 post_head->set_is_rce_post_loop();
1323
1324 // Now force out all loop-invariant dominating tests. The optimizer
1325 // finds some, but we _know_ they are all useless.
1326 peeled_dom_test_elim(loop, old_new);
1327 loop->record_for_igvn();
1328 }
1329
1330
1331 //------------------------------insert_post_loop-------------------------------
1332 // Insert post loops. Add a post loop to the given loop passed.
1333 Node *PhaseIdealLoop::insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
1334 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
1335 Node *incr, Node *limit, CountedLoopNode *&post_head) {
1336 IfNode* outer_main_end = main_end;
1337 IdealLoopTree* outer_loop = loop;
1338 if (main_head->is_strip_mined()) {
1339 main_head->verify_strip_mined(1);
1340 outer_main_end = main_head->outer_loop_end();
1341 outer_loop = loop->_parent;
1342 assert(outer_loop->_head == main_head->in(LoopNode::EntryControl), "broken loop tree");
1343 }
1344
1345 //------------------------------
1346 // Step A: Create a new post-Loop.
1347 Node* main_exit = outer_main_end->proj_out(false);
1348 assert(main_exit->Opcode() == Op_IfFalse, "");
1349 int dd_main_exit = dom_depth(main_exit);
1350
1351 // Step A1: Clone the loop body of main. The clone becomes the post-loop.
1352 // The main loop pre-header illegally has 2 control users (old & new loops).
1353 clone_loop(loop, old_new, dd_main_exit, ControlAroundStripMined);
1354 assert(old_new[main_end->_idx]->Opcode() == Op_CountedLoopEnd, "");
1355 post_head = old_new[main_head->_idx]->as_CountedLoop();
1356 post_head->set_normal_loop();
1357 post_head->set_post_loop(main_head);
1358
1359 // Reduce the post-loop trip count.
1360 CountedLoopEndNode* post_end = old_new[main_end->_idx]->as_CountedLoopEnd();
1361 post_end->_prob = PROB_FAIR;
1362
1363 // Build the main-loop normal exit.
1364 IfFalseNode *new_main_exit = new IfFalseNode(outer_main_end);
1365 _igvn.register_new_node_with_optimizer(new_main_exit);
1366 set_idom(new_main_exit, outer_main_end, dd_main_exit);
1367 set_loop(new_main_exit, outer_loop->_parent);
1368
1369 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
1370 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
1371 // (the previous loop trip-counter exit value) because we will be changing
1372 // the exit value (via additional unrolling) so we cannot constant-fold away the zero
1373 // trip guard until all unrolling is done.
1374 Node *zer_opaq = new Opaque1Node(C, incr);
1375 Node *zer_cmp = new CmpINode(zer_opaq, limit);
1376 Node *zer_bol = new BoolNode(zer_cmp, main_end->test_trip());
1377 register_new_node(zer_opaq, new_main_exit);
1378 register_new_node(zer_cmp, new_main_exit);
1379 register_new_node(zer_bol, new_main_exit);
1380
1381 // Build the IfNode
1382 IfNode *zer_iff = new IfNode(new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN);
1383 _igvn.register_new_node_with_optimizer(zer_iff);
1384 set_idom(zer_iff, new_main_exit, dd_main_exit);
1385 set_loop(zer_iff, outer_loop->_parent);
1386
1387 // Plug in the false-path, taken if we need to skip this post-loop
1388 _igvn.replace_input_of(main_exit, 0, zer_iff);
1389 set_idom(main_exit, zer_iff, dd_main_exit);
1390 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
1391 // Make the true-path, must enter this post loop
1392 Node *zer_taken = new IfTrueNode(zer_iff);
1393 _igvn.register_new_node_with_optimizer(zer_taken);
1394 set_idom(zer_taken, zer_iff, dd_main_exit);
1395 set_loop(zer_taken, outer_loop->_parent);
1396 // Plug in the true path
1397 _igvn.hash_delete(post_head);
1398 post_head->set_req(LoopNode::EntryControl, zer_taken);
1399 set_idom(post_head, zer_taken, dd_main_exit);
1400
1401 Arena *a = Thread::current()->resource_area();
1402 VectorSet visited(a);
1403 Node_Stack clones(a, main_head->back_control()->outcnt());
1404 // Step A3: Make the fall-in values to the post-loop come from the
1405 // fall-out values of the main-loop.
1406 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
1407 Node* main_phi = main_head->fast_out(i);
1408 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0) {
1409 Node *cur_phi = old_new[main_phi->_idx];
1410 Node *fallnew = clone_up_backedge_goo(main_head->back_control(),
1411 post_head->init_control(),
1412 main_phi->in(LoopNode::LoopBackControl),
1413 visited, clones);
1414 _igvn.hash_delete(cur_phi);
1415 cur_phi->set_req(LoopNode::EntryControl, fallnew);
1416 }
1417 }
1418
1419 // CastII for the new post loop:
1420 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1421 assert(inserted, "no castII inserted");
1422
1423 return new_main_exit;
1424 }
1425
1426 //------------------------------is_invariant-----------------------------
1427 // Return true if n is invariant
1428 bool IdealLoopTree::is_invariant(Node* n) const {
1429 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1430 if (n_c->is_top()) return false;
1431 return !is_member(_phase->get_loop(n_c));
1432 }
1433
1434
1435 //------------------------------do_unroll--------------------------------------
1436 // Unroll the loop body one step - make each trip do 2 iterations.
1437 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1438 assert(LoopUnrollLimit, "");
1439 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1440 CountedLoopEndNode *loop_end = loop_head->loopexit();
1441 assert(loop_end, "");
1442 #ifndef PRODUCT
1443 if (PrintOpto && VerifyLoopOptimizations) {
1444 tty->print("Unrolling ");
1445 loop->dump_head();
1446 } else if (TraceLoopOpts) {
1447 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1448 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1449 } else {
1450 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1451 }
1452 loop->dump_head();
1453 }
1454
1455 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
1456 Arena* arena = Thread::current()->resource_area();
1457 Node_Stack stack(arena, C->live_nodes() >> 2);
1458 Node_List rpo_list;
1459 VectorSet visited(arena);
1460 visited.set(loop_head->_idx);
1461 rpo( loop_head, stack, visited, rpo_list );
1462 dump(loop, rpo_list.size(), rpo_list );
1463 }
1464 #endif
1465
1466 // Remember loop node count before unrolling to detect
1467 // if rounds of unroll,optimize are making progress
1468 loop_head->set_node_count_before_unroll(loop->_body.size());
1469
1470 Node *ctrl = loop_head->skip_strip_mined()->in(LoopNode::EntryControl);
1471 Node *limit = loop_head->limit();
1472 Node *init = loop_head->init_trip();
1473 Node *stride = loop_head->stride();
1474
1475 Node *opaq = NULL;
1476 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1477 // Search for zero-trip guard.
1478
1479 // Check the shape of the graph at the loop entry. If an inappropriate
1480 // graph shape is encountered, the compiler bails out loop unrolling;
1481 // compilation of the method will still succeed.
1482 if (!is_canonical_loop_entry(loop_head)) {
1483 return;
1484 }
1485 opaq = ctrl->in(0)->in(1)->in(1)->in(2);
1486 // Zero-trip test uses an 'opaque' node which is not shared.
1487 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1488 }
1489
1490 C->set_major_progress();
1491
1492 Node* new_limit = NULL;
1493 int stride_con = stride->get_int();
1494 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1495 uint old_trip_count = loop_head->trip_count();
1496 // Verify that unroll policy result is still valid.
1497 assert(old_trip_count > 1 &&
1498 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1499
1500 // Adjust loop limit to keep valid iterations number after unroll.
1501 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1502 // which may overflow.
1503 if (!adjust_min_trip) {
1504 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1505 "odd trip count for maximally unroll");
1506 // Don't need to adjust limit for maximally unroll since trip count is even.
1507 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1508 // Loop's limit is constant. Loop's init could be constant when pre-loop
1509 // become peeled iteration.
1510 jlong init_con = init->get_int();
1511 // We can keep old loop limit if iterations count stays the same:
1512 // old_trip_count == new_trip_count * 2
1513 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1514 // so we also don't need to adjust zero trip test.
1515 jlong limit_con = limit->get_int();
1516 // (stride_con*2) not overflow since stride_con <= 8.
1517 int new_stride_con = stride_con * 2;
1518 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1519 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1520 // New trip count should satisfy next conditions.
1521 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1522 uint new_trip_count = (uint)trip_count;
1523 adjust_min_trip = (old_trip_count != new_trip_count*2);
1524 }
1525
1526 if (adjust_min_trip) {
1527 // Step 2: Adjust the trip limit if it is called for.
1528 // The adjustment amount is -stride. Need to make sure if the
1529 // adjustment underflows or overflows, then the main loop is skipped.
1530 Node* cmp = loop_end->cmp_node();
1531 assert(cmp->in(2) == limit, "sanity");
1532 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1533
1534 // Verify that policy_unroll result is still valid.
1535 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1536 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1537 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1538
1539 if (limit->is_Con()) {
1540 // The check in policy_unroll and the assert above guarantee
1541 // no underflow if limit is constant.
1542 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1543 set_ctrl(new_limit, C->root());
1544 } else {
1545 // Limit is not constant.
1546 if (loop_head->unrolled_count() == 1) { // only for first unroll
1547 // Separate limit by Opaque node in case it is an incremented
1548 // variable from previous loop to avoid using pre-incremented
1549 // value which could increase register pressure.
1550 // Otherwise reorg_offsets() optimization will create a separate
1551 // Opaque node for each use of trip-counter and as result
1552 // zero trip guard limit will be different from loop limit.
1553 assert(has_ctrl(opaq), "should have it");
1554 Node* opaq_ctrl = get_ctrl(opaq);
1555 limit = new Opaque2Node( C, limit );
1556 register_new_node( limit, opaq_ctrl );
1557 }
1558 if ((stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo)) ||
1559 (stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi))) {
1560 // No underflow.
1561 new_limit = new SubINode(limit, stride);
1562 } else {
1563 // (limit - stride) may underflow.
1564 // Clamp the adjustment value with MININT or MAXINT:
1565 //
1566 // new_limit = limit-stride
1567 // if (stride > 0)
1568 // new_limit = (limit < new_limit) ? MININT : new_limit;
1569 // else
1570 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1571 //
1572 BoolTest::mask bt = loop_end->test_trip();
1573 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1574 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1575 set_ctrl(adj_max, C->root());
1576 Node* old_limit = NULL;
1577 Node* adj_limit = NULL;
1578 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1579 if (loop_head->unrolled_count() > 1 &&
1580 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1581 limit->in(CMoveNode::IfTrue) == adj_max &&
1582 bol->as_Bool()->_test._test == bt &&
1583 bol->in(1)->Opcode() == Op_CmpI &&
1584 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1585 // Loop was unrolled before.
1586 // Optimize the limit to avoid nested CMove:
1587 // use original limit as old limit.
1588 old_limit = bol->in(1)->in(1);
1589 // Adjust previous adjusted limit.
1590 adj_limit = limit->in(CMoveNode::IfFalse);
1591 adj_limit = new SubINode(adj_limit, stride);
1592 } else {
1593 old_limit = limit;
1594 adj_limit = new SubINode(limit, stride);
1595 }
1596 assert(old_limit != NULL && adj_limit != NULL, "");
1597 register_new_node( adj_limit, ctrl ); // adjust amount
1598 Node* adj_cmp = new CmpINode(old_limit, adj_limit);
1599 register_new_node( adj_cmp, ctrl );
1600 Node* adj_bool = new BoolNode(adj_cmp, bt);
1601 register_new_node( adj_bool, ctrl );
1602 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1603 }
1604 register_new_node(new_limit, ctrl);
1605 }
1606 assert(new_limit != NULL, "");
1607 // Replace in loop test.
1608 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1609 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1610 // Don't need to create new test since only one user.
1611 _igvn.hash_delete(cmp);
1612 cmp->set_req(2, new_limit);
1613 } else {
1614 // Create new test since it is shared.
1615 Node* ctrl2 = loop_end->in(0);
1616 Node* cmp2 = cmp->clone();
1617 cmp2->set_req(2, new_limit);
1618 register_new_node(cmp2, ctrl2);
1619 Node* bol2 = loop_end->in(1)->clone();
1620 bol2->set_req(1, cmp2);
1621 register_new_node(bol2, ctrl2);
1622 _igvn.replace_input_of(loop_end, 1, bol2);
1623 }
1624 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1625 // Make it a 1-trip test (means at least 2 trips).
1626
1627 // Guard test uses an 'opaque' node which is not shared. Hence I
1628 // can edit it's inputs directly. Hammer in the new limit for the
1629 // minimum-trip guard.
1630 assert(opaq->outcnt() == 1, "");
1631 _igvn.replace_input_of(opaq, 1, new_limit);
1632 }
1633
1634 // Adjust max trip count. The trip count is intentionally rounded
1635 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1636 // the main, unrolled, part of the loop will never execute as it is protected
1637 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1638 // and later determined that part of the unrolled loop was dead.
1639 loop_head->set_trip_count(old_trip_count / 2);
1640
1641 // Double the count of original iterations in the unrolled loop body.
1642 loop_head->double_unrolled_count();
1643
1644 // ---------
1645 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1646 // represents the odd iterations; since the loop trips an even number of
1647 // times its backedge is never taken. Kill the backedge.
1648 uint dd = dom_depth(loop_head);
1649 clone_loop(loop, old_new, dd, IgnoreStripMined);
1650
1651 // Make backedges of the clone equal to backedges of the original.
1652 // Make the fall-in from the original come from the fall-out of the clone.
1653 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1654 Node* phi = loop_head->fast_out(j);
1655 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1656 Node *newphi = old_new[phi->_idx];
1657 _igvn.hash_delete( phi );
1658 _igvn.hash_delete( newphi );
1659
1660 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1661 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1662 phi ->set_req(LoopNode::LoopBackControl, C->top());
1663 }
1664 }
1665 Node *clone_head = old_new[loop_head->_idx];
1666 _igvn.hash_delete( clone_head );
1667 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1668 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1669 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1670 loop->_head = clone_head; // New loop header
1671
1672 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1673 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1674
1675 // Kill the clone's backedge
1676 Node *newcle = old_new[loop_end->_idx];
1677 _igvn.hash_delete( newcle );
1678 Node *one = _igvn.intcon(1);
1679 set_ctrl(one, C->root());
1680 newcle->set_req(1, one);
1681 // Force clone into same loop body
1682 uint max = loop->_body.size();
1683 for( uint k = 0; k < max; k++ ) {
1684 Node *old = loop->_body.at(k);
1685 Node *nnn = old_new[old->_idx];
1686 loop->_body.push(nnn);
1687 if (!has_ctrl(old))
1688 set_loop(nnn, loop);
1689 }
1690
1691 loop->record_for_igvn();
1692 loop_head->clear_strip_mined();
1693
1694 #ifndef PRODUCT
1695 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
1696 tty->print("\nnew loop after unroll\n"); loop->dump_head();
1697 for (uint i = 0; i < loop->_body.size(); i++) {
1698 loop->_body.at(i)->dump();
1699 }
1700 if(C->clone_map().is_debug()) {
1701 tty->print("\nCloneMap\n");
1702 Dict* dict = C->clone_map().dict();
1703 DictI i(dict);
1704 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size());
1705 for (int ii = 0; i.test(); ++i, ++ii) {
1706 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key));
1707 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen());
1708 if (ii % 10 == 9) {
1709 tty->print_cr(" ");
1710 }
1711 }
1712 tty->print_cr(" ");
1713 }
1714 }
1715 #endif
1716
1717 }
1718
1719 //------------------------------do_maximally_unroll----------------------------
1720
1721 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1722 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1723 assert(cl->has_exact_trip_count(), "trip count is not exact");
1724 assert(cl->trip_count() > 0, "");
1725 #ifndef PRODUCT
1726 if (TraceLoopOpts) {
1727 tty->print("MaxUnroll %d ", cl->trip_count());
1728 loop->dump_head();
1729 }
1730 #endif
1731
1732 // If loop is tripping an odd number of times, peel odd iteration
1733 if ((cl->trip_count() & 1) == 1) {
1734 do_peeling(loop, old_new);
1735 }
1736
1737 // Now its tripping an even number of times remaining. Double loop body.
1738 // Do not adjust pre-guards; they are not needed and do not exist.
1739 if (cl->trip_count() > 0) {
1740 assert((cl->trip_count() & 1) == 0, "missed peeling");
1741 do_unroll(loop, old_new, false);
1742 }
1743 }
1744
1745 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) {
1746 if (SuperWordReductions == false) return;
1747
1748 CountedLoopNode* loop_head = loop->_head->as_CountedLoop();
1749 if (loop_head->unrolled_count() > 1) {
1750 return;
1751 }
1752
1753 Node* trip_phi = loop_head->phi();
1754 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) {
1755 Node* phi = loop_head->fast_out(i);
1756 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) {
1757 // For definitions which are loop inclusive and not tripcounts.
1758 Node* def_node = phi->in(LoopNode::LoopBackControl);
1759
1760 if (def_node != NULL) {
1761 Node* n_ctrl = get_ctrl(def_node);
1762 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) {
1763 // Now test it to see if it fits the standard pattern for a reduction operator.
1764 int opc = def_node->Opcode();
1765 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) {
1766 if (!def_node->is_reduction()) { // Not marked yet
1767 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it
1768 bool ok = false;
1769 for (unsigned j = 1; j < def_node->req(); j++) {
1770 Node* in = def_node->in(j);
1771 if (in == phi) {
1772 ok = true;
1773 break;
1774 }
1775 }
1776
1777 // do nothing if we did not match the initial criteria
1778 if (ok == false) {
1779 continue;
1780 }
1781
1782 // The result of the reduction must not be used in the loop
1783 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) {
1784 Node* u = def_node->fast_out(i);
1785 if (!loop->is_member(get_loop(ctrl_or_self(u)))) {
1786 continue;
1787 }
1788 if (u == phi) {
1789 continue;
1790 }
1791 ok = false;
1792 }
1793
1794 // iff the uses conform
1795 if (ok) {
1796 def_node->add_flag(Node::Flag_is_reduction);
1797 loop_head->mark_has_reductions();
1798 }
1799 }
1800 }
1801 }
1802 }
1803 }
1804 }
1805 }
1806
1807 //------------------------------adjust_limit-----------------------------------
1808 // Helper function for add_constraint().
1809 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1810 // Compute "I :: (limit-offset)/scale"
1811 Node *con = new SubINode(rc_limit, offset);
1812 register_new_node(con, pre_ctrl);
1813 Node *X = new DivINode(0, con, scale);
1814 register_new_node(X, pre_ctrl);
1815
1816 // Adjust loop limit
1817 loop_limit = (stride_con > 0)
1818 ? (Node*)(new MinINode(loop_limit, X))
1819 : (Node*)(new MaxINode(loop_limit, X));
1820 register_new_node(loop_limit, pre_ctrl);
1821 return loop_limit;
1822 }
1823
1824 //------------------------------add_constraint---------------------------------
1825 // Constrain the main loop iterations so the conditions:
1826 // low_limit <= scale_con * I + offset < upper_limit
1827 // always holds true. That is, either increase the number of iterations in
1828 // the pre-loop or the post-loop until the condition holds true in the main
1829 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1830 // stride and scale are constants (offset and limit often are).
1831 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
1832 // For positive stride, the pre-loop limit always uses a MAX function
1833 // and the main loop a MIN function. For negative stride these are
1834 // reversed.
1835
1836 // Also for positive stride*scale the affine function is increasing, so the
1837 // pre-loop must check for underflow and the post-loop for overflow.
1838 // Negative stride*scale reverses this; pre-loop checks for overflow and
1839 // post-loop for underflow.
1840
1841 Node *scale = _igvn.intcon(scale_con);
1842 set_ctrl(scale, C->root());
1843
1844 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1845 // The overflow limit: scale*I+offset < upper_limit
1846 // For main-loop compute
1847 // ( if (scale > 0) /* and stride > 0 */
1848 // I < (upper_limit-offset)/scale
1849 // else /* scale < 0 and stride < 0 */
1850 // I > (upper_limit-offset)/scale
1851 // )
1852 //
1853 // (upper_limit-offset) may overflow or underflow.
1854 // But it is fine since main loop will either have
1855 // less iterations or will be skipped in such case.
1856 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1857
1858 // The underflow limit: low_limit <= scale*I+offset.
1859 // For pre-loop compute
1860 // NOT(scale*I+offset >= low_limit)
1861 // scale*I+offset < low_limit
1862 // ( if (scale > 0) /* and stride > 0 */
1863 // I < (low_limit-offset)/scale
1864 // else /* scale < 0 and stride < 0 */
1865 // I > (low_limit-offset)/scale
1866 // )
1867
1868 if (low_limit->get_int() == -max_jint) {
1869 // We need this guard when scale*pre_limit+offset >= limit
1870 // due to underflow. So we need execute pre-loop until
1871 // scale*I+offset >= min_int. But (min_int-offset) will
1872 // underflow when offset > 0 and X will be > original_limit
1873 // when stride > 0. To avoid it we replace positive offset with 0.
1874 //
1875 // Also (min_int+1 == -max_int) is used instead of min_int here
1876 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1877 Node* shift = _igvn.intcon(31);
1878 set_ctrl(shift, C->root());
1879 Node* sign = new RShiftINode(offset, shift);
1880 register_new_node(sign, pre_ctrl);
1881 offset = new AndINode(offset, sign);
1882 register_new_node(offset, pre_ctrl);
1883 } else {
1884 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1885 // The only problem we have here when offset == min_int
1886 // since (0-min_int) == min_int. It may be fine for stride > 0
1887 // but for stride < 0 X will be < original_limit. To avoid it
1888 // max(pre_limit, original_limit) is used in do_range_check().
1889 }
1890 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1891 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1892
1893 } else { // stride_con*scale_con < 0
1894 // For negative stride*scale pre-loop checks for overflow and
1895 // post-loop for underflow.
1896 //
1897 // The overflow limit: scale*I+offset < upper_limit
1898 // For pre-loop compute
1899 // NOT(scale*I+offset < upper_limit)
1900 // scale*I+offset >= upper_limit
1901 // scale*I+offset+1 > upper_limit
1902 // ( if (scale < 0) /* and stride > 0 */
1903 // I < (upper_limit-(offset+1))/scale
1904 // else /* scale > 0 and stride < 0 */
1905 // I > (upper_limit-(offset+1))/scale
1906 // )
1907 //
1908 // (upper_limit-offset-1) may underflow or overflow.
1909 // To avoid it min(pre_limit, original_limit) is used
1910 // in do_range_check() for stride > 0 and max() for < 0.
1911 Node *one = _igvn.intcon(1);
1912 set_ctrl(one, C->root());
1913
1914 Node *plus_one = new AddINode(offset, one);
1915 register_new_node( plus_one, pre_ctrl );
1916 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1917 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1918
1919 if (low_limit->get_int() == -max_jint) {
1920 // We need this guard when scale*main_limit+offset >= limit
1921 // due to underflow. So we need execute main-loop while
1922 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1923 // underflow when (offset+1) > 0 and X will be < main_limit
1924 // when scale < 0 (and stride > 0). To avoid it we replace
1925 // positive (offset+1) with 0.
1926 //
1927 // Also (min_int+1 == -max_int) is used instead of min_int here
1928 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1929 Node* shift = _igvn.intcon(31);
1930 set_ctrl(shift, C->root());
1931 Node* sign = new RShiftINode(plus_one, shift);
1932 register_new_node(sign, pre_ctrl);
1933 plus_one = new AndINode(plus_one, sign);
1934 register_new_node(plus_one, pre_ctrl);
1935 } else {
1936 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1937 // The only problem we have here when offset == max_int
1938 // since (max_int+1) == min_int and (0-min_int) == min_int.
1939 // But it is fine since main loop will either have
1940 // less iterations or will be skipped in such case.
1941 }
1942 // The underflow limit: low_limit <= scale*I+offset.
1943 // For main-loop compute
1944 // scale*I+offset+1 > low_limit
1945 // ( if (scale < 0) /* and stride > 0 */
1946 // I < (low_limit-(offset+1))/scale
1947 // else /* scale > 0 and stride < 0 */
1948 // I > (low_limit-(offset+1))/scale
1949 // )
1950
1951 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1952 }
1953 }
1954
1955
1956 //------------------------------is_scaled_iv---------------------------------
1957 // Return true if exp is a constant times an induction var
1958 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1959 if (exp == iv) {
1960 if (p_scale != NULL) {
1961 *p_scale = 1;
1962 }
1963 return true;
1964 }
1965 int opc = exp->Opcode();
1966 if (opc == Op_MulI) {
1967 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1968 if (p_scale != NULL) {
1969 *p_scale = exp->in(2)->get_int();
1970 }
1971 return true;
1972 }
1973 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1974 if (p_scale != NULL) {
1975 *p_scale = exp->in(1)->get_int();
1976 }
1977 return true;
1978 }
1979 } else if (opc == Op_LShiftI) {
1980 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1981 if (p_scale != NULL) {
1982 *p_scale = 1 << exp->in(2)->get_int();
1983 }
1984 return true;
1985 }
1986 }
1987 return false;
1988 }
1989
1990 //-----------------------------is_scaled_iv_plus_offset------------------------------
1991 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1992 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1993 if (is_scaled_iv(exp, iv, p_scale)) {
1994 if (p_offset != NULL) {
1995 Node *zero = _igvn.intcon(0);
1996 set_ctrl(zero, C->root());
1997 *p_offset = zero;
1998 }
1999 return true;
2000 }
2001 int opc = exp->Opcode();
2002 if (opc == Op_AddI) {
2003 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
2004 if (p_offset != NULL) {
2005 *p_offset = exp->in(2);
2006 }
2007 return true;
2008 }
2009 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
2010 if (p_offset != NULL) {
2011 *p_offset = exp->in(1);
2012 }
2013 return true;
2014 }
2015 if (exp->in(2)->is_Con()) {
2016 Node* offset2 = NULL;
2017 if (depth < 2 &&
2018 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
2019 p_offset != NULL ? &offset2 : NULL, depth+1)) {
2020 if (p_offset != NULL) {
2021 Node *ctrl_off2 = get_ctrl(offset2);
2022 Node* offset = new AddINode(offset2, exp->in(2));
2023 register_new_node(offset, ctrl_off2);
2024 *p_offset = offset;
2025 }
2026 return true;
2027 }
2028 }
2029 } else if (opc == Op_SubI) {
2030 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
2031 if (p_offset != NULL) {
2032 Node *zero = _igvn.intcon(0);
2033 set_ctrl(zero, C->root());
2034 Node *ctrl_off = get_ctrl(exp->in(2));
2035 Node* offset = new SubINode(zero, exp->in(2));
2036 register_new_node(offset, ctrl_off);
2037 *p_offset = offset;
2038 }
2039 return true;
2040 }
2041 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
2042 if (p_offset != NULL) {
2043 *p_scale *= -1;
2044 *p_offset = exp->in(1);
2045 }
2046 return true;
2047 }
2048 }
2049 return false;
2050 }
2051
2052 //------------------------------do_range_check---------------------------------
2053 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
2054 int PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
2055 #ifndef PRODUCT
2056 if (PrintOpto && VerifyLoopOptimizations) {
2057 tty->print("Range Check Elimination ");
2058 loop->dump_head();
2059 } else if (TraceLoopOpts) {
2060 tty->print("RangeCheck ");
2061 loop->dump_head();
2062 }
2063 #endif
2064 assert(RangeCheckElimination, "");
2065 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2066 // If we fail before trying to eliminate range checks, set multiversion state
2067 int closed_range_checks = 1;
2068
2069 // protect against stride not being a constant
2070 if (!cl->stride_is_con())
2071 return closed_range_checks;
2072
2073 // Find the trip counter; we are iteration splitting based on it
2074 Node *trip_counter = cl->phi();
2075 // Find the main loop limit; we will trim it's iterations
2076 // to not ever trip end tests
2077 Node *main_limit = cl->limit();
2078
2079 // Check graph shape. Cannot optimize a loop if zero-trip
2080 // Opaque1 node is optimized away and then another round
2081 // of loop opts attempted.
2082 if (!is_canonical_loop_entry(cl)) {
2083 return closed_range_checks;
2084 }
2085
2086 // Need to find the main-loop zero-trip guard
2087 Node *ctrl = cl->skip_strip_mined()->in(LoopNode::EntryControl);
2088 Node *iffm = ctrl->in(0);
2089 Node *opqzm = iffm->in(1)->in(1)->in(2);
2090 assert(opqzm->in(1) == main_limit, "do not understand situation");
2091
2092 // Find the pre-loop limit; we will expand its iterations to
2093 // not ever trip low tests.
2094 Node *p_f = iffm->in(0);
2095 // pre loop may have been optimized out
2096 if (p_f->Opcode() != Op_IfFalse) {
2097 return closed_range_checks;
2098 }
2099 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2100 assert(pre_end->loopnode()->is_pre_loop(), "");
2101 Node *pre_opaq1 = pre_end->limit();
2102 // Occasionally it's possible for a pre-loop Opaque1 node to be
2103 // optimized away and then another round of loop opts attempted.
2104 // We can not optimize this particular loop in that case.
2105 if (pre_opaq1->Opcode() != Op_Opaque1)
2106 return closed_range_checks;
2107 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2108 Node *pre_limit = pre_opaq->in(1);
2109
2110 // Where do we put new limit calculations
2111 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2112
2113 // Ensure the original loop limit is available from the
2114 // pre-loop Opaque1 node.
2115 Node *orig_limit = pre_opaq->original_loop_limit();
2116 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
2117 return closed_range_checks;
2118
2119 // Must know if its a count-up or count-down loop
2120
2121 int stride_con = cl->stride_con();
2122 Node *zero = _igvn.intcon(0);
2123 Node *one = _igvn.intcon(1);
2124 // Use symmetrical int range [-max_jint,max_jint]
2125 Node *mini = _igvn.intcon(-max_jint);
2126 set_ctrl(zero, C->root());
2127 set_ctrl(one, C->root());
2128 set_ctrl(mini, C->root());
2129
2130 // Range checks that do not dominate the loop backedge (ie.
2131 // conditionally executed) can lengthen the pre loop limit beyond
2132 // the original loop limit. To prevent this, the pre limit is
2133 // (for stride > 0) MINed with the original loop limit (MAXed
2134 // stride < 0) when some range_check (rc) is conditionally
2135 // executed.
2136 bool conditional_rc = false;
2137
2138 // Count number of range checks and reduce by load range limits, if zero,
2139 // the loop is in canonical form to multiversion.
2140 closed_range_checks = 0;
2141
2142 // Check loop body for tests of trip-counter plus loop-invariant vs loop-variant.
2143 for( uint i = 0; i < loop->_body.size(); i++ ) {
2144 Node *iff = loop->_body[i];
2145 if (iff->Opcode() == Op_If ||
2146 iff->Opcode() == Op_RangeCheck) { // Test?
2147 // Test is an IfNode, has 2 projections. If BOTH are in the loop
2148 // we need loop unswitching instead of iteration splitting.
2149 closed_range_checks++;
2150 Node *exit = loop->is_loop_exit(iff);
2151 if( !exit ) continue;
2152 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
2153
2154 // Get boolean condition to test
2155 Node *i1 = iff->in(1);
2156 if( !i1->is_Bool() ) continue;
2157 BoolNode *bol = i1->as_Bool();
2158 BoolTest b_test = bol->_test;
2159 // Flip sense of test if exit condition is flipped
2160 if( flip )
2161 b_test = b_test.negate();
2162
2163 // Get compare
2164 Node *cmp = bol->in(1);
2165
2166 // Look for trip_counter + offset vs limit
2167 Node *rc_exp = cmp->in(1);
2168 Node *limit = cmp->in(2);
2169 jint scale_con= 1; // Assume trip counter not scaled
2170
2171 Node *limit_c = get_ctrl(limit);
2172 if( loop->is_member(get_loop(limit_c) ) ) {
2173 // Compare might have operands swapped; commute them
2174 b_test = b_test.commute();
2175 rc_exp = cmp->in(2);
2176 limit = cmp->in(1);
2177 limit_c = get_ctrl(limit);
2178 if( loop->is_member(get_loop(limit_c) ) )
2179 continue; // Both inputs are loop varying; cannot RCE
2180 }
2181 // Here we know 'limit' is loop invariant
2182
2183 // 'limit' maybe pinned below the zero trip test (probably from a
2184 // previous round of rce), in which case, it can't be used in the
2185 // zero trip test expression which must occur before the zero test's if.
2186 if( limit_c == ctrl ) {
2187 continue; // Don't rce this check but continue looking for other candidates.
2188 }
2189
2190 // Check for scaled induction variable plus an offset
2191 Node *offset = NULL;
2192
2193 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
2194 continue;
2195 }
2196
2197 Node *offset_c = get_ctrl(offset);
2198 if( loop->is_member( get_loop(offset_c) ) )
2199 continue; // Offset is not really loop invariant
2200 // Here we know 'offset' is loop invariant.
2201
2202 // As above for the 'limit', the 'offset' maybe pinned below the
2203 // zero trip test.
2204 if( offset_c == ctrl ) {
2205 continue; // Don't rce this check but continue looking for other candidates.
2206 }
2207 #ifdef ASSERT
2208 if (TraceRangeLimitCheck) {
2209 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
2210 bol->dump(2);
2211 }
2212 #endif
2213 // At this point we have the expression as:
2214 // scale_con * trip_counter + offset :: limit
2215 // where scale_con, offset and limit are loop invariant. Trip_counter
2216 // monotonically increases by stride_con, a constant. Both (or either)
2217 // stride_con and scale_con can be negative which will flip about the
2218 // sense of the test.
2219
2220 // Adjust pre and main loop limits to guard the correct iteration set
2221 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
2222 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
2223 // The underflow and overflow limits: 0 <= scale*I+offset < limit
2224 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
2225 // (0-offset)/scale could be outside of loop iterations range.
2226 conditional_rc = true;
2227 } else {
2228 if (PrintOpto) {
2229 tty->print_cr("missed RCE opportunity");
2230 }
2231 continue; // In release mode, ignore it
2232 }
2233 } else { // Otherwise work on normal compares
2234 switch( b_test._test ) {
2235 case BoolTest::gt:
2236 // Fall into GE case
2237 case BoolTest::ge:
2238 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
2239 scale_con = -scale_con;
2240 offset = new SubINode( zero, offset );
2241 register_new_node( offset, pre_ctrl );
2242 limit = new SubINode( zero, limit );
2243 register_new_node( limit, pre_ctrl );
2244 // Fall into LE case
2245 case BoolTest::le:
2246 if (b_test._test != BoolTest::gt) {
2247 // Convert X <= Y to X < Y+1
2248 limit = new AddINode( limit, one );
2249 register_new_node( limit, pre_ctrl );
2250 }
2251 // Fall into LT case
2252 case BoolTest::lt:
2253 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
2254 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
2255 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
2256 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
2257 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
2258 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
2259 // still be outside of loop range.
2260 conditional_rc = true;
2261 break;
2262 default:
2263 if (PrintOpto) {
2264 tty->print_cr("missed RCE opportunity");
2265 }
2266 continue; // Unhandled case
2267 }
2268 }
2269
2270 // Kill the eliminated test
2271 C->set_major_progress();
2272 Node *kill_con = _igvn.intcon( 1-flip );
2273 set_ctrl(kill_con, C->root());
2274 _igvn.replace_input_of(iff, 1, kill_con);
2275 // Find surviving projection
2276 assert(iff->is_If(), "");
2277 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2278 // Find loads off the surviving projection; remove their control edge
2279 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2280 Node* cd = dp->fast_out(i); // Control-dependent node
2281 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop
2282 // Allow the load to float around in the loop, or before it
2283 // but NOT before the pre-loop.
2284 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
2285 --i;
2286 --imax;
2287 }
2288 }
2289 if (limit->Opcode() == Op_LoadRange) {
2290 closed_range_checks--;
2291 }
2292
2293 } // End of is IF
2294
2295 }
2296
2297 // Update loop limits
2298 if (conditional_rc) {
2299 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit)
2300 : (Node*)new MaxINode(pre_limit, orig_limit);
2301 register_new_node(pre_limit, pre_ctrl);
2302 }
2303 _igvn.replace_input_of(pre_opaq, 1, pre_limit);
2304
2305 // Note:: we are making the main loop limit no longer precise;
2306 // need to round up based on stride.
2307 cl->set_nonexact_trip_count();
2308 Node *main_cle = cl->loopexit();
2309 Node *main_bol = main_cle->in(1);
2310 // Hacking loop bounds; need private copies of exit test
2311 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2312 main_bol = main_bol->clone();// Clone a private BoolNode
2313 register_new_node( main_bol, main_cle->in(0) );
2314 _igvn.replace_input_of(main_cle, 1, main_bol);
2315 }
2316 Node *main_cmp = main_bol->in(1);
2317 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2318 main_cmp = main_cmp->clone();// Clone a private CmpNode
2319 register_new_node( main_cmp, main_cle->in(0) );
2320 _igvn.replace_input_of(main_bol, 1, main_cmp);
2321 }
2322 // Hack the now-private loop bounds
2323 _igvn.replace_input_of(main_cmp, 2, main_limit);
2324 // The OpaqueNode is unshared by design
2325 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2326 _igvn.replace_input_of(opqzm, 1, main_limit);
2327
2328 return closed_range_checks;
2329 }
2330
2331 //------------------------------has_range_checks-------------------------------
2332 // Check to see if RCE cleaned the current loop of range-checks.
2333 void PhaseIdealLoop::has_range_checks(IdealLoopTree *loop) {
2334 assert(RangeCheckElimination, "");
2335
2336 // skip if not a counted loop
2337 if (!loop->is_counted()) return;
2338
2339 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2340
2341 // skip this loop if it is already checked
2342 if (cl->has_been_range_checked()) return;
2343
2344 // Now check for existence of range checks
2345 for (uint i = 0; i < loop->_body.size(); i++) {
2346 Node *iff = loop->_body[i];
2347 int iff_opc = iff->Opcode();
2348 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
2349 cl->mark_has_range_checks();
2350 break;
2351 }
2352 }
2353 cl->set_has_been_range_checked();
2354 }
2355
2356 //-------------------------multi_version_post_loops----------------------------
2357 // Check the range checks that remain, if simple, use the bounds to guard
2358 // which version to a post loop we execute, one with range checks or one without
2359 bool PhaseIdealLoop::multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop) {
2360 bool multi_version_succeeded = false;
2361 assert(RangeCheckElimination, "");
2362 CountedLoopNode *legacy_cl = legacy_loop->_head->as_CountedLoop();
2363 assert(legacy_cl->is_post_loop(), "");
2364
2365 // Check for existence of range checks using the unique instance to make a guard with
2366 Unique_Node_List worklist;
2367 for (uint i = 0; i < legacy_loop->_body.size(); i++) {
2368 Node *iff = legacy_loop->_body[i];
2369 int iff_opc = iff->Opcode();
2370 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
2371 worklist.push(iff);
2372 }
2373 }
2374
2375 // Find RCE'd post loop so that we can stage its guard.
2376 if (!is_canonical_loop_entry(legacy_cl)) return multi_version_succeeded;
2377 Node* ctrl = legacy_cl->in(LoopNode::EntryControl);
2378 Node* iffm = ctrl->in(0);
2379
2380 // Now we test that both the post loops are connected
2381 Node* post_loop_region = iffm->in(0);
2382 if (post_loop_region == NULL) return multi_version_succeeded;
2383 if (!post_loop_region->is_Region()) return multi_version_succeeded;
2384 Node* covering_region = post_loop_region->in(RegionNode::Control+1);
2385 if (covering_region == NULL) return multi_version_succeeded;
2386 if (!covering_region->is_Region()) return multi_version_succeeded;
2387 Node* p_f = covering_region->in(RegionNode::Control);
2388 if (p_f == NULL) return multi_version_succeeded;
2389 if (!p_f->is_IfFalse()) return multi_version_succeeded;
2390 if (!p_f->in(0)->is_CountedLoopEnd()) return multi_version_succeeded;
2391 CountedLoopEndNode* rce_loop_end = p_f->in(0)->as_CountedLoopEnd();
2392 if (rce_loop_end == NULL) return multi_version_succeeded;
2393 CountedLoopNode* rce_cl = rce_loop_end->loopnode();
2394 if (rce_cl == NULL || !rce_cl->is_post_loop()) return multi_version_succeeded;
2395 CountedLoopNode *known_rce_cl = rce_loop->_head->as_CountedLoop();
2396 if (rce_cl != known_rce_cl) return multi_version_succeeded;
2397
2398 // Then we fetch the cover entry test
2399 ctrl = rce_cl->in(LoopNode::EntryControl);
2400 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return multi_version_succeeded;
2401
2402 #ifndef PRODUCT
2403 if (TraceLoopOpts) {
2404 tty->print("PostMultiVersion\n");
2405 rce_loop->dump_head();
2406 legacy_loop->dump_head();
2407 }
2408 #endif
2409
2410 // Now fetch the limit we want to compare against
2411 Node *limit = rce_cl->limit();
2412 bool first_time = true;
2413
2414 // If we got this far, we identified the post loop which has been RCE'd and
2415 // we have a work list. Now we will try to transform the if guard to cause
2416 // the loop pair to be multi version executed with the determination left to runtime
2417 // or the optimizer if full information is known about the given arrays at compile time.
2418 Node *last_min = NULL;
2419 multi_version_succeeded = true;
2420 while (worklist.size()) {
2421 Node* rc_iffm = worklist.pop();
2422 if (rc_iffm->is_If()) {
2423 Node *rc_bolzm = rc_iffm->in(1);
2424 if (rc_bolzm->is_Bool()) {
2425 Node *rc_cmpzm = rc_bolzm->in(1);
2426 if (rc_cmpzm->is_Cmp()) {
2427 Node *rc_left = rc_cmpzm->in(2);
2428 if (rc_left->Opcode() != Op_LoadRange) {
2429 multi_version_succeeded = false;
2430 break;
2431 }
2432 if (first_time) {
2433 last_min = rc_left;
2434 first_time = false;
2435 } else {
2436 Node *cur_min = new MinINode(last_min, rc_left);
2437 last_min = cur_min;
2438 _igvn.register_new_node_with_optimizer(last_min);
2439 }
2440 }
2441 }
2442 }
2443 }
2444
2445 // All we have to do is update the limit of the rce loop
2446 // with the min of our expression and the current limit.
2447 // We will use this expression to replace the current limit.
2448 if (last_min && multi_version_succeeded) {
2449 Node *cur_min = new MinINode(last_min, limit);
2450 _igvn.register_new_node_with_optimizer(cur_min);
2451 Node *cmp_node = rce_loop_end->cmp_node();
2452 _igvn.replace_input_of(cmp_node, 2, cur_min);
2453 set_ctrl(cur_min, ctrl);
2454 set_loop(cur_min, rce_loop->_parent);
2455
2456 legacy_cl->mark_is_multiversioned();
2457 rce_cl->mark_is_multiversioned();
2458 multi_version_succeeded = true;
2459
2460 C->set_major_progress();
2461 }
2462
2463 return multi_version_succeeded;
2464 }
2465
2466 //-------------------------poison_rce_post_loop--------------------------------
2467 // Causes the rce'd post loop to be optimized away if multiversioning fails
2468 void PhaseIdealLoop::poison_rce_post_loop(IdealLoopTree *rce_loop) {
2469 CountedLoopNode *rce_cl = rce_loop->_head->as_CountedLoop();
2470 Node* ctrl = rce_cl->in(LoopNode::EntryControl);
2471 if (ctrl->is_IfTrue() || ctrl->is_IfFalse()) {
2472 Node* iffm = ctrl->in(0);
2473 if (iffm->is_If()) {
2474 Node* cur_bool = iffm->in(1);
2475 if (cur_bool->is_Bool()) {
2476 Node* cur_cmp = cur_bool->in(1);
2477 if (cur_cmp->is_Cmp()) {
2478 BoolTest::mask new_test = BoolTest::gt;
2479 BoolNode *new_bool = new BoolNode(cur_cmp, new_test);
2480 _igvn.replace_node(cur_bool, new_bool);
2481 _igvn._worklist.push(new_bool);
2482 Node* left_op = cur_cmp->in(1);
2483 _igvn.replace_input_of(cur_cmp, 2, left_op);
2484 C->set_major_progress();
2485 }
2486 }
2487 }
2488 }
2489 }
2490
2491 //------------------------------DCE_loop_body----------------------------------
2492 // Remove simplistic dead code from loop body
2493 void IdealLoopTree::DCE_loop_body() {
2494 for( uint i = 0; i < _body.size(); i++ )
2495 if( _body.at(i)->outcnt() == 0 )
2496 _body.map( i--, _body.pop() );
2497 }
2498
2499
2500 //------------------------------adjust_loop_exit_prob--------------------------
2501 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2502 // Replace with a 1-in-10 exit guess.
2503 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2504 Node *test = tail();
2505 while( test != _head ) {
2506 uint top = test->Opcode();
2507 if( top == Op_IfTrue || top == Op_IfFalse ) {
2508 int test_con = ((ProjNode*)test)->_con;
2509 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2510 IfNode *iff = test->in(0)->as_If();
2511 if( iff->outcnt() == 2 ) { // Ignore dead tests
2512 Node *bol = iff->in(1);
2513 if( bol && bol->req() > 1 && bol->in(1) &&
2514 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2515 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2516 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2517 (bol->in(1)->Opcode() == Op_CompareAndExchangeB ) ||
2518 (bol->in(1)->Opcode() == Op_CompareAndExchangeS ) ||
2519 (bol->in(1)->Opcode() == Op_CompareAndExchangeI ) ||
2520 (bol->in(1)->Opcode() == Op_CompareAndExchangeL ) ||
2521 (bol->in(1)->Opcode() == Op_CompareAndExchangeP ) ||
2522 (bol->in(1)->Opcode() == Op_CompareAndExchangeN ) ||
2523 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapB ) ||
2524 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapS ) ||
2525 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapI ) ||
2526 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapL ) ||
2527 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapP ) ||
2528 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapN ) ||
2529 (bol->in(1)->Opcode() == Op_CompareAndSwapB ) ||
2530 (bol->in(1)->Opcode() == Op_CompareAndSwapS ) ||
2531 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2532 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2533 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2534 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2535 return; // Allocation loops RARELY take backedge
2536 // Find the OTHER exit path from the IF
2537 Node* ex = iff->proj_out(1-test_con);
2538 float p = iff->_prob;
2539 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2540 if( top == Op_IfTrue ) {
2541 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2542 iff->_prob = PROB_STATIC_FREQUENT;
2543 }
2544 } else {
2545 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2546 iff->_prob = PROB_STATIC_INFREQUENT;
2547 }
2548 }
2549 }
2550 }
2551 }
2552 test = phase->idom(test);
2553 }
2554 }
2555
2556 #ifdef ASSERT
2557 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) {
2558 Node *ctrl = cl->skip_strip_mined()->in(LoopNode::EntryControl);
2559 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
2560 Node *iffm = ctrl->in(0);
2561 assert(iffm->Opcode() == Op_If, "");
2562 Node *p_f = iffm->in(0);
2563 assert(p_f->Opcode() == Op_IfFalse, "");
2564 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2565 assert(pre_end->loopnode()->is_pre_loop(), "");
2566 return pre_end->loopnode();
2567 }
2568 #endif
2569
2570 // Remove the main and post loops and make the pre loop execute all
2571 // iterations. Useful when the pre loop is found empty.
2572 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) {
2573 CountedLoopEndNode* pre_end = cl->loopexit();
2574 Node* pre_cmp = pre_end->cmp_node();
2575 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) {
2576 // Only safe to remove the main loop if the compiler optimized it
2577 // out based on an unknown number of iterations
2578 return;
2579 }
2580
2581 // Can we find the main loop?
2582 if (_next == NULL) {
2583 return;
2584 }
2585
2586 Node* next_head = _next->_head;
2587 if (!next_head->is_CountedLoop()) {
2588 return;
2589 }
2590
2591 CountedLoopNode* main_head = next_head->as_CountedLoop();
2592 if (!main_head->is_main_loop()) {
2593 return;
2594 }
2595
2596 assert(locate_pre_from_main(main_head) == cl, "bad main loop");
2597 Node* main_iff = main_head->skip_strip_mined()->in(LoopNode::EntryControl)->in(0);
2598
2599 // Remove the Opaque1Node of the pre loop and make it execute all iterations
2600 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2));
2601 // Remove the Opaque1Node of the main loop so it can be optimized out
2602 Node* main_cmp = main_iff->in(1)->in(1);
2603 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?");
2604 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1));
2605 }
2606
2607 //------------------------------policy_do_remove_empty_loop--------------------
2608 // Micro-benchmark spamming. Policy is to always remove empty loops.
2609 // The 'DO' part is to replace the trip counter with the value it will
2610 // have on the last iteration. This will break the loop.
2611 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2612 // Minimum size must be empty loop
2613 if (_body.size() > EMPTY_LOOP_SIZE)
2614 return false;
2615
2616 if (!_head->is_CountedLoop())
2617 return false; // Dead loop
2618 CountedLoopNode *cl = _head->as_CountedLoop();
2619 if (!cl->is_valid_counted_loop())
2620 return false; // Malformed loop
2621 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2622 return false; // Infinite loop
2623
2624 if (cl->is_pre_loop()) {
2625 // If the loop we are removing is a pre-loop then the main and
2626 // post loop can be removed as well
2627 remove_main_post_loops(cl, phase);
2628 }
2629
2630 #ifdef ASSERT
2631 // Ensure only one phi which is the iv.
2632 Node* iv = NULL;
2633 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2634 Node* n = cl->fast_out(i);
2635 if (n->Opcode() == Op_Phi) {
2636 assert(iv == NULL, "Too many phis" );
2637 iv = n;
2638 }
2639 }
2640 assert(iv == cl->phi(), "Wrong phi" );
2641 #endif
2642
2643 // main and post loops have explicitly created zero trip guard
2644 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2645 if (needs_guard) {
2646 // Skip guard if values not overlap.
2647 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2648 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2649 int stride_con = cl->stride_con();
2650 if (stride_con > 0) {
2651 needs_guard = (init_t->_hi >= limit_t->_lo);
2652 } else {
2653 needs_guard = (init_t->_lo <= limit_t->_hi);
2654 }
2655 }
2656 if (needs_guard) {
2657 // Check for an obvious zero trip guard.
2658 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->skip_strip_mined()->in(LoopNode::EntryControl));
2659 if (inctrl->Opcode() == Op_IfTrue || inctrl->Opcode() == Op_IfFalse) {
2660 bool maybe_swapped = (inctrl->Opcode() == Op_IfFalse);
2661 // The test should look like just the backedge of a CountedLoop
2662 Node* iff = inctrl->in(0);
2663 if (iff->is_If()) {
2664 Node* bol = iff->in(1);
2665 if (bol->is_Bool()) {
2666 BoolTest test = bol->as_Bool()->_test;
2667 if (maybe_swapped) {
2668 test._test = test.commute();
2669 test._test = test.negate();
2670 }
2671 if (test._test == cl->loopexit()->test_trip()) {
2672 Node* cmp = bol->in(1);
2673 int init_idx = maybe_swapped ? 2 : 1;
2674 int limit_idx = maybe_swapped ? 1 : 2;
2675 if (cmp->is_Cmp() && cmp->in(init_idx) == cl->init_trip() && cmp->in(limit_idx) == cl->limit()) {
2676 needs_guard = false;
2677 }
2678 }
2679 }
2680 }
2681 }
2682 }
2683
2684 #ifndef PRODUCT
2685 if (PrintOpto) {
2686 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2687 this->dump_head();
2688 } else if (TraceLoopOpts) {
2689 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2690 this->dump_head();
2691 }
2692 #endif
2693
2694 if (needs_guard) {
2695 // Peel the loop to ensure there's a zero trip guard
2696 Node_List old_new;
2697 phase->do_peeling(this, old_new);
2698 }
2699
2700 // Replace the phi at loop head with the final value of the last
2701 // iteration. Then the CountedLoopEnd will collapse (backedge never
2702 // taken) and all loop-invariant uses of the exit values will be correct.
2703 Node *phi = cl->phi();
2704 Node *exact_limit = phase->exact_limit(this);
2705 if (exact_limit != cl->limit()) {
2706 // We also need to replace the original limit to collapse loop exit.
2707 Node* cmp = cl->loopexit()->cmp_node();
2708 assert(cl->limit() == cmp->in(2), "sanity");
2709 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2710 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2711 }
2712 // Note: the final value after increment should not overflow since
2713 // counted loop has limit check predicate.
2714 Node *final = new SubINode( exact_limit, cl->stride() );
2715 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2716 phase->_igvn.replace_node(phi,final);
2717 phase->C->set_major_progress();
2718 return true;
2719 }
2720
2721 //------------------------------policy_do_one_iteration_loop-------------------
2722 // Convert one iteration loop into normal code.
2723 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2724 if (!_head->as_Loop()->is_valid_counted_loop())
2725 return false; // Only for counted loop
2726
2727 CountedLoopNode *cl = _head->as_CountedLoop();
2728 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2729 return false;
2730 }
2731
2732 #ifndef PRODUCT
2733 if(TraceLoopOpts) {
2734 tty->print("OneIteration ");
2735 this->dump_head();
2736 }
2737 #endif
2738
2739 Node *init_n = cl->init_trip();
2740 #ifdef ASSERT
2741 // Loop boundaries should be constant since trip count is exact.
2742 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2743 #endif
2744 // Replace the phi at loop head with the value of the init_trip.
2745 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2746 // and all loop-invariant uses of the exit values will be correct.
2747 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2748 phase->C->set_major_progress();
2749 return true;
2750 }
2751
2752 //=============================================================================
2753 //------------------------------iteration_split_impl---------------------------
2754 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2755 // Compute loop trip count if possible.
2756 compute_trip_count(phase);
2757
2758 // Convert one iteration loop into normal code.
2759 if (policy_do_one_iteration_loop(phase))
2760 return true;
2761
2762 // Check and remove empty loops (spam micro-benchmarks)
2763 if (policy_do_remove_empty_loop(phase))
2764 return true; // Here we removed an empty loop
2765
2766 bool should_peel = policy_peeling(phase); // Should we peel?
2767
2768 bool should_unswitch = policy_unswitching(phase);
2769
2770 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2771 // This removes loop-invariant tests (usually null checks).
2772 if (!_head->is_CountedLoop()) { // Non-counted loop
2773 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2774 // Partial peel succeeded so terminate this round of loop opts
2775 return false;
2776 }
2777 if (should_peel) { // Should we peel?
2778 if (PrintOpto) { tty->print_cr("should_peel"); }
2779 phase->do_peeling(this,old_new);
2780 } else if (should_unswitch) {
2781 phase->do_unswitching(this, old_new);
2782 }
2783 return true;
2784 }
2785 CountedLoopNode *cl = _head->as_CountedLoop();
2786
2787 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2788
2789 // Do nothing special to pre- and post- loops
2790 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2791
2792 // Compute loop trip count from profile data
2793 compute_profile_trip_cnt(phase);
2794
2795 // Before attempting fancy unrolling, RCE or alignment, see if we want
2796 // to completely unroll this loop or do loop unswitching.
2797 if (cl->is_normal_loop()) {
2798 if (should_unswitch) {
2799 phase->do_unswitching(this, old_new);
2800 return true;
2801 }
2802 bool should_maximally_unroll = policy_maximally_unroll(phase);
2803 if (should_maximally_unroll) {
2804 // Here we did some unrolling and peeling. Eventually we will
2805 // completely unroll this loop and it will no longer be a loop.
2806 phase->do_maximally_unroll(this,old_new);
2807 return true;
2808 }
2809 }
2810
2811 // Skip next optimizations if running low on nodes. Note that
2812 // policy_unswitching and policy_maximally_unroll have this check.
2813 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes();
2814 if ((int)(2 * _body.size()) > nodes_left) {
2815 return true;
2816 }
2817
2818 // Counted loops may be peeled, may need some iterations run up
2819 // front for RCE, and may want to align loop refs to a cache
2820 // line. Thus we clone a full loop up front whose trip count is
2821 // at least 1 (if peeling), but may be several more.
2822
2823 // The main loop will start cache-line aligned with at least 1
2824 // iteration of the unrolled body (zero-trip test required) and
2825 // will have some range checks removed.
2826
2827 // A post-loop will finish any odd iterations (leftover after
2828 // unrolling), plus any needed for RCE purposes.
2829
2830 bool should_unroll = policy_unroll(phase);
2831
2832 bool should_rce = policy_range_check(phase);
2833
2834 bool should_align = policy_align(phase);
2835
2836 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2837 // need a pre-loop. We may still need to peel an initial iteration but
2838 // we will not be needing an unknown number of pre-iterations.
2839 //
2840 // Basically, if may_rce_align reports FALSE first time through,
2841 // we will not be able to later do RCE or Aligning on this loop.
2842 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2843
2844 // If we have any of these conditions (RCE, alignment, unrolling) met, then
2845 // we switch to the pre-/main-/post-loop model. This model also covers
2846 // peeling.
2847 if (should_rce || should_align || should_unroll) {
2848 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
2849 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2850
2851 // Adjust the pre- and main-loop limits to let the pre and post loops run
2852 // with full checks, but the main-loop with no checks. Remove said
2853 // checks from the main body.
2854 if (should_rce) {
2855 if (phase->do_range_check(this, old_new) != 0) {
2856 cl->mark_has_range_checks();
2857 }
2858 } else if (PostLoopMultiversioning) {
2859 phase->has_range_checks(this);
2860 }
2861
2862 if (should_unroll && !should_peel && PostLoopMultiversioning) {
2863 // Try to setup multiversioning on main loops before they are unrolled
2864 if (cl->is_main_loop() && (cl->unrolled_count() == 1)) {
2865 phase->insert_scalar_rced_post_loop(this, old_new);
2866 }
2867 }
2868
2869 // Double loop body for unrolling. Adjust the minimum-trip test (will do
2870 // twice as many iterations as before) and the main body limit (only do
2871 // an even number of trips). If we are peeling, we might enable some RCE
2872 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2873 // peeling.
2874 if (should_unroll && !should_peel) {
2875 if (SuperWordLoopUnrollAnalysis) {
2876 phase->insert_vector_post_loop(this, old_new);
2877 }
2878 phase->do_unroll(this, old_new, true);
2879 }
2880
2881 // Adjust the pre-loop limits to align the main body
2882 // iterations.
2883 if (should_align)
2884 Unimplemented();
2885
2886 } else { // Else we have an unchanged counted loop
2887 if (should_peel) // Might want to peel but do nothing else
2888 phase->do_peeling(this,old_new);
2889 }
2890 return true;
2891 }
2892
2893
2894 //=============================================================================
2895 //------------------------------iteration_split--------------------------------
2896 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2897 // Recursively iteration split nested loops
2898 if (_child && !_child->iteration_split(phase, old_new))
2899 return false;
2900
2901 // Clean out prior deadwood
2902 DCE_loop_body();
2903
2904
2905 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2906 // Replace with a 1-in-10 exit guess.
2907 if (_parent /*not the root loop*/ &&
2908 !_irreducible &&
2909 // Also ignore the occasional dead backedge
2910 !tail()->is_top()) {
2911 adjust_loop_exit_prob(phase);
2912 }
2913
2914 // Gate unrolling, RCE and peeling efforts.
2915 if (!_child && // If not an inner loop, do not split
2916 !_irreducible &&
2917 _allow_optimizations &&
2918 !tail()->is_top()) { // Also ignore the occasional dead backedge
2919 if (!_has_call) {
2920 if (!iteration_split_impl(phase, old_new)) {
2921 return false;
2922 }
2923 } else if (policy_unswitching(phase)) {
2924 phase->do_unswitching(this, old_new);
2925 }
2926 }
2927
2928 // Minor offset re-organization to remove loop-fallout uses of
2929 // trip counter when there was no major reshaping.
2930 phase->reorg_offsets(this);
2931
2932 if (_next && !_next->iteration_split(phase, old_new))
2933 return false;
2934 return true;
2935 }
2936
2937
2938 //=============================================================================
2939 // Process all the loops in the loop tree and replace any fill
2940 // patterns with an intrinsic version.
2941 bool PhaseIdealLoop::do_intrinsify_fill() {
2942 bool changed = false;
2943 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2944 IdealLoopTree* lpt = iter.current();
2945 changed |= intrinsify_fill(lpt);
2946 }
2947 return changed;
2948 }
2949
2950
2951 // Examine an inner loop looking for a a single store of an invariant
2952 // value in a unit stride loop,
2953 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2954 Node*& shift, Node*& con) {
2955 const char* msg = NULL;
2956 Node* msg_node = NULL;
2957
2958 store_value = NULL;
2959 con = NULL;
2960 shift = NULL;
2961
2962 // Process the loop looking for stores. If there are multiple
2963 // stores or extra control flow give at this point.
2964 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2965 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2966 Node* n = lpt->_body.at(i);
2967 if (n->outcnt() == 0) continue; // Ignore dead
2968 if (n->is_Store()) {
2969 if (store != NULL) {
2970 msg = "multiple stores";
2971 break;
2972 }
2973 int opc = n->Opcode();
2974 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
2975 msg = "oop fills not handled";
2976 break;
2977 }
2978 Node* value = n->in(MemNode::ValueIn);
2979 if (!lpt->is_invariant(value)) {
2980 msg = "variant store value";
2981 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2982 msg = "not array address";
2983 }
2984 store = n;
2985 store_value = value;
2986 } else if (n->is_If() && n != head->loopexit()) {
2987 msg = "extra control flow";
2988 msg_node = n;
2989 }
2990 }
2991
2992 if (store == NULL) {
2993 // No store in loop
2994 return false;
2995 }
2996
2997 if (msg == NULL && head->stride_con() != 1) {
2998 // could handle negative strides too
2999 if (head->stride_con() < 0) {
3000 msg = "negative stride";
3001 } else {
3002 msg = "non-unit stride";
3003 }
3004 }
3005
3006 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
3007 msg = "can't handle store address";
3008 msg_node = store->in(MemNode::Address);
3009 }
3010
3011 if (msg == NULL &&
3012 (!store->in(MemNode::Memory)->is_Phi() ||
3013 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
3014 msg = "store memory isn't proper phi";
3015 msg_node = store->in(MemNode::Memory);
3016 }
3017
3018 // Make sure there is an appropriate fill routine
3019 BasicType t = store->as_Mem()->memory_type();
3020 const char* fill_name;
3021 if (msg == NULL &&
3022 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
3023 msg = "unsupported store";
3024 msg_node = store;
3025 }
3026
3027 if (msg != NULL) {
3028 #ifndef PRODUCT
3029 if (TraceOptimizeFill) {
3030 tty->print_cr("not fill intrinsic candidate: %s", msg);
3031 if (msg_node != NULL) msg_node->dump();
3032 }
3033 #endif
3034 return false;
3035 }
3036
3037 // Make sure the address expression can be handled. It should be
3038 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()).
3039 Node* elements[4];
3040 Node* cast = NULL;
3041 Node* conv = NULL;
3042 bool found_index = false;
3043 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
3044 for (int e = 0; e < count; e++) {
3045 Node* n = elements[e];
3046 if (n->is_Con() && con == NULL) {
3047 con = n;
3048 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
3049 Node* value = n->in(1);
3050 #ifdef _LP64
3051 if (value->Opcode() == Op_ConvI2L) {
3052 conv = value;
3053 value = value->in(1);
3054 }
3055 if (value->Opcode() == Op_CastII &&
3056 value->as_CastII()->has_range_check()) {
3057 // Skip range check dependent CastII nodes
3058 cast = value;
3059 value = value->in(1);
3060 }
3061 #endif
3062 if (value != head->phi()) {
3063 msg = "unhandled shift in address";
3064 } else {
3065 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
3066 msg = "scale doesn't match";
3067 } else {
3068 found_index = true;
3069 shift = n;
3070 }
3071 }
3072 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
3073 conv = n;
3074 n = n->in(1);
3075 if (n->Opcode() == Op_CastII &&
3076 n->as_CastII()->has_range_check()) {
3077 // Skip range check dependent CastII nodes
3078 cast = n;
3079 n = n->in(1);
3080 }
3081 if (n == head->phi()) {
3082 found_index = true;
3083 } else {
3084 msg = "unhandled input to ConvI2L";
3085 }
3086 } else if (n == head->phi()) {
3087 // no shift, check below for allowed cases
3088 found_index = true;
3089 } else {
3090 msg = "unhandled node in address";
3091 msg_node = n;
3092 }
3093 }
3094
3095 if (count == -1) {
3096 msg = "malformed address expression";
3097 msg_node = store;
3098 }
3099
3100 if (!found_index) {
3101 msg = "missing use of index";
3102 }
3103
3104 // byte sized items won't have a shift
3105 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
3106 msg = "can't find shift";
3107 msg_node = store;
3108 }
3109
3110 if (msg != NULL) {
3111 #ifndef PRODUCT
3112 if (TraceOptimizeFill) {
3113 tty->print_cr("not fill intrinsic: %s", msg);
3114 if (msg_node != NULL) msg_node->dump();
3115 }
3116 #endif
3117 return false;
3118 }
3119
3120 // No make sure all the other nodes in the loop can be handled
3121 VectorSet ok(Thread::current()->resource_area());
3122
3123 // store related values are ok
3124 ok.set(store->_idx);
3125 ok.set(store->in(MemNode::Memory)->_idx);
3126
3127 CountedLoopEndNode* loop_exit = head->loopexit();
3128 guarantee(loop_exit != NULL, "no loop exit node");
3129
3130 // Loop structure is ok
3131 ok.set(head->_idx);
3132 ok.set(loop_exit->_idx);
3133 ok.set(head->phi()->_idx);
3134 ok.set(head->incr()->_idx);
3135 ok.set(loop_exit->cmp_node()->_idx);
3136 ok.set(loop_exit->in(1)->_idx);
3137
3138 // Address elements are ok
3139 if (con) ok.set(con->_idx);
3140 if (shift) ok.set(shift->_idx);
3141 if (cast) ok.set(cast->_idx);
3142 if (conv) ok.set(conv->_idx);
3143
3144 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3145 Node* n = lpt->_body.at(i);
3146 if (n->outcnt() == 0) continue; // Ignore dead
3147 if (ok.test(n->_idx)) continue;
3148 // Backedge projection is ok
3149 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
3150 if (!n->is_AddP()) {
3151 msg = "unhandled node";
3152 msg_node = n;
3153 break;
3154 }
3155 }
3156
3157 // Make sure no unexpected values are used outside the loop
3158 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3159 Node* n = lpt->_body.at(i);
3160 // These values can be replaced with other nodes if they are used
3161 // outside the loop.
3162 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
3163 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
3164 Node* use = iter.get();
3165 if (!lpt->_body.contains(use)) {
3166 msg = "node is used outside loop";
3167 // lpt->_body.dump();
3168 msg_node = n;
3169 break;
3170 }
3171 }
3172 }
3173
3174 #ifdef ASSERT
3175 if (TraceOptimizeFill) {
3176 if (msg != NULL) {
3177 tty->print_cr("no fill intrinsic: %s", msg);
3178 if (msg_node != NULL) msg_node->dump();
3179 } else {
3180 tty->print_cr("fill intrinsic for:");
3181 }
3182 store->dump();
3183 if (Verbose) {
3184 lpt->_body.dump();
3185 }
3186 }
3187 #endif
3188
3189 return msg == NULL;
3190 }
3191
3192
3193
3194 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
3195 // Only for counted inner loops
3196 if (!lpt->is_counted() || !lpt->is_inner()) {
3197 return false;
3198 }
3199
3200 // Must have constant stride
3201 CountedLoopNode* head = lpt->_head->as_CountedLoop();
3202 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
3203 return false;
3204 }
3205
3206 head->verify_strip_mined(1);
3207
3208 // Check that the body only contains a store of a loop invariant
3209 // value that is indexed by the loop phi.
3210 Node* store = NULL;
3211 Node* store_value = NULL;
3212 Node* shift = NULL;
3213 Node* offset = NULL;
3214 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
3215 return false;
3216 }
3217
3218 Node* exit = head->loopexit()->proj_out(0);
3219 if (exit == NULL) {
3220 return false;
3221 }
3222
3223 #ifndef PRODUCT
3224 if (TraceLoopOpts) {
3225 tty->print("ArrayFill ");
3226 lpt->dump_head();
3227 }
3228 #endif
3229
3230 // Now replace the whole loop body by a call to a fill routine that
3231 // covers the same region as the loop.
3232 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
3233
3234 // Build an expression for the beginning of the copy region
3235 Node* index = head->init_trip();
3236 #ifdef _LP64
3237 index = new ConvI2LNode(index);
3238 _igvn.register_new_node_with_optimizer(index);
3239 #endif
3240 if (shift != NULL) {
3241 // byte arrays don't require a shift but others do.
3242 index = new LShiftXNode(index, shift->in(2));
3243 _igvn.register_new_node_with_optimizer(index);
3244 }
3245 index = new AddPNode(base, base, index);
3246 _igvn.register_new_node_with_optimizer(index);
3247 Node* from = new AddPNode(base, index, offset);
3248 _igvn.register_new_node_with_optimizer(from);
3249 // Compute the number of elements to copy
3250 Node* len = new SubINode(head->limit(), head->init_trip());
3251 _igvn.register_new_node_with_optimizer(len);
3252
3253 BasicType t = store->as_Mem()->memory_type();
3254 bool aligned = false;
3255 if (offset != NULL && head->init_trip()->is_Con()) {
3256 int element_size = type2aelembytes(t);
3257 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
3258 }
3259
3260 // Build a call to the fill routine
3261 const char* fill_name;
3262 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
3263 assert(fill != NULL, "what?");
3264
3265 // Convert float/double to int/long for fill routines
3266 if (t == T_FLOAT) {
3267 store_value = new MoveF2INode(store_value);
3268 _igvn.register_new_node_with_optimizer(store_value);
3269 } else if (t == T_DOUBLE) {
3270 store_value = new MoveD2LNode(store_value);
3271 _igvn.register_new_node_with_optimizer(store_value);
3272 }
3273
3274 Node* mem_phi = store->in(MemNode::Memory);
3275 Node* result_ctrl;
3276 Node* result_mem;
3277 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
3278 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill,
3279 fill_name, TypeAryPtr::get_array_body_type(t));
3280 uint cnt = 0;
3281 call->init_req(TypeFunc::Parms + cnt++, from);
3282 call->init_req(TypeFunc::Parms + cnt++, store_value);
3283 #ifdef _LP64
3284 len = new ConvI2LNode(len);
3285 _igvn.register_new_node_with_optimizer(len);
3286 #endif
3287 call->init_req(TypeFunc::Parms + cnt++, len);
3288 #ifdef _LP64
3289 call->init_req(TypeFunc::Parms + cnt++, C->top());
3290 #endif
3291 call->init_req(TypeFunc::Control, head->init_control());
3292 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O.
3293 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl));
3294 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr));
3295 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
3296 _igvn.register_new_node_with_optimizer(call);
3297 result_ctrl = new ProjNode(call,TypeFunc::Control);
3298 _igvn.register_new_node_with_optimizer(result_ctrl);
3299 result_mem = new ProjNode(call,TypeFunc::Memory);
3300 _igvn.register_new_node_with_optimizer(result_mem);
3301
3302 /* Disable following optimization until proper fix (add missing checks).
3303
3304 // If this fill is tightly coupled to an allocation and overwrites
3305 // the whole body, allow it to take over the zeroing.
3306 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
3307 if (alloc != NULL && alloc->is_AllocateArray()) {
3308 Node* length = alloc->as_AllocateArray()->Ideal_length();
3309 if (head->limit() == length &&
3310 head->init_trip() == _igvn.intcon(0)) {
3311 if (TraceOptimizeFill) {
3312 tty->print_cr("Eliminated zeroing in allocation");
3313 }
3314 alloc->maybe_set_complete(&_igvn);
3315 } else {
3316 #ifdef ASSERT
3317 if (TraceOptimizeFill) {
3318 tty->print_cr("filling array but bounds don't match");
3319 alloc->dump();
3320 head->init_trip()->dump();
3321 head->limit()->dump();
3322 length->dump();
3323 }
3324 #endif
3325 }
3326 }
3327 */
3328
3329 if (head->is_strip_mined()) {
3330 // Inner strip mined loop goes away so get rid of outer strip
3331 // mined loop
3332 Node* outer_sfpt = head->outer_safepoint();
3333 Node* in = outer_sfpt->in(0);
3334 Node* outer_out = head->outer_loop_exit();
3335 lazy_replace(outer_out, in);
3336 _igvn.replace_input_of(outer_sfpt, 0, C->top());
3337 }
3338
3339 // Redirect the old control and memory edges that are outside the loop.
3340 // Sometimes the memory phi of the head is used as the outgoing
3341 // state of the loop. It's safe in this case to replace it with the
3342 // result_mem.
3343 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
3344 lazy_replace(exit, result_ctrl);
3345 _igvn.replace_node(store, result_mem);
3346 // Any uses the increment outside of the loop become the loop limit.
3347 _igvn.replace_node(head->incr(), head->limit());
3348
3349 // Disconnect the head from the loop.
3350 for (uint i = 0; i < lpt->_body.size(); i++) {
3351 Node* n = lpt->_body.at(i);
3352 _igvn.replace_node(n, C->top());
3353 }
3354
3355 return true;
3356 }