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