1 /*
2 * Copyright (c) 2000, 2015, 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 }
71
72 //------------------------------compute_exact_trip_count-----------------------
73 // Compute loop exact trip count if possible. Do not recalculate trip count for
74 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
75 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) {
76 if (!_head->as_Loop()->is_valid_counted_loop()) {
77 return;
78 }
79 CountedLoopNode* cl = _head->as_CountedLoop();
80 // Trip count may become nonexact for iteration split loops since
81 // RCE modifies limits. Note, _trip_count value is not reset since
82 // it is used to limit unrolling of main loop.
83 cl->set_nonexact_trip_count();
84
85 // Loop's test should be part of loop.
86 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
87 return; // Infinite loop
88
89 #ifdef ASSERT
90 BoolTest::mask bt = cl->loopexit()->test_trip();
91 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
92 bt == BoolTest::ne, "canonical test is expected");
93 #endif
94
95 Node* init_n = cl->init_trip();
96 Node* limit_n = cl->limit();
97 if (init_n != NULL && init_n->is_Con() &&
98 limit_n != NULL && limit_n->is_Con()) {
99 // Use longs to avoid integer overflow.
100 int stride_con = cl->stride_con();
101 jlong init_con = cl->init_trip()->get_int();
102 jlong limit_con = cl->limit()->get_int();
103 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
104 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
105 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
106 // Set exact trip count.
107 cl->set_exact_trip_count((uint)trip_count);
108 }
109 }
110 }
111
112 //------------------------------compute_profile_trip_cnt----------------------------
113 // Compute loop trip count from profile data as
114 // (backedge_count + loop_exit_count) / loop_exit_count
115 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
116 if (!_head->is_CountedLoop()) {
117 return;
118 }
119 CountedLoopNode* head = _head->as_CountedLoop();
120 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
121 return; // Already computed
122 }
123 float trip_cnt = (float)max_jint; // default is big
124
125 Node* back = head->in(LoopNode::LoopBackControl);
126 while (back != head) {
127 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
128 back->in(0) &&
129 back->in(0)->is_If() &&
130 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
131 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
132 break;
133 }
134 back = phase->idom(back);
135 }
136 if (back != head) {
137 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
138 back->in(0), "if-projection exists");
139 IfNode* back_if = back->in(0)->as_If();
140 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
141
142 // Now compute a loop exit count
143 float loop_exit_cnt = 0.0f;
144 for( uint i = 0; i < _body.size(); i++ ) {
145 Node *n = _body[i];
146 if( n->is_If() ) {
147 IfNode *iff = n->as_If();
148 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
149 Node *exit = is_loop_exit(iff);
150 if( exit ) {
151 float exit_prob = iff->_prob;
152 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
153 if (exit_prob > PROB_MIN) {
154 float exit_cnt = iff->_fcnt * exit_prob;
155 loop_exit_cnt += exit_cnt;
156 }
157 }
158 }
159 }
160 }
161 if (loop_exit_cnt > 0.0f) {
162 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
163 } else {
164 // No exit count so use
165 trip_cnt = loop_back_cnt;
166 }
167 }
168 #ifndef PRODUCT
169 if (TraceProfileTripCount) {
170 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
171 }
172 #endif
173 head->set_profile_trip_cnt(trip_cnt);
174 }
175
176 //---------------------is_invariant_addition-----------------------------
177 // Return nonzero index of invariant operand for an Add or Sub
178 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
179 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
180 int op = n->Opcode();
181 if (op == Op_AddI || op == Op_SubI) {
182 bool in1_invar = this->is_invariant(n->in(1));
183 bool in2_invar = this->is_invariant(n->in(2));
184 if (in1_invar && !in2_invar) return 1;
185 if (!in1_invar && in2_invar) return 2;
186 }
187 return 0;
188 }
189
190 //---------------------reassociate_add_sub-----------------------------
191 // Reassociate invariant add and subtract expressions:
192 //
193 // inv1 + (x + inv2) => ( inv1 + inv2) + x
194 // (x + inv2) + inv1 => ( inv1 + inv2) + x
195 // inv1 + (x - inv2) => ( inv1 - inv2) + x
196 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
197 // (x + inv2) - inv1 => (-inv1 + inv2) + x
198 // (x - inv2) + inv1 => ( inv1 - inv2) + x
199 // (x - inv2) - inv1 => (-inv1 - inv2) + x
200 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
201 // inv1 - (x - inv2) => ( inv1 + inv2) - x
202 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
203 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
204 // inv1 - (x + inv2) => ( inv1 - inv2) - x
205 //
206 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
207 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
208 if (is_invariant(n1)) return NULL;
209 int inv1_idx = is_invariant_addition(n1, phase);
210 if (!inv1_idx) return NULL;
211 // Don't mess with add of constant (igvn moves them to expression tree root.)
212 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
213 Node* inv1 = n1->in(inv1_idx);
214 Node* n2 = n1->in(3 - inv1_idx);
215 int inv2_idx = is_invariant_addition(n2, phase);
216 if (!inv2_idx) return NULL;
217 Node* x = n2->in(3 - inv2_idx);
218 Node* inv2 = n2->in(inv2_idx);
219
220 bool neg_x = n2->is_Sub() && inv2_idx == 1;
221 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
222 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
223 if (n1->is_Sub() && inv1_idx == 1) {
224 neg_x = !neg_x;
225 neg_inv2 = !neg_inv2;
226 }
227 Node* inv1_c = phase->get_ctrl(inv1);
228 Node* inv2_c = phase->get_ctrl(inv2);
229 Node* n_inv1;
230 if (neg_inv1) {
231 Node *zero = phase->_igvn.intcon(0);
232 phase->set_ctrl(zero, phase->C->root());
233 n_inv1 = new SubINode(zero, inv1);
234 phase->register_new_node(n_inv1, inv1_c);
235 } else {
236 n_inv1 = inv1;
237 }
238 Node* inv;
239 if (neg_inv2) {
240 inv = new SubINode(n_inv1, inv2);
241 } else {
242 inv = new AddINode(n_inv1, inv2);
243 }
244 phase->register_new_node(inv, phase->get_early_ctrl(inv));
245
246 Node* addx;
247 if (neg_x) {
248 addx = new SubINode(inv, x);
249 } else {
250 addx = new AddINode(x, inv);
251 }
252 phase->register_new_node(addx, phase->get_ctrl(x));
253 phase->_igvn.replace_node(n1, addx);
254 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
255 _body.yank(n1);
256 return addx;
257 }
258
259 //---------------------reassociate_invariants-----------------------------
260 // Reassociate invariant expressions:
261 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
262 for (int i = _body.size() - 1; i >= 0; i--) {
263 Node *n = _body.at(i);
264 for (int j = 0; j < 5; j++) {
265 Node* nn = reassociate_add_sub(n, phase);
266 if (nn == NULL) break;
267 n = nn; // again
268 };
269 }
270 }
271
272 //------------------------------policy_peeling---------------------------------
273 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
274 // make some loop-invariant test (usually a null-check) happen before the loop.
275 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
276 Node *test = ((IdealLoopTree*)this)->tail();
277 int body_size = ((IdealLoopTree*)this)->_body.size();
278 // Peeling does loop cloning which can result in O(N^2) node construction
279 if( body_size > 255 /* Prevent overflow for large body_size */
280 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) {
281 return false; // too large to safely clone
282 }
283
284 // check for vectorized loops, any peeling done was already applied
285 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false;
286
287 while( test != _head ) { // Scan till run off top of loop
288 if( test->is_If() ) { // Test?
289 Node *ctrl = phase->get_ctrl(test->in(1));
290 if (ctrl->is_top())
291 return false; // Found dead test on live IF? No peeling!
292 // Standard IF only has one input value to check for loop invariance
293 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
294 // Condition is not a member of this loop?
295 if( !is_member(phase->get_loop(ctrl)) &&
296 is_loop_exit(test) )
297 return true; // Found reason to peel!
298 }
299 // Walk up dominators to loop _head looking for test which is
300 // executed on every path thru loop.
301 test = phase->idom(test);
302 }
303 return false;
304 }
305
306 //------------------------------peeled_dom_test_elim---------------------------
307 // If we got the effect of peeling, either by actually peeling or by making
308 // a pre-loop which must execute at least once, we can remove all
309 // loop-invariant dominated tests in the main body.
310 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
311 bool progress = true;
312 while( progress ) {
313 progress = false; // Reset for next iteration
314 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
315 Node *test = prev->in(0);
316 while( test != loop->_head ) { // Scan till run off top of loop
317
318 int p_op = prev->Opcode();
319 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
320 test->is_If() && // Test?
321 !test->in(1)->is_Con() && // And not already obvious?
322 // Condition is not a member of this loop?
323 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
324 // Walk loop body looking for instances of this test
325 for( uint i = 0; i < loop->_body.size(); i++ ) {
326 Node *n = loop->_body.at(i);
327 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
328 // IfNode was dominated by version in peeled loop body
329 progress = true;
330 dominated_by( old_new[prev->_idx], n );
331 }
332 }
333 }
334 prev = test;
335 test = idom(test);
336 } // End of scan tests in loop
337
338 } // End of while( progress )
339 }
340
341 //------------------------------do_peeling-------------------------------------
342 // Peel the first iteration of the given loop.
343 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
344 // The pre-loop illegally has 2 control users (old & new loops).
345 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
346 // Do this by making the old-loop fall-in edges act as if they came
347 // around the loopback from the prior iteration (follow the old-loop
348 // backedges) and then map to the new peeled iteration. This leaves
349 // the pre-loop with only 1 user (the new peeled iteration), but the
350 // peeled-loop backedge has 2 users.
351 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
352 // extra backedge user.
353 //
354 // orig
355 //
356 // stmt1
357 // |
358 // v
359 // loop predicate
360 // |
361 // v
362 // loop<----+
363 // | |
364 // stmt2 |
365 // | |
366 // v |
367 // if ^
368 // / \ |
369 // / \ |
370 // v v |
371 // false true |
372 // / \ |
373 // / ----+
374 // |
375 // v
376 // exit
377 //
378 //
379 // after clone loop
380 //
381 // stmt1
382 // |
383 // v
384 // loop predicate
385 // / \
386 // clone / \ orig
387 // / \
388 // / \
389 // v v
390 // +---->loop clone loop<----+
391 // | | | |
392 // | stmt2 clone stmt2 |
393 // | | | |
394 // | v v |
395 // ^ if clone If ^
396 // | / \ / \ |
397 // | / \ / \ |
398 // | v v v v |
399 // | true false false true |
400 // | / \ / \ |
401 // +---- \ / ----+
402 // \ /
403 // 1v v2
404 // region
405 // |
406 // v
407 // exit
408 //
409 //
410 // after peel and predicate move
411 //
412 // stmt1
413 // /
414 // /
415 // clone / orig
416 // /
417 // / +----------+
418 // / | |
419 // / loop predicate |
420 // / | |
421 // v v |
422 // TOP-->loop clone loop<----+ |
423 // | | | |
424 // stmt2 clone stmt2 | |
425 // | | | ^
426 // v v | |
427 // if clone If ^ |
428 // / \ / \ | |
429 // / \ / \ | |
430 // v v v v | |
431 // true false false true | |
432 // | \ / \ | |
433 // | \ / ----+ ^
434 // | \ / |
435 // | 1v v2 |
436 // v region |
437 // | | |
438 // | v |
439 // | exit |
440 // | |
441 // +--------------->-----------------+
442 //
443 //
444 // final graph
445 //
446 // stmt1
447 // |
448 // v
449 // stmt2 clone
450 // |
451 // v
452 // if clone
453 // / |
454 // / |
455 // v v
456 // false true
457 // | |
458 // | v
459 // | loop predicate
460 // | |
461 // | v
462 // | loop<----+
463 // | | |
464 // | stmt2 |
465 // | | |
466 // | v |
467 // v if ^
468 // | / \ |
469 // | / \ |
470 // | v v |
471 // | false true |
472 // | | \ |
473 // v v --+
474 // region
475 // |
476 // v
477 // exit
478 //
479 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
480
481 C->set_major_progress();
482 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
483 // 'pre' loop from the main and the 'pre' can no longer have its
484 // iterations adjusted. Therefore, we need to declare this loop as
485 // no longer a 'main' loop; it will need new pre and post loops before
486 // we can do further RCE.
487 #ifndef PRODUCT
488 if (TraceLoopOpts) {
489 tty->print("Peel ");
490 loop->dump_head();
491 }
492 #endif
493 Node* head = loop->_head;
494 bool counted_loop = head->is_CountedLoop();
495 if (counted_loop) {
496 CountedLoopNode *cl = head->as_CountedLoop();
497 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
498 cl->set_trip_count(cl->trip_count() - 1);
499 if (cl->is_main_loop()) {
500 cl->set_normal_loop();
501 #ifndef PRODUCT
502 if (PrintOpto && VerifyLoopOptimizations) {
503 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
504 loop->dump_head();
505 }
506 #endif
507 }
508 }
509 Node* entry = head->in(LoopNode::EntryControl);
510
511 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
512 // The pre-loop illegally has 2 control users (old & new loops).
513 clone_loop( loop, old_new, dom_depth(head) );
514
515 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
516 // Do this by making the old-loop fall-in edges act as if they came
517 // around the loopback from the prior iteration (follow the old-loop
518 // backedges) and then map to the new peeled iteration. This leaves
519 // the pre-loop with only 1 user (the new peeled iteration), but the
520 // peeled-loop backedge has 2 users.
521 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
522 _igvn.hash_delete(head);
523 head->set_req(LoopNode::EntryControl, new_entry);
524 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
525 Node* old = head->fast_out(j);
526 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
527 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
528 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
529 // Then loop body backedge value remains the same.
530 new_exit_value = old->in(LoopNode::LoopBackControl);
531 _igvn.hash_delete(old);
532 old->set_req(LoopNode::EntryControl, new_exit_value);
533 }
534 }
535
536
537 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
538 // extra backedge user.
539 Node* new_head = old_new[head->_idx];
540 _igvn.hash_delete(new_head);
541 new_head->set_req(LoopNode::LoopBackControl, C->top());
542 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
543 Node* use = new_head->fast_out(j2);
544 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
545 _igvn.hash_delete(use);
546 use->set_req(LoopNode::LoopBackControl, C->top());
547 }
548 }
549
550
551 // Step 4: Correct dom-depth info. Set to loop-head depth.
552 int dd = dom_depth(head);
553 set_idom(head, head->in(1), dd);
554 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
555 Node *old = loop->_body.at(j3);
556 Node *nnn = old_new[old->_idx];
557 if (!has_ctrl(nnn))
558 set_idom(nnn, idom(nnn), dd-1);
559 }
560
561 // Now force out all loop-invariant dominating tests. The optimizer
562 // finds some, but we _know_ they are all useless.
563 peeled_dom_test_elim(loop,old_new);
564
565 loop->record_for_igvn();
566 }
567
568 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
569
570 //------------------------------policy_maximally_unroll------------------------
571 // Calculate exact loop trip count and return true if loop can be maximally
572 // unrolled.
573 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
574 CountedLoopNode *cl = _head->as_CountedLoop();
575 assert(cl->is_normal_loop(), "");
576 if (!cl->is_valid_counted_loop())
577 return false; // Malformed counted loop
578
579 if (!cl->has_exact_trip_count()) {
580 // Trip count is not exact.
581 return false;
582 }
583
584 uint trip_count = cl->trip_count();
585 // Note, max_juint is used to indicate unknown trip count.
586 assert(trip_count > 1, "one iteration loop should be optimized out already");
587 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
588
589 // Real policy: if we maximally unroll, does it get too big?
590 // Allow the unrolled mess to get larger than standard loop
591 // size. After all, it will no longer be a loop.
592 uint body_size = _body.size();
593 uint unroll_limit = (uint)LoopUnrollLimit * 4;
594 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
595 if (trip_count > unroll_limit || body_size > unroll_limit) {
596 return false;
597 }
598
599 // Fully unroll a loop with few iterations regardless next
600 // conditions since following loop optimizations will split
601 // such loop anyway (pre-main-post).
602 if (trip_count <= 3)
603 return true;
604
605 // Take into account that after unroll conjoined heads and tails will fold,
606 // otherwise policy_unroll() may allow more unrolling than max unrolling.
607 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
608 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
609 if (body_size != tst_body_size) // Check for int overflow
610 return false;
611 if (new_body_size > unroll_limit ||
612 // Unrolling can result in a large amount of node construction
613 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) {
614 return false;
615 }
616
617 // Do not unroll a loop with String intrinsics code.
618 // String intrinsics are large and have loops.
619 for (uint k = 0; k < _body.size(); k++) {
620 Node* n = _body.at(k);
621 switch (n->Opcode()) {
622 case Op_StrComp:
623 case Op_StrEquals:
624 case Op_StrIndexOf:
625 case Op_EncodeISOArray:
626 case Op_AryEq: {
627 return false;
628 }
629 #if INCLUDE_RTM_OPT
630 case Op_FastLock:
631 case Op_FastUnlock: {
632 // Don't unroll RTM locking code because it is large.
633 if (UseRTMLocking) {
634 return false;
635 }
636 }
637 #endif
638 } // switch
639 }
640
641 return true; // Do maximally unroll
642 }
643
644
645 //------------------------------policy_unroll----------------------------------
646 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
647 // the loop is a CountedLoop and the body is small enough.
648 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
649
650 CountedLoopNode *cl = _head->as_CountedLoop();
651 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
652
653 if (!cl->is_valid_counted_loop())
654 return false; // Malformed counted loop
655
656 // Protect against over-unrolling.
657 // After split at least one iteration will be executed in pre-loop.
658 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
659
660 _local_loop_unroll_limit = LoopUnrollLimit;
661 _local_loop_unroll_factor = 4;
662 int future_unroll_ct = cl->unrolled_count() * 2;
663 if (!cl->do_unroll_only()) {
664 if (future_unroll_ct > LoopMaxUnroll) return false;
665 } else {
666 // obey user constraints on vector mapped loops with additional unrolling applied
667 if ((future_unroll_ct / cl->slp_max_unroll()) > LoopMaxUnroll) return false;
668 }
669
670 // Check for initial stride being a small enough constant
671 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
672
673 // Don't unroll if the next round of unrolling would push us
674 // over the expected trip count of the loop. One is subtracted
675 // from the expected trip count because the pre-loop normally
676 // executes 1 iteration.
677 if (UnrollLimitForProfileCheck > 0 &&
678 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
679 future_unroll_ct > UnrollLimitForProfileCheck &&
680 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
681 return false;
682 }
683
684 // When unroll count is greater than LoopUnrollMin, don't unroll if:
685 // the residual iterations are more than 10% of the trip count
686 // and rounds of "unroll,optimize" are not making significant progress
687 // Progress defined as current size less than 20% larger than previous size.
688 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
689 future_unroll_ct > LoopUnrollMin &&
690 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
691 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
692 return false;
693 }
694
695 Node *init_n = cl->init_trip();
696 Node *limit_n = cl->limit();
697 int stride_con = cl->stride_con();
698 // Non-constant bounds.
699 // Protect against over-unrolling when init or/and limit are not constant
700 // (so that trip_count's init value is maxint) but iv range is known.
701 if (init_n == NULL || !init_n->is_Con() ||
702 limit_n == NULL || !limit_n->is_Con()) {
703 Node* phi = cl->phi();
704 if (phi != NULL) {
705 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
706 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
707 int next_stride = stride_con * 2; // stride after this unroll
708 if (next_stride > 0) {
709 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
710 iv_type->_lo + next_stride > iv_type->_hi) {
711 return false; // over-unrolling
712 }
713 } else if (next_stride < 0) {
714 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
715 iv_type->_hi + next_stride < iv_type->_lo) {
716 return false; // over-unrolling
717 }
718 }
719 }
720 }
721
722 // After unroll limit will be adjusted: new_limit = limit-stride.
723 // Bailout if adjustment overflow.
724 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
725 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
726 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
727 return false; // overflow
728
729 // Adjust body_size to determine if we unroll or not
730 uint body_size = _body.size();
731 // Key test to unroll loop in CRC32 java code
732 int xors_in_loop = 0;
733 // Also count ModL, DivL and MulL which expand mightly
734 for (uint k = 0; k < _body.size(); k++) {
735 Node* n = _body.at(k);
736 switch (n->Opcode()) {
737 case Op_XorI: xors_in_loop++; break; // CRC32 java code
738 case Op_ModL: body_size += 30; break;
739 case Op_DivL: body_size += 30; break;
740 case Op_MulL: body_size += 10; break;
741 case Op_StrComp:
742 case Op_StrEquals:
743 case Op_StrIndexOf:
744 case Op_EncodeISOArray:
745 case Op_AryEq: {
746 // Do not unroll a loop with String intrinsics code.
747 // String intrinsics are large and have loops.
748 return false;
749 }
750 #if INCLUDE_RTM_OPT
751 case Op_FastLock:
752 case Op_FastUnlock: {
753 // Don't unroll RTM locking code because it is large.
754 if (UseRTMLocking) {
755 return false;
756 }
757 }
758 #endif
759 } // switch
760 }
761
762 if (UseSuperWord) {
763 if (!cl->is_reduction_loop()) {
764 phase->mark_reductions(this);
765 }
766
767 // Only attempt slp analysis when user controls do not prohibit it
768 if (LoopMaxUnroll > _local_loop_unroll_factor) {
769 // Once policy_slp_analysis succeeds, mark the loop with the
770 // maximal unroll factor so that we minimize analysis passes
771 if (future_unroll_ct >= _local_loop_unroll_factor) {
772 policy_unroll_slp_analysis(cl, phase, future_unroll_ct);
773 }
774 }
775 }
776
777 int slp_max_unroll_factor = cl->slp_max_unroll();
778 if (cl->has_passed_slp()) {
779 if (slp_max_unroll_factor >= future_unroll_ct) return true;
780 // Normal case: loop too big
781 return false;
782 }
783
784 // Check for being too big
785 if (body_size > (uint)_local_loop_unroll_limit) {
786 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
787 // Normal case: loop too big
788 return false;
789 }
790
791 if(cl->do_unroll_only()) {
792 NOT_PRODUCT(if (TraceSuperWordLoopUnrollAnalysis) tty->print_cr("policy_unroll passed vector loop(vlen=%d,factor = %d)\n", slp_max_unroll_factor, future_unroll_ct));
793 }
794
795 // Unroll once! (Each trip will soon do double iterations)
796 return true;
797 }
798
799 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) {
800 // Enable this functionality target by target as needed
801 if (SuperWordLoopUnrollAnalysis) {
802 if (!cl->was_slp_analyzed()) {
803 SuperWord sw(phase);
804 sw.transform_loop(this, false);
805
806 // If the loop is slp canonical analyze it
807 if (sw.early_return() == false) {
808 sw.unrolling_analysis(_local_loop_unroll_factor);
809 }
810 }
811
812 if (cl->has_passed_slp()) {
813 int slp_max_unroll_factor = cl->slp_max_unroll();
814 if (slp_max_unroll_factor >= future_unroll_ct) {
815 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor;
816 if (new_limit > LoopUnrollLimit) {
817 NOT_PRODUCT(if (TraceSuperWordLoopUnrollAnalysis) tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit));
818 _local_loop_unroll_limit = new_limit;
819 }
820 }
821 }
822 }
823 }
824
825 //------------------------------policy_align-----------------------------------
826 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
827 // expression that does the alignment. Note that only one array base can be
828 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
829 // if we vectorize short memory ops into longer memory ops, we may want to
830 // increase alignment.
831 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
832 return false;
833 }
834
835 //------------------------------policy_range_check-----------------------------
836 // Return TRUE or FALSE if the loop should be range-check-eliminated.
837 // Actually we do iteration-splitting, a more powerful form of RCE.
838 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
839 if (!RangeCheckElimination) return false;
840
841 CountedLoopNode *cl = _head->as_CountedLoop();
842 // If we unrolled with no intention of doing RCE and we later
843 // changed our minds, we got no pre-loop. Either we need to
844 // make a new pre-loop, or we gotta disallow RCE.
845 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
846 Node *trip_counter = cl->phi();
847
848 // check for vectorized loops, some opts are no longer needed
849 if (cl->do_unroll_only()) return false;
850
851 // Check loop body for tests of trip-counter plus loop-invariant vs
852 // loop-invariant.
853 for (uint i = 0; i < _body.size(); i++) {
854 Node *iff = _body[i];
855 if (iff->Opcode() == Op_If) { // Test?
856
857 // Comparing trip+off vs limit
858 Node *bol = iff->in(1);
859 if (bol->req() != 2) continue; // dead constant test
860 if (!bol->is_Bool()) {
861 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
862 continue;
863 }
864 if (bol->as_Bool()->_test._test == BoolTest::ne)
865 continue; // not RC
866
867 Node *cmp = bol->in(1);
868 Node *rc_exp = cmp->in(1);
869 Node *limit = cmp->in(2);
870
871 Node *limit_c = phase->get_ctrl(limit);
872 if( limit_c == phase->C->top() )
873 return false; // Found dead test on live IF? No RCE!
874 if( is_member(phase->get_loop(limit_c) ) ) {
875 // Compare might have operands swapped; commute them
876 rc_exp = cmp->in(2);
877 limit = cmp->in(1);
878 limit_c = phase->get_ctrl(limit);
879 if( is_member(phase->get_loop(limit_c) ) )
880 continue; // Both inputs are loop varying; cannot RCE
881 }
882
883 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
884 continue;
885 }
886 // Yeah! Found a test like 'trip+off vs limit'
887 // Test is an IfNode, has 2 projections. If BOTH are in the loop
888 // we need loop unswitching instead of iteration splitting.
889 if( is_loop_exit(iff) )
890 return true; // Found reason to split iterations
891 } // End of is IF
892 }
893
894 return false;
895 }
896
897 //------------------------------policy_peel_only-------------------------------
898 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
899 // for unrolling loops with NO array accesses.
900 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
901 // check for vectorized loops, any peeling done was already applied
902 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false;
903
904 for( uint i = 0; i < _body.size(); i++ )
905 if( _body[i]->is_Mem() )
906 return false;
907
908 // No memory accesses at all!
909 return true;
910 }
911
912 //------------------------------clone_up_backedge_goo--------------------------
913 // If Node n lives in the back_ctrl block and cannot float, we clone a private
914 // version of n in preheader_ctrl block and return that, otherwise return n.
915 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
916 if( get_ctrl(n) != back_ctrl ) return n;
917
918 // Only visit once
919 if (visited.test_set(n->_idx)) {
920 Node *x = clones.find(n->_idx);
921 if (x != NULL)
922 return x;
923 return n;
924 }
925
926 Node *x = NULL; // If required, a clone of 'n'
927 // Check for 'n' being pinned in the backedge.
928 if( n->in(0) && n->in(0) == back_ctrl ) {
929 assert(clones.find(n->_idx) == NULL, "dead loop");
930 assert(n->Opcode() != Op_ShenandoahWriteBarrier, "no wbs yet");
931 x = n->clone(); // Clone a copy of 'n' to preheader
932 clones.push(x, n->_idx);
933 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
934 }
935
936 // Recursive fixup any other input edges into x.
937 // If there are no changes we can just return 'n', otherwise
938 // we need to clone a private copy and change it.
939 for( uint i = 1; i < n->req(); i++ ) {
940 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
941 if( g != n->in(i) ) {
942 if( !x ) {
943 assert(clones.find(n->_idx) == NULL, "dead loop");
944 x = n->clone();
945 clones.push(x, n->_idx);
946 }
947 x->set_req(i, g);
948 }
949 }
950 if( x ) { // x can legally float to pre-header location
951 register_new_node( x, preheader_ctrl );
952 return x;
953 } else { // raise n to cover LCA of uses
954 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
955 }
956 return n;
957 }
958
959 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) {
960 Node* castii = new CastIINode(incr, TypeInt::INT, true);
961 castii->set_req(0, ctrl);
962 register_new_node(castii, ctrl);
963 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) {
964 Node* n = incr->fast_out(i);
965 if (n->is_Phi() && n->in(0) == loop) {
966 int nrep = n->replace_edge(incr, castii);
967 return true;
968 }
969 }
970 return false;
971 }
972
973 //------------------------------insert_pre_post_loops--------------------------
974 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
975 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
976 // alignment. Useful to unroll loops that do no array accesses.
977 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
978
979 #ifndef PRODUCT
980 if (TraceLoopOpts) {
981 if (peel_only)
982 tty->print("PeelMainPost ");
983 else
984 tty->print("PreMainPost ");
985 loop->dump_head();
986 }
987 #endif
988 C->set_major_progress();
989
990 // Find common pieces of the loop being guarded with pre & post loops
991 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
992 assert( main_head->is_normal_loop(), "" );
993 CountedLoopEndNode *main_end = main_head->loopexit();
994 guarantee(main_end != NULL, "no loop exit node");
995 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
996 uint dd_main_head = dom_depth(main_head);
997 uint max = main_head->outcnt();
998
999 Node *pre_header= main_head->in(LoopNode::EntryControl);
1000 Node *init = main_head->init_trip();
1001 Node *incr = main_end ->incr();
1002 Node *limit = main_end ->limit();
1003 Node *stride = main_end ->stride();
1004 Node *cmp = main_end ->cmp_node();
1005 BoolTest::mask b_test = main_end->test_trip();
1006
1007 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
1008 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
1009 if( bol->outcnt() != 1 ) {
1010 bol = bol->clone();
1011 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
1012 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol);
1013 }
1014 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
1015 if( cmp->outcnt() != 1 ) {
1016 cmp = cmp->clone();
1017 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
1018 _igvn.replace_input_of(bol, 1, cmp);
1019 }
1020
1021 //------------------------------
1022 // Step A: Create Post-Loop.
1023 Node* main_exit = main_end->proj_out(false);
1024 assert( main_exit->Opcode() == Op_IfFalse, "" );
1025 int dd_main_exit = dom_depth(main_exit);
1026
1027 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
1028 // loop pre-header illegally has 2 control users (old & new loops).
1029 clone_loop( loop, old_new, dd_main_exit );
1030 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
1031 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
1032 post_head->set_post_loop(main_head);
1033
1034 // Reduce the post-loop trip count.
1035 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1036 post_end->_prob = PROB_FAIR;
1037
1038 // Build the main-loop normal exit.
1039 IfFalseNode *new_main_exit = new IfFalseNode(main_end);
1040 _igvn.register_new_node_with_optimizer( new_main_exit );
1041 set_idom(new_main_exit, main_end, dd_main_exit );
1042 set_loop(new_main_exit, loop->_parent);
1043
1044 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
1045 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
1046 // (the main-loop trip-counter exit value) because we will be changing
1047 // the exit value (via unrolling) so we cannot constant-fold away the zero
1048 // trip guard until all unrolling is done.
1049 Node *zer_opaq = new Opaque1Node(C, incr);
1050 Node *zer_cmp = new CmpINode( zer_opaq, limit );
1051 Node *zer_bol = new BoolNode( zer_cmp, b_test );
1052 register_new_node( zer_opaq, new_main_exit );
1053 register_new_node( zer_cmp , new_main_exit );
1054 register_new_node( zer_bol , new_main_exit );
1055
1056 // Build the IfNode
1057 IfNode *zer_iff = new IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
1058 _igvn.register_new_node_with_optimizer( zer_iff );
1059 set_idom(zer_iff, new_main_exit, dd_main_exit);
1060 set_loop(zer_iff, loop->_parent);
1061
1062 // Plug in the false-path, taken if we need to skip post-loop
1063 _igvn.replace_input_of(main_exit, 0, zer_iff);
1064 set_idom(main_exit, zer_iff, dd_main_exit);
1065 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
1066 // Make the true-path, must enter the post loop
1067 Node *zer_taken = new IfTrueNode( zer_iff );
1068 _igvn.register_new_node_with_optimizer( zer_taken );
1069 set_idom(zer_taken, zer_iff, dd_main_exit);
1070 set_loop(zer_taken, loop->_parent);
1071 // Plug in the true path
1072 _igvn.hash_delete( post_head );
1073 post_head->set_req(LoopNode::EntryControl, zer_taken);
1074 set_idom(post_head, zer_taken, dd_main_exit);
1075
1076 Arena *a = Thread::current()->resource_area();
1077 VectorSet visited(a);
1078 Node_Stack clones(a, main_head->back_control()->outcnt());
1079 // Step A3: Make the fall-in values to the post-loop come from the
1080 // fall-out values of the main-loop.
1081 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
1082 Node* main_phi = main_head->fast_out(i);
1083 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
1084 Node *post_phi = old_new[main_phi->_idx];
1085 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
1086 post_head->init_control(),
1087 main_phi->in(LoopNode::LoopBackControl),
1088 visited, clones);
1089 _igvn.hash_delete(post_phi);
1090 post_phi->set_req( LoopNode::EntryControl, fallmain );
1091 }
1092 }
1093
1094 // Update local caches for next stanza
1095 main_exit = new_main_exit;
1096
1097
1098 //------------------------------
1099 // Step B: Create Pre-Loop.
1100
1101 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1102 // loop pre-header illegally has 2 control users (old & new loops).
1103 clone_loop( loop, old_new, dd_main_head );
1104 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1105 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1106 pre_head->set_pre_loop(main_head);
1107 Node *pre_incr = old_new[incr->_idx];
1108
1109 // Reduce the pre-loop trip count.
1110 pre_end->_prob = PROB_FAIR;
1111
1112 // Find the pre-loop normal exit.
1113 Node* pre_exit = pre_end->proj_out(false);
1114 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1115 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end);
1116 _igvn.register_new_node_with_optimizer( new_pre_exit );
1117 set_idom(new_pre_exit, pre_end, dd_main_head);
1118 set_loop(new_pre_exit, loop->_parent);
1119
1120 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1121 // pre-loop, the main-loop may not execute at all. Later in life this
1122 // zero-trip guard will become the minimum-trip guard when we unroll
1123 // the main-loop.
1124 Node *min_opaq = new Opaque1Node(C, limit);
1125 Node *min_cmp = new CmpINode( pre_incr, min_opaq );
1126 Node *min_bol = new BoolNode( min_cmp, b_test );
1127 register_new_node( min_opaq, new_pre_exit );
1128 register_new_node( min_cmp , new_pre_exit );
1129 register_new_node( min_bol , new_pre_exit );
1130
1131 // Build the IfNode (assume the main-loop is executed always).
1132 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1133 _igvn.register_new_node_with_optimizer( min_iff );
1134 set_idom(min_iff, new_pre_exit, dd_main_head);
1135 set_loop(min_iff, loop->_parent);
1136
1137 // Plug in the false-path, taken if we need to skip main-loop
1138 _igvn.hash_delete( pre_exit );
1139 pre_exit->set_req(0, min_iff);
1140 set_idom(pre_exit, min_iff, dd_main_head);
1141 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1142 // Make the true-path, must enter the main loop
1143 Node *min_taken = new IfTrueNode( min_iff );
1144 _igvn.register_new_node_with_optimizer( min_taken );
1145 set_idom(min_taken, min_iff, dd_main_head);
1146 set_loop(min_taken, loop->_parent);
1147 // Plug in the true path
1148 _igvn.hash_delete( main_head );
1149 main_head->set_req(LoopNode::EntryControl, min_taken);
1150 set_idom(main_head, min_taken, dd_main_head);
1151
1152 visited.Clear();
1153 clones.clear();
1154 // Step B3: Make the fall-in values to the main-loop come from the
1155 // fall-out values of the pre-loop.
1156 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1157 Node* main_phi = main_head->fast_out(i2);
1158 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1159 Node *pre_phi = old_new[main_phi->_idx];
1160 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1161 main_head->init_control(),
1162 pre_phi->in(LoopNode::LoopBackControl),
1163 visited, clones);
1164 _igvn.hash_delete(main_phi);
1165 main_phi->set_req( LoopNode::EntryControl, fallpre );
1166 }
1167 }
1168
1169 // Nodes inside the loop may be control dependent on a predicate
1170 // that was moved before the preloop. If the back branch of the main
1171 // or post loops becomes dead, those nodes won't be dependent on the
1172 // test that guards that loop nest anymore which could lead to an
1173 // incorrect array access because it executes independently of the
1174 // test that was guarding the loop nest. We add a special CastII on
1175 // the if branch that enters the loop, between the input induction
1176 // variable value and the induction variable Phi to preserve correct
1177 // dependencies.
1178
1179 // CastII for the post loop:
1180 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1181 assert(inserted, "no castII inserted");
1182
1183 // CastII for the main loop:
1184 inserted = cast_incr_before_loop(pre_incr, min_taken, main_head);
1185 assert(inserted, "no castII inserted");
1186
1187 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1188 // RCE and alignment may change this later.
1189 Node *cmp_end = pre_end->cmp_node();
1190 assert( cmp_end->in(2) == limit, "" );
1191 Node *pre_limit = new AddINode( init, stride );
1192
1193 // Save the original loop limit in this Opaque1 node for
1194 // use by range check elimination.
1195 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit);
1196
1197 register_new_node( pre_limit, pre_head->in(0) );
1198 register_new_node( pre_opaq , pre_head->in(0) );
1199
1200 // Since no other users of pre-loop compare, I can hack limit directly
1201 assert( cmp_end->outcnt() == 1, "no other users" );
1202 _igvn.hash_delete(cmp_end);
1203 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1204
1205 // Special case for not-equal loop bounds:
1206 // Change pre loop test, main loop test, and the
1207 // main loop guard test to use lt or gt depending on stride
1208 // direction:
1209 // positive stride use <
1210 // negative stride use >
1211 //
1212 // not-equal test is kept for post loop to handle case
1213 // when init > limit when stride > 0 (and reverse).
1214
1215 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1216
1217 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1218 // Modify pre loop end condition
1219 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1220 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test);
1221 register_new_node( new_bol0, pre_head->in(0) );
1222 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0);
1223 // Modify main loop guard condition
1224 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1225 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test);
1226 register_new_node( new_bol1, new_pre_exit );
1227 _igvn.hash_delete(min_iff);
1228 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1229 // Modify main loop end condition
1230 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1231 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test);
1232 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1233 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2);
1234 }
1235
1236 // Flag main loop
1237 main_head->set_main_loop();
1238 if( peel_only ) main_head->set_main_no_pre_loop();
1239
1240 // Subtract a trip count for the pre-loop.
1241 main_head->set_trip_count(main_head->trip_count() - 1);
1242
1243 // It's difficult to be precise about the trip-counts
1244 // for the pre/post loops. They are usually very short,
1245 // so guess that 4 trips is a reasonable value.
1246 post_head->set_profile_trip_cnt(4.0);
1247 pre_head->set_profile_trip_cnt(4.0);
1248
1249 // Now force out all loop-invariant dominating tests. The optimizer
1250 // finds some, but we _know_ they are all useless.
1251 peeled_dom_test_elim(loop,old_new);
1252 loop->record_for_igvn();
1253 }
1254
1255 //------------------------------is_invariant-----------------------------
1256 // Return true if n is invariant
1257 bool IdealLoopTree::is_invariant(Node* n) const {
1258 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1259 if (n_c->is_top()) return false;
1260 return !is_member(_phase->get_loop(n_c));
1261 }
1262
1263
1264 //------------------------------do_unroll--------------------------------------
1265 // Unroll the loop body one step - make each trip do 2 iterations.
1266 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1267 assert(LoopUnrollLimit, "");
1268 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1269 CountedLoopEndNode *loop_end = loop_head->loopexit();
1270 assert(loop_end, "");
1271 #ifndef PRODUCT
1272 if (PrintOpto && VerifyLoopOptimizations) {
1273 tty->print("Unrolling ");
1274 loop->dump_head();
1275 } else if (TraceLoopOpts) {
1276 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1277 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1278 } else {
1279 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1280 }
1281 loop->dump_head();
1282 }
1283
1284 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) {
1285 Arena* arena = Thread::current()->resource_area();
1286 Node_Stack stack(arena, C->live_nodes() >> 2);
1287 Node_List rpo_list;
1288 VectorSet visited(arena);
1289 visited.set(loop_head->_idx);
1290 rpo( loop_head, stack, visited, rpo_list );
1291 dump(loop, rpo_list.size(), rpo_list );
1292 }
1293 #endif
1294
1295 // Remember loop node count before unrolling to detect
1296 // if rounds of unroll,optimize are making progress
1297 loop_head->set_node_count_before_unroll(loop->_body.size());
1298
1299 Node *ctrl = loop_head->in(LoopNode::EntryControl);
1300 Node *limit = loop_head->limit();
1301 Node *init = loop_head->init_trip();
1302 Node *stride = loop_head->stride();
1303
1304 Node *opaq = NULL;
1305 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1306 // Search for zero-trip guard.
1307 assert( loop_head->is_main_loop(), "" );
1308 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1309 Node *iff = ctrl->in(0);
1310 assert( iff->Opcode() == Op_If, "" );
1311 Node *bol = iff->in(1);
1312 assert( bol->Opcode() == Op_Bool, "" );
1313 Node *cmp = bol->in(1);
1314 assert( cmp->Opcode() == Op_CmpI, "" );
1315 opaq = cmp->in(2);
1316 // Occasionally it's possible for a zero-trip guard Opaque1 node to be
1317 // optimized away and then another round of loop opts attempted.
1318 // We can not optimize this particular loop in that case.
1319 if (opaq->Opcode() != Op_Opaque1)
1320 return; // Cannot find zero-trip guard! Bail out!
1321 // Zero-trip test uses an 'opaque' node which is not shared.
1322 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1323 }
1324
1325 C->set_major_progress();
1326
1327 Node* new_limit = NULL;
1328 if (UnrollLimitCheck) {
1329 int stride_con = stride->get_int();
1330 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1331 uint old_trip_count = loop_head->trip_count();
1332 // Verify that unroll policy result is still valid.
1333 assert(old_trip_count > 1 &&
1334 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1335
1336 // Adjust loop limit to keep valid iterations number after unroll.
1337 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1338 // which may overflow.
1339 if (!adjust_min_trip) {
1340 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1341 "odd trip count for maximally unroll");
1342 // Don't need to adjust limit for maximally unroll since trip count is even.
1343 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1344 // Loop's limit is constant. Loop's init could be constant when pre-loop
1345 // become peeled iteration.
1346 jlong init_con = init->get_int();
1347 // We can keep old loop limit if iterations count stays the same:
1348 // old_trip_count == new_trip_count * 2
1349 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1350 // so we also don't need to adjust zero trip test.
1351 jlong limit_con = limit->get_int();
1352 // (stride_con*2) not overflow since stride_con <= 8.
1353 int new_stride_con = stride_con * 2;
1354 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1355 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1356 // New trip count should satisfy next conditions.
1357 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1358 uint new_trip_count = (uint)trip_count;
1359 adjust_min_trip = (old_trip_count != new_trip_count*2);
1360 }
1361
1362 if (adjust_min_trip) {
1363 // Step 2: Adjust the trip limit if it is called for.
1364 // The adjustment amount is -stride. Need to make sure if the
1365 // adjustment underflows or overflows, then the main loop is skipped.
1366 Node* cmp = loop_end->cmp_node();
1367 assert(cmp->in(2) == limit, "sanity");
1368 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1369
1370 // Verify that policy_unroll result is still valid.
1371 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1372 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1373 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1374
1375 if (limit->is_Con()) {
1376 // The check in policy_unroll and the assert above guarantee
1377 // no underflow if limit is constant.
1378 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1379 set_ctrl(new_limit, C->root());
1380 } else {
1381 // Limit is not constant.
1382 if (loop_head->unrolled_count() == 1) { // only for first unroll
1383 // Separate limit by Opaque node in case it is an incremented
1384 // variable from previous loop to avoid using pre-incremented
1385 // value which could increase register pressure.
1386 // Otherwise reorg_offsets() optimization will create a separate
1387 // Opaque node for each use of trip-counter and as result
1388 // zero trip guard limit will be different from loop limit.
1389 assert(has_ctrl(opaq), "should have it");
1390 Node* opaq_ctrl = get_ctrl(opaq);
1391 limit = new Opaque2Node( C, limit );
1392 register_new_node( limit, opaq_ctrl );
1393 }
1394 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) ||
1395 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) {
1396 // No underflow.
1397 new_limit = new SubINode(limit, stride);
1398 } else {
1399 // (limit - stride) may underflow.
1400 // Clamp the adjustment value with MININT or MAXINT:
1401 //
1402 // new_limit = limit-stride
1403 // if (stride > 0)
1404 // new_limit = (limit < new_limit) ? MININT : new_limit;
1405 // else
1406 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1407 //
1408 BoolTest::mask bt = loop_end->test_trip();
1409 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1410 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1411 set_ctrl(adj_max, C->root());
1412 Node* old_limit = NULL;
1413 Node* adj_limit = NULL;
1414 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1415 if (loop_head->unrolled_count() > 1 &&
1416 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1417 limit->in(CMoveNode::IfTrue) == adj_max &&
1418 bol->as_Bool()->_test._test == bt &&
1419 bol->in(1)->Opcode() == Op_CmpI &&
1420 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1421 // Loop was unrolled before.
1422 // Optimize the limit to avoid nested CMove:
1423 // use original limit as old limit.
1424 old_limit = bol->in(1)->in(1);
1425 // Adjust previous adjusted limit.
1426 adj_limit = limit->in(CMoveNode::IfFalse);
1427 adj_limit = new SubINode(adj_limit, stride);
1428 } else {
1429 old_limit = limit;
1430 adj_limit = new SubINode(limit, stride);
1431 }
1432 assert(old_limit != NULL && adj_limit != NULL, "");
1433 register_new_node( adj_limit, ctrl ); // adjust amount
1434 Node* adj_cmp = new CmpINode(old_limit, adj_limit);
1435 register_new_node( adj_cmp, ctrl );
1436 Node* adj_bool = new BoolNode(adj_cmp, bt);
1437 register_new_node( adj_bool, ctrl );
1438 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1439 }
1440 register_new_node(new_limit, ctrl);
1441 }
1442 assert(new_limit != NULL, "");
1443 // Replace in loop test.
1444 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1445 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1446 // Don't need to create new test since only one user.
1447 _igvn.hash_delete(cmp);
1448 cmp->set_req(2, new_limit);
1449 } else {
1450 // Create new test since it is shared.
1451 Node* ctrl2 = loop_end->in(0);
1452 Node* cmp2 = cmp->clone();
1453 cmp2->set_req(2, new_limit);
1454 register_new_node(cmp2, ctrl2);
1455 Node* bol2 = loop_end->in(1)->clone();
1456 bol2->set_req(1, cmp2);
1457 register_new_node(bol2, ctrl2);
1458 _igvn.replace_input_of(loop_end, 1, bol2);
1459 }
1460 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1461 // Make it a 1-trip test (means at least 2 trips).
1462
1463 // Guard test uses an 'opaque' node which is not shared. Hence I
1464 // can edit it's inputs directly. Hammer in the new limit for the
1465 // minimum-trip guard.
1466 assert(opaq->outcnt() == 1, "");
1467 _igvn.replace_input_of(opaq, 1, new_limit);
1468 }
1469
1470 // Adjust max trip count. The trip count is intentionally rounded
1471 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1472 // the main, unrolled, part of the loop will never execute as it is protected
1473 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1474 // and later determined that part of the unrolled loop was dead.
1475 loop_head->set_trip_count(old_trip_count / 2);
1476
1477 // Double the count of original iterations in the unrolled loop body.
1478 loop_head->double_unrolled_count();
1479
1480 } else { // LoopLimitCheck
1481
1482 // Adjust max trip count. The trip count is intentionally rounded
1483 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1484 // the main, unrolled, part of the loop will never execute as it is protected
1485 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1486 // and later determined that part of the unrolled loop was dead.
1487 loop_head->set_trip_count(loop_head->trip_count() / 2);
1488
1489 // Double the count of original iterations in the unrolled loop body.
1490 loop_head->double_unrolled_count();
1491
1492 // -----------
1493 // Step 2: Cut back the trip counter for an unroll amount of 2.
1494 // Loop will normally trip (limit - init)/stride_con. Since it's a
1495 // CountedLoop this is exact (stride divides limit-init exactly).
1496 // We are going to double the loop body, so we want to knock off any
1497 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
1498 Node *span = new SubINode( limit, init );
1499 register_new_node( span, ctrl );
1500 Node *trip = new DivINode( 0, span, stride );
1501 register_new_node( trip, ctrl );
1502 Node *mtwo = _igvn.intcon(-2);
1503 set_ctrl(mtwo, C->root());
1504 Node *rond = new AndINode( trip, mtwo );
1505 register_new_node( rond, ctrl );
1506 Node *spn2 = new MulINode( rond, stride );
1507 register_new_node( spn2, ctrl );
1508 new_limit = new AddINode( spn2, init );
1509 register_new_node( new_limit, ctrl );
1510
1511 // Hammer in the new limit
1512 Node *ctrl2 = loop_end->in(0);
1513 Node *cmp2 = new CmpINode( loop_head->incr(), new_limit );
1514 register_new_node( cmp2, ctrl2 );
1515 Node *bol2 = new BoolNode( cmp2, loop_end->test_trip() );
1516 register_new_node( bol2, ctrl2 );
1517 _igvn.replace_input_of(loop_end, CountedLoopEndNode::TestValue, bol2);
1518
1519 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1520 // Make it a 1-trip test (means at least 2 trips).
1521 if( adjust_min_trip ) {
1522 assert( new_limit != NULL, "" );
1523 // Guard test uses an 'opaque' node which is not shared. Hence I
1524 // can edit it's inputs directly. Hammer in the new limit for the
1525 // minimum-trip guard.
1526 assert( opaq->outcnt() == 1, "" );
1527 _igvn.hash_delete(opaq);
1528 opaq->set_req(1, new_limit);
1529 }
1530 } // LoopLimitCheck
1531
1532 // ---------
1533 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1534 // represents the odd iterations; since the loop trips an even number of
1535 // times its backedge is never taken. Kill the backedge.
1536 uint dd = dom_depth(loop_head);
1537 clone_loop( loop, old_new, dd );
1538
1539 // Make backedges of the clone equal to backedges of the original.
1540 // Make the fall-in from the original come from the fall-out of the clone.
1541 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1542 Node* phi = loop_head->fast_out(j);
1543 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1544 Node *newphi = old_new[phi->_idx];
1545 _igvn.hash_delete( phi );
1546 _igvn.hash_delete( newphi );
1547
1548 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1549 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1550 phi ->set_req(LoopNode::LoopBackControl, C->top());
1551 }
1552 }
1553 Node *clone_head = old_new[loop_head->_idx];
1554 _igvn.hash_delete( clone_head );
1555 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1556 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1557 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1558 loop->_head = clone_head; // New loop header
1559
1560 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1561 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1562
1563 // Kill the clone's backedge
1564 Node *newcle = old_new[loop_end->_idx];
1565 _igvn.hash_delete( newcle );
1566 Node *one = _igvn.intcon(1);
1567 set_ctrl(one, C->root());
1568 newcle->set_req(1, one);
1569 // Force clone into same loop body
1570 uint max = loop->_body.size();
1571 for( uint k = 0; k < max; k++ ) {
1572 Node *old = loop->_body.at(k);
1573 Node *nnn = old_new[old->_idx];
1574 loop->_body.push(nnn);
1575 if (!has_ctrl(old))
1576 set_loop(nnn, loop);
1577 }
1578
1579 loop->record_for_igvn();
1580
1581 #ifndef PRODUCT
1582 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) {
1583 tty->print("\nnew loop after unroll\n"); loop->dump_head();
1584 for (uint i = 0; i < loop->_body.size(); i++) {
1585 loop->_body.at(i)->dump();
1586 }
1587 if(C->clone_map().is_debug()) {
1588 tty->print("\nCloneMap\n");
1589 Dict* dict = C->clone_map().dict();
1590 DictI i(dict);
1591 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size());
1592 for (int ii = 0; i.test(); ++i, ++ii) {
1593 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key));
1594 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen());
1595 if (ii % 10 == 9) {
1596 tty->print_cr(" ");
1597 }
1598 }
1599 tty->print_cr(" ");
1600 }
1601 }
1602 #endif
1603
1604 }
1605
1606 //------------------------------do_maximally_unroll----------------------------
1607
1608 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1609 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1610 assert(cl->has_exact_trip_count(), "trip count is not exact");
1611 assert(cl->trip_count() > 0, "");
1612 #ifndef PRODUCT
1613 if (TraceLoopOpts) {
1614 tty->print("MaxUnroll %d ", cl->trip_count());
1615 loop->dump_head();
1616 }
1617 #endif
1618
1619 // If loop is tripping an odd number of times, peel odd iteration
1620 if ((cl->trip_count() & 1) == 1) {
1621 do_peeling(loop, old_new);
1622 }
1623
1624 // Now its tripping an even number of times remaining. Double loop body.
1625 // Do not adjust pre-guards; they are not needed and do not exist.
1626 if (cl->trip_count() > 0) {
1627 assert((cl->trip_count() & 1) == 0, "missed peeling");
1628 do_unroll(loop, old_new, false);
1629 }
1630 }
1631
1632 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) {
1633 if (SuperWordReductions == false) return;
1634
1635 CountedLoopNode* loop_head = loop->_head->as_CountedLoop();
1636 if (loop_head->unrolled_count() > 1) {
1637 return;
1638 }
1639
1640 Node* trip_phi = loop_head->phi();
1641 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) {
1642 Node* phi = loop_head->fast_out(i);
1643 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) {
1644 // For definitions which are loop inclusive and not tripcounts.
1645 Node* def_node = phi->in(LoopNode::LoopBackControl);
1646
1647 if (def_node != NULL) {
1648 Node* n_ctrl = get_ctrl(def_node);
1649 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) {
1650 // Now test it to see if it fits the standard pattern for a reduction operator.
1651 int opc = def_node->Opcode();
1652 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) {
1653 if (!def_node->is_reduction()) { // Not marked yet
1654 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it
1655 bool ok = false;
1656 for (unsigned j = 1; j < def_node->req(); j++) {
1657 Node* in = def_node->in(j);
1658 if (in == phi) {
1659 ok = true;
1660 break;
1661 }
1662 }
1663
1664 // do nothing if we did not match the initial criteria
1665 if (ok == false) {
1666 continue;
1667 }
1668
1669 // The result of the reduction must not be used in the loop
1670 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) {
1671 Node* u = def_node->fast_out(i);
1672 if (has_ctrl(u) && !loop->is_member(get_loop(get_ctrl(u)))) {
1673 continue;
1674 }
1675 if (u == phi) {
1676 continue;
1677 }
1678 ok = false;
1679 }
1680
1681 // iff the uses conform
1682 if (ok) {
1683 def_node->add_flag(Node::Flag_is_reduction);
1684 loop_head->mark_has_reductions();
1685 }
1686 }
1687 }
1688 }
1689 }
1690 }
1691 }
1692 }
1693
1694 //------------------------------dominates_backedge---------------------------------
1695 // Returns true if ctrl is executed on every complete iteration
1696 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1697 assert(ctrl->is_CFG(), "must be control");
1698 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1699 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1700 }
1701
1702 //------------------------------adjust_limit-----------------------------------
1703 // Helper function for add_constraint().
1704 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1705 // Compute "I :: (limit-offset)/scale"
1706 Node *con = new SubINode(rc_limit, offset);
1707 register_new_node(con, pre_ctrl);
1708 Node *X = new DivINode(0, con, scale);
1709 register_new_node(X, pre_ctrl);
1710
1711 // Adjust loop limit
1712 loop_limit = (stride_con > 0)
1713 ? (Node*)(new MinINode(loop_limit, X))
1714 : (Node*)(new MaxINode(loop_limit, X));
1715 register_new_node(loop_limit, pre_ctrl);
1716 return loop_limit;
1717 }
1718
1719 //------------------------------add_constraint---------------------------------
1720 // Constrain the main loop iterations so the conditions:
1721 // low_limit <= scale_con * I + offset < upper_limit
1722 // always holds true. That is, either increase the number of iterations in
1723 // the pre-loop or the post-loop until the condition holds true in the main
1724 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1725 // stride and scale are constants (offset and limit often are).
1726 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 ) {
1727 // For positive stride, the pre-loop limit always uses a MAX function
1728 // and the main loop a MIN function. For negative stride these are
1729 // reversed.
1730
1731 // Also for positive stride*scale the affine function is increasing, so the
1732 // pre-loop must check for underflow and the post-loop for overflow.
1733 // Negative stride*scale reverses this; pre-loop checks for overflow and
1734 // post-loop for underflow.
1735
1736 Node *scale = _igvn.intcon(scale_con);
1737 set_ctrl(scale, C->root());
1738
1739 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1740 // The overflow limit: scale*I+offset < upper_limit
1741 // For main-loop compute
1742 // ( if (scale > 0) /* and stride > 0 */
1743 // I < (upper_limit-offset)/scale
1744 // else /* scale < 0 and stride < 0 */
1745 // I > (upper_limit-offset)/scale
1746 // )
1747 //
1748 // (upper_limit-offset) may overflow or underflow.
1749 // But it is fine since main loop will either have
1750 // less iterations or will be skipped in such case.
1751 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1752
1753 // The underflow limit: low_limit <= scale*I+offset.
1754 // For pre-loop compute
1755 // NOT(scale*I+offset >= low_limit)
1756 // scale*I+offset < low_limit
1757 // ( if (scale > 0) /* and stride > 0 */
1758 // I < (low_limit-offset)/scale
1759 // else /* scale < 0 and stride < 0 */
1760 // I > (low_limit-offset)/scale
1761 // )
1762
1763 if (low_limit->get_int() == -max_jint) {
1764 if (!RangeLimitCheck) return;
1765 // We need this guard when scale*pre_limit+offset >= limit
1766 // due to underflow. So we need execute pre-loop until
1767 // scale*I+offset >= min_int. But (min_int-offset) will
1768 // underflow when offset > 0 and X will be > original_limit
1769 // when stride > 0. To avoid it we replace positive offset with 0.
1770 //
1771 // Also (min_int+1 == -max_int) is used instead of min_int here
1772 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1773 Node* shift = _igvn.intcon(31);
1774 set_ctrl(shift, C->root());
1775 Node* sign = new RShiftINode(offset, shift);
1776 register_new_node(sign, pre_ctrl);
1777 offset = new AndINode(offset, sign);
1778 register_new_node(offset, pre_ctrl);
1779 } else {
1780 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1781 // The only problem we have here when offset == min_int
1782 // since (0-min_int) == min_int. It may be fine for stride > 0
1783 // but for stride < 0 X will be < original_limit. To avoid it
1784 // max(pre_limit, original_limit) is used in do_range_check().
1785 }
1786 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1787 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1788
1789 } else { // stride_con*scale_con < 0
1790 // For negative stride*scale pre-loop checks for overflow and
1791 // post-loop for underflow.
1792 //
1793 // The overflow limit: scale*I+offset < upper_limit
1794 // For pre-loop compute
1795 // NOT(scale*I+offset < upper_limit)
1796 // scale*I+offset >= upper_limit
1797 // scale*I+offset+1 > upper_limit
1798 // ( if (scale < 0) /* and stride > 0 */
1799 // I < (upper_limit-(offset+1))/scale
1800 // else /* scale > 0 and stride < 0 */
1801 // I > (upper_limit-(offset+1))/scale
1802 // )
1803 //
1804 // (upper_limit-offset-1) may underflow or overflow.
1805 // To avoid it min(pre_limit, original_limit) is used
1806 // in do_range_check() for stride > 0 and max() for < 0.
1807 Node *one = _igvn.intcon(1);
1808 set_ctrl(one, C->root());
1809
1810 Node *plus_one = new AddINode(offset, one);
1811 register_new_node( plus_one, pre_ctrl );
1812 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1813 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1814
1815 if (low_limit->get_int() == -max_jint) {
1816 if (!RangeLimitCheck) return;
1817 // We need this guard when scale*main_limit+offset >= limit
1818 // due to underflow. So we need execute main-loop while
1819 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1820 // underflow when (offset+1) > 0 and X will be < main_limit
1821 // when scale < 0 (and stride > 0). To avoid it we replace
1822 // positive (offset+1) with 0.
1823 //
1824 // Also (min_int+1 == -max_int) is used instead of min_int here
1825 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1826 Node* shift = _igvn.intcon(31);
1827 set_ctrl(shift, C->root());
1828 Node* sign = new RShiftINode(plus_one, shift);
1829 register_new_node(sign, pre_ctrl);
1830 plus_one = new AndINode(plus_one, sign);
1831 register_new_node(plus_one, pre_ctrl);
1832 } else {
1833 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1834 // The only problem we have here when offset == max_int
1835 // since (max_int+1) == min_int and (0-min_int) == min_int.
1836 // But it is fine since main loop will either have
1837 // less iterations or will be skipped in such case.
1838 }
1839 // The underflow limit: low_limit <= scale*I+offset.
1840 // For main-loop compute
1841 // scale*I+offset+1 > low_limit
1842 // ( if (scale < 0) /* and stride > 0 */
1843 // I < (low_limit-(offset+1))/scale
1844 // else /* scale > 0 and stride < 0 */
1845 // I > (low_limit-(offset+1))/scale
1846 // )
1847
1848 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1849 }
1850 }
1851
1852
1853 //------------------------------is_scaled_iv---------------------------------
1854 // Return true if exp is a constant times an induction var
1855 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1856 if (exp == iv) {
1857 if (p_scale != NULL) {
1858 *p_scale = 1;
1859 }
1860 return true;
1861 }
1862 int opc = exp->Opcode();
1863 if (opc == Op_MulI) {
1864 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1865 if (p_scale != NULL) {
1866 *p_scale = exp->in(2)->get_int();
1867 }
1868 return true;
1869 }
1870 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1871 if (p_scale != NULL) {
1872 *p_scale = exp->in(1)->get_int();
1873 }
1874 return true;
1875 }
1876 } else if (opc == Op_LShiftI) {
1877 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1878 if (p_scale != NULL) {
1879 *p_scale = 1 << exp->in(2)->get_int();
1880 }
1881 return true;
1882 }
1883 }
1884 return false;
1885 }
1886
1887 //-----------------------------is_scaled_iv_plus_offset------------------------------
1888 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1889 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1890 if (is_scaled_iv(exp, iv, p_scale)) {
1891 if (p_offset != NULL) {
1892 Node *zero = _igvn.intcon(0);
1893 set_ctrl(zero, C->root());
1894 *p_offset = zero;
1895 }
1896 return true;
1897 }
1898 int opc = exp->Opcode();
1899 if (opc == Op_AddI) {
1900 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1901 if (p_offset != NULL) {
1902 *p_offset = exp->in(2);
1903 }
1904 return true;
1905 }
1906 if (exp->in(2)->is_Con()) {
1907 Node* offset2 = NULL;
1908 if (depth < 2 &&
1909 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1910 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1911 if (p_offset != NULL) {
1912 Node *ctrl_off2 = get_ctrl(offset2);
1913 Node* offset = new AddINode(offset2, exp->in(2));
1914 register_new_node(offset, ctrl_off2);
1915 *p_offset = offset;
1916 }
1917 return true;
1918 }
1919 }
1920 } else if (opc == Op_SubI) {
1921 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1922 if (p_offset != NULL) {
1923 Node *zero = _igvn.intcon(0);
1924 set_ctrl(zero, C->root());
1925 Node *ctrl_off = get_ctrl(exp->in(2));
1926 Node* offset = new SubINode(zero, exp->in(2));
1927 register_new_node(offset, ctrl_off);
1928 *p_offset = offset;
1929 }
1930 return true;
1931 }
1932 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1933 if (p_offset != NULL) {
1934 *p_scale *= -1;
1935 *p_offset = exp->in(1);
1936 }
1937 return true;
1938 }
1939 }
1940 return false;
1941 }
1942
1943 //------------------------------do_range_check---------------------------------
1944 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1945 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1946 #ifndef PRODUCT
1947 if (PrintOpto && VerifyLoopOptimizations) {
1948 tty->print("Range Check Elimination ");
1949 loop->dump_head();
1950 } else if (TraceLoopOpts) {
1951 tty->print("RangeCheck ");
1952 loop->dump_head();
1953 }
1954 #endif
1955 assert(RangeCheckElimination, "");
1956 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1957 assert(cl->is_main_loop(), "");
1958
1959 // protect against stride not being a constant
1960 if (!cl->stride_is_con())
1961 return;
1962
1963 // Find the trip counter; we are iteration splitting based on it
1964 Node *trip_counter = cl->phi();
1965 // Find the main loop limit; we will trim it's iterations
1966 // to not ever trip end tests
1967 Node *main_limit = cl->limit();
1968
1969 // Need to find the main-loop zero-trip guard
1970 Node *ctrl = cl->in(LoopNode::EntryControl);
1971 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1972 Node *iffm = ctrl->in(0);
1973 assert(iffm->Opcode() == Op_If, "");
1974 Node *bolzm = iffm->in(1);
1975 assert(bolzm->Opcode() == Op_Bool, "");
1976 Node *cmpzm = bolzm->in(1);
1977 assert(cmpzm->is_Cmp(), "");
1978 Node *opqzm = cmpzm->in(2);
1979 // Can not optimize a loop if zero-trip Opaque1 node is optimized
1980 // away and then another round of loop opts attempted.
1981 if (opqzm->Opcode() != Op_Opaque1)
1982 return;
1983 assert(opqzm->in(1) == main_limit, "do not understand situation");
1984
1985 // Find the pre-loop limit; we will expand its iterations to
1986 // not ever trip low tests.
1987 Node *p_f = iffm->in(0);
1988 // pre loop may have been optimized out
1989 if (p_f->Opcode() != Op_IfFalse) {
1990 return;
1991 }
1992 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1993 assert(pre_end->loopnode()->is_pre_loop(), "");
1994 Node *pre_opaq1 = pre_end->limit();
1995 // Occasionally it's possible for a pre-loop Opaque1 node to be
1996 // optimized away and then another round of loop opts attempted.
1997 // We can not optimize this particular loop in that case.
1998 if (pre_opaq1->Opcode() != Op_Opaque1)
1999 return;
2000 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2001 Node *pre_limit = pre_opaq->in(1);
2002
2003 // Where do we put new limit calculations
2004 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2005
2006 // Ensure the original loop limit is available from the
2007 // pre-loop Opaque1 node.
2008 Node *orig_limit = pre_opaq->original_loop_limit();
2009 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
2010 return;
2011
2012 // Must know if its a count-up or count-down loop
2013
2014 int stride_con = cl->stride_con();
2015 Node *zero = _igvn.intcon(0);
2016 Node *one = _igvn.intcon(1);
2017 // Use symmetrical int range [-max_jint,max_jint]
2018 Node *mini = _igvn.intcon(-max_jint);
2019 set_ctrl(zero, C->root());
2020 set_ctrl(one, C->root());
2021 set_ctrl(mini, C->root());
2022
2023 // Range checks that do not dominate the loop backedge (ie.
2024 // conditionally executed) can lengthen the pre loop limit beyond
2025 // the original loop limit. To prevent this, the pre limit is
2026 // (for stride > 0) MINed with the original loop limit (MAXed
2027 // stride < 0) when some range_check (rc) is conditionally
2028 // executed.
2029 bool conditional_rc = false;
2030
2031 // Check loop body for tests of trip-counter plus loop-invariant vs
2032 // loop-invariant.
2033 for( uint i = 0; i < loop->_body.size(); i++ ) {
2034 Node *iff = loop->_body[i];
2035 if( iff->Opcode() == Op_If ) { // Test?
2036
2037 // Test is an IfNode, has 2 projections. If BOTH are in the loop
2038 // we need loop unswitching instead of iteration splitting.
2039 Node *exit = loop->is_loop_exit(iff);
2040 if( !exit ) continue;
2041 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
2042
2043 // Get boolean condition to test
2044 Node *i1 = iff->in(1);
2045 if( !i1->is_Bool() ) continue;
2046 BoolNode *bol = i1->as_Bool();
2047 BoolTest b_test = bol->_test;
2048 // Flip sense of test if exit condition is flipped
2049 if( flip )
2050 b_test = b_test.negate();
2051
2052 // Get compare
2053 Node *cmp = bol->in(1);
2054
2055 // Look for trip_counter + offset vs limit
2056 Node *rc_exp = cmp->in(1);
2057 Node *limit = cmp->in(2);
2058 jint scale_con= 1; // Assume trip counter not scaled
2059
2060 Node *limit_c = get_ctrl(limit);
2061 if( loop->is_member(get_loop(limit_c) ) ) {
2062 // Compare might have operands swapped; commute them
2063 b_test = b_test.commute();
2064 rc_exp = cmp->in(2);
2065 limit = cmp->in(1);
2066 limit_c = get_ctrl(limit);
2067 if( loop->is_member(get_loop(limit_c) ) )
2068 continue; // Both inputs are loop varying; cannot RCE
2069 }
2070 // Here we know 'limit' is loop invariant
2071
2072 // 'limit' maybe pinned below the zero trip test (probably from a
2073 // previous round of rce), in which case, it can't be used in the
2074 // zero trip test expression which must occur before the zero test's if.
2075 if( limit_c == ctrl ) {
2076 continue; // Don't rce this check but continue looking for other candidates.
2077 }
2078
2079 // Check for scaled induction variable plus an offset
2080 Node *offset = NULL;
2081
2082 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
2083 continue;
2084 }
2085
2086 Node *offset_c = get_ctrl(offset);
2087 if( loop->is_member( get_loop(offset_c) ) )
2088 continue; // Offset is not really loop invariant
2089 // Here we know 'offset' is loop invariant.
2090
2091 // As above for the 'limit', the 'offset' maybe pinned below the
2092 // zero trip test.
2093 if( offset_c == ctrl ) {
2094 continue; // Don't rce this check but continue looking for other candidates.
2095 }
2096 #ifdef ASSERT
2097 if (TraceRangeLimitCheck) {
2098 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
2099 bol->dump(2);
2100 }
2101 #endif
2102 // At this point we have the expression as:
2103 // scale_con * trip_counter + offset :: limit
2104 // where scale_con, offset and limit are loop invariant. Trip_counter
2105 // monotonically increases by stride_con, a constant. Both (or either)
2106 // stride_con and scale_con can be negative which will flip about the
2107 // sense of the test.
2108
2109 // Adjust pre and main loop limits to guard the correct iteration set
2110 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
2111 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
2112 // The underflow and overflow limits: 0 <= scale*I+offset < limit
2113 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
2114 if (!conditional_rc) {
2115 // (0-offset)/scale could be outside of loop iterations range.
2116 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
2117 }
2118 } else {
2119 #ifndef PRODUCT
2120 if( PrintOpto )
2121 tty->print_cr("missed RCE opportunity");
2122 #endif
2123 continue; // In release mode, ignore it
2124 }
2125 } else { // Otherwise work on normal compares
2126 switch( b_test._test ) {
2127 case BoolTest::gt:
2128 // Fall into GE case
2129 case BoolTest::ge:
2130 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
2131 scale_con = -scale_con;
2132 offset = new SubINode( zero, offset );
2133 register_new_node( offset, pre_ctrl );
2134 limit = new SubINode( zero, limit );
2135 register_new_node( limit, pre_ctrl );
2136 // Fall into LE case
2137 case BoolTest::le:
2138 if (b_test._test != BoolTest::gt) {
2139 // Convert X <= Y to X < Y+1
2140 limit = new AddINode( limit, one );
2141 register_new_node( limit, pre_ctrl );
2142 }
2143 // Fall into LT case
2144 case BoolTest::lt:
2145 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
2146 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
2147 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
2148 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
2149 if (!conditional_rc) {
2150 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
2151 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
2152 // still be outside of loop range.
2153 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
2154 }
2155 break;
2156 default:
2157 #ifndef PRODUCT
2158 if( PrintOpto )
2159 tty->print_cr("missed RCE opportunity");
2160 #endif
2161 continue; // Unhandled case
2162 }
2163 }
2164
2165 // Kill the eliminated test
2166 C->set_major_progress();
2167 Node *kill_con = _igvn.intcon( 1-flip );
2168 set_ctrl(kill_con, C->root());
2169 _igvn.replace_input_of(iff, 1, kill_con);
2170 // Find surviving projection
2171 assert(iff->is_If(), "");
2172 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2173 // Find loads off the surviving projection; remove their control edge
2174 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2175 Node* cd = dp->fast_out(i); // Control-dependent node
2176 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop
2177 // Allow the load to float around in the loop, or before it
2178 // but NOT before the pre-loop.
2179 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
2180 --i;
2181 --imax;
2182 }
2183 }
2184
2185 } // End of is IF
2186
2187 }
2188
2189 // Update loop limits
2190 if (conditional_rc) {
2191 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit)
2192 : (Node*)new MaxINode(pre_limit, orig_limit);
2193 register_new_node(pre_limit, pre_ctrl);
2194 }
2195 _igvn.replace_input_of(pre_opaq, 1, pre_limit);
2196
2197 // Note:: we are making the main loop limit no longer precise;
2198 // need to round up based on stride.
2199 cl->set_nonexact_trip_count();
2200 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
2201 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
2202 // Hopefully, compiler will optimize for powers of 2.
2203 Node *ctrl = get_ctrl(main_limit);
2204 Node *stride = cl->stride();
2205 Node *init = cl->init_trip()->uncast();
2206 Node *span = new SubINode(main_limit,init);
2207 register_new_node(span,ctrl);
2208 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
2209 Node *add = new AddINode(span,rndup);
2210 register_new_node(add,ctrl);
2211 Node *div = new DivINode(0,add,stride);
2212 register_new_node(div,ctrl);
2213 Node *mul = new MulINode(div,stride);
2214 register_new_node(mul,ctrl);
2215 Node *newlim = new AddINode(mul,init);
2216 register_new_node(newlim,ctrl);
2217 main_limit = newlim;
2218 }
2219
2220 Node *main_cle = cl->loopexit();
2221 Node *main_bol = main_cle->in(1);
2222 // Hacking loop bounds; need private copies of exit test
2223 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2224 main_bol = main_bol->clone();// Clone a private BoolNode
2225 register_new_node( main_bol, main_cle->in(0) );
2226 _igvn.replace_input_of(main_cle, 1, main_bol);
2227 }
2228 Node *main_cmp = main_bol->in(1);
2229 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2230 main_cmp = main_cmp->clone();// Clone a private CmpNode
2231 register_new_node( main_cmp, main_cle->in(0) );
2232 _igvn.replace_input_of(main_bol, 1, main_cmp);
2233 }
2234 // Hack the now-private loop bounds
2235 _igvn.replace_input_of(main_cmp, 2, main_limit);
2236 // The OpaqueNode is unshared by design
2237 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2238 _igvn.replace_input_of(opqzm, 1, main_limit);
2239 }
2240
2241 //------------------------------DCE_loop_body----------------------------------
2242 // Remove simplistic dead code from loop body
2243 void IdealLoopTree::DCE_loop_body() {
2244 for( uint i = 0; i < _body.size(); i++ )
2245 if( _body.at(i)->outcnt() == 0 )
2246 _body.map( i--, _body.pop() );
2247 }
2248
2249
2250 //------------------------------adjust_loop_exit_prob--------------------------
2251 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2252 // Replace with a 1-in-10 exit guess.
2253 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2254 Node *test = tail();
2255 while( test != _head ) {
2256 uint top = test->Opcode();
2257 if( top == Op_IfTrue || top == Op_IfFalse ) {
2258 int test_con = ((ProjNode*)test)->_con;
2259 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2260 IfNode *iff = test->in(0)->as_If();
2261 if( iff->outcnt() == 2 ) { // Ignore dead tests
2262 Node *bol = iff->in(1);
2263 if( bol && bol->req() > 1 && bol->in(1) &&
2264 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2265 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2266 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2267 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2268 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2269 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2270 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2271 return; // Allocation loops RARELY take backedge
2272 // Find the OTHER exit path from the IF
2273 Node* ex = iff->proj_out(1-test_con);
2274 float p = iff->_prob;
2275 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2276 if( top == Op_IfTrue ) {
2277 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2278 iff->_prob = PROB_STATIC_FREQUENT;
2279 }
2280 } else {
2281 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2282 iff->_prob = PROB_STATIC_INFREQUENT;
2283 }
2284 }
2285 }
2286 }
2287 }
2288 test = phase->idom(test);
2289 }
2290 }
2291
2292 #ifdef ASSERT
2293 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) {
2294 Node *ctrl = cl->in(LoopNode::EntryControl);
2295 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
2296 Node *iffm = ctrl->in(0);
2297 assert(iffm->Opcode() == Op_If, "");
2298 Node *p_f = iffm->in(0);
2299 assert(p_f->Opcode() == Op_IfFalse, "");
2300 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2301 assert(pre_end->loopnode()->is_pre_loop(), "");
2302 return pre_end->loopnode();
2303 }
2304 #endif
2305
2306 // Remove the main and post loops and make the pre loop execute all
2307 // iterations. Useful when the pre loop is found empty.
2308 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) {
2309 CountedLoopEndNode* pre_end = cl->loopexit();
2310 Node* pre_cmp = pre_end->cmp_node();
2311 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) {
2312 // Only safe to remove the main loop if the compiler optimized it
2313 // out based on an unknown number of iterations
2314 return;
2315 }
2316
2317 // Can we find the main loop?
2318 if (_next == NULL) {
2319 return;
2320 }
2321
2322 Node* next_head = _next->_head;
2323 if (!next_head->is_CountedLoop()) {
2324 return;
2325 }
2326
2327 CountedLoopNode* main_head = next_head->as_CountedLoop();
2328 if (!main_head->is_main_loop()) {
2329 return;
2330 }
2331
2332 assert(locate_pre_from_main(main_head) == cl, "bad main loop");
2333 Node* main_iff = main_head->in(LoopNode::EntryControl)->in(0);
2334
2335 // Remove the Opaque1Node of the pre loop and make it execute all iterations
2336 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2));
2337 // Remove the Opaque1Node of the main loop so it can be optimized out
2338 Node* main_cmp = main_iff->in(1)->in(1);
2339 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?");
2340 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1));
2341 }
2342
2343 //------------------------------policy_do_remove_empty_loop--------------------
2344 // Micro-benchmark spamming. Policy is to always remove empty loops.
2345 // The 'DO' part is to replace the trip counter with the value it will
2346 // have on the last iteration. This will break the loop.
2347 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2348 // Minimum size must be empty loop
2349 if (_body.size() > EMPTY_LOOP_SIZE)
2350 return false;
2351
2352 if (!_head->is_CountedLoop())
2353 return false; // Dead loop
2354 CountedLoopNode *cl = _head->as_CountedLoop();
2355 if (!cl->is_valid_counted_loop())
2356 return false; // Malformed loop
2357 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2358 return false; // Infinite loop
2359
2360 if (cl->is_pre_loop()) {
2361 // If the loop we are removing is a pre-loop then the main and
2362 // post loop can be removed as well
2363 remove_main_post_loops(cl, phase);
2364 }
2365
2366 #ifdef ASSERT
2367 // Ensure only one phi which is the iv.
2368 Node* iv = NULL;
2369 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2370 Node* n = cl->fast_out(i);
2371 if (n->Opcode() == Op_Phi) {
2372 assert(iv == NULL, "Too many phis" );
2373 iv = n;
2374 }
2375 }
2376 assert(iv == cl->phi(), "Wrong phi" );
2377 #endif
2378
2379 // main and post loops have explicitly created zero trip guard
2380 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2381 if (needs_guard) {
2382 // Skip guard if values not overlap.
2383 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2384 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2385 int stride_con = cl->stride_con();
2386 if (stride_con > 0) {
2387 needs_guard = (init_t->_hi >= limit_t->_lo);
2388 } else {
2389 needs_guard = (init_t->_lo <= limit_t->_hi);
2390 }
2391 }
2392 if (needs_guard) {
2393 // Check for an obvious zero trip guard.
2394 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2395 if (inctrl->Opcode() == Op_IfTrue) {
2396 // The test should look like just the backedge of a CountedLoop
2397 Node* iff = inctrl->in(0);
2398 if (iff->is_If()) {
2399 Node* bol = iff->in(1);
2400 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
2401 Node* cmp = bol->in(1);
2402 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
2403 needs_guard = false;
2404 }
2405 }
2406 }
2407 }
2408 }
2409
2410 #ifndef PRODUCT
2411 if (PrintOpto) {
2412 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2413 this->dump_head();
2414 } else if (TraceLoopOpts) {
2415 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2416 this->dump_head();
2417 }
2418 #endif
2419
2420 if (needs_guard) {
2421 // Peel the loop to ensure there's a zero trip guard
2422 Node_List old_new;
2423 phase->do_peeling(this, old_new);
2424 }
2425
2426 // Replace the phi at loop head with the final value of the last
2427 // iteration. Then the CountedLoopEnd will collapse (backedge never
2428 // taken) and all loop-invariant uses of the exit values will be correct.
2429 Node *phi = cl->phi();
2430 Node *exact_limit = phase->exact_limit(this);
2431 if (exact_limit != cl->limit()) {
2432 // We also need to replace the original limit to collapse loop exit.
2433 Node* cmp = cl->loopexit()->cmp_node();
2434 assert(cl->limit() == cmp->in(2), "sanity");
2435 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2436 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2437 }
2438 // Note: the final value after increment should not overflow since
2439 // counted loop has limit check predicate.
2440 Node *final = new SubINode( exact_limit, cl->stride() );
2441 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2442 phase->_igvn.replace_node(phi,final);
2443 phase->C->set_major_progress();
2444 return true;
2445 }
2446
2447 //------------------------------policy_do_one_iteration_loop-------------------
2448 // Convert one iteration loop into normal code.
2449 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2450 if (!_head->as_Loop()->is_valid_counted_loop())
2451 return false; // Only for counted loop
2452
2453 CountedLoopNode *cl = _head->as_CountedLoop();
2454 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2455 return false;
2456 }
2457
2458 #ifndef PRODUCT
2459 if(TraceLoopOpts) {
2460 tty->print("OneIteration ");
2461 this->dump_head();
2462 }
2463 #endif
2464
2465 Node *init_n = cl->init_trip();
2466 #ifdef ASSERT
2467 // Loop boundaries should be constant since trip count is exact.
2468 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2469 #endif
2470 // Replace the phi at loop head with the value of the init_trip.
2471 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2472 // and all loop-invariant uses of the exit values will be correct.
2473 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2474 phase->C->set_major_progress();
2475 return true;
2476 }
2477
2478 //=============================================================================
2479 //------------------------------iteration_split_impl---------------------------
2480 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2481 // Compute exact loop trip count if possible.
2482 compute_exact_trip_count(phase);
2483
2484 // Convert one iteration loop into normal code.
2485 if (policy_do_one_iteration_loop(phase))
2486 return true;
2487
2488 // Check and remove empty loops (spam micro-benchmarks)
2489 if (policy_do_remove_empty_loop(phase))
2490 return true; // Here we removed an empty loop
2491
2492 bool should_peel = policy_peeling(phase); // Should we peel?
2493
2494 bool should_unswitch = policy_unswitching(phase);
2495
2496 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2497 // This removes loop-invariant tests (usually null checks).
2498 if (!_head->is_CountedLoop()) { // Non-counted loop
2499 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2500 // Partial peel succeeded so terminate this round of loop opts
2501 return false;
2502 }
2503 if (should_peel) { // Should we peel?
2504 #ifndef PRODUCT
2505 if (PrintOpto) tty->print_cr("should_peel");
2506 #endif
2507 phase->do_peeling(this,old_new);
2508 } else if (should_unswitch) {
2509 phase->do_unswitching(this, old_new);
2510 }
2511 return true;
2512 }
2513 CountedLoopNode *cl = _head->as_CountedLoop();
2514
2515 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2516
2517 // Do nothing special to pre- and post- loops
2518 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2519
2520 // Compute loop trip count from profile data
2521 compute_profile_trip_cnt(phase);
2522
2523 // Before attempting fancy unrolling, RCE or alignment, see if we want
2524 // to completely unroll this loop or do loop unswitching.
2525 if (cl->is_normal_loop()) {
2526 if (should_unswitch) {
2527 phase->do_unswitching(this, old_new);
2528 return true;
2529 }
2530 bool should_maximally_unroll = policy_maximally_unroll(phase);
2531 if (should_maximally_unroll) {
2532 // Here we did some unrolling and peeling. Eventually we will
2533 // completely unroll this loop and it will no longer be a loop.
2534 phase->do_maximally_unroll(this,old_new);
2535 return true;
2536 }
2537 }
2538
2539 // Skip next optimizations if running low on nodes. Note that
2540 // policy_unswitching and policy_maximally_unroll have this check.
2541 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes();
2542 if ((int)(2 * _body.size()) > nodes_left) {
2543 return true;
2544 }
2545
2546 // Counted loops may be peeled, may need some iterations run up
2547 // front for RCE, and may want to align loop refs to a cache
2548 // line. Thus we clone a full loop up front whose trip count is
2549 // at least 1 (if peeling), but may be several more.
2550
2551 // The main loop will start cache-line aligned with at least 1
2552 // iteration of the unrolled body (zero-trip test required) and
2553 // will have some range checks removed.
2554
2555 // A post-loop will finish any odd iterations (leftover after
2556 // unrolling), plus any needed for RCE purposes.
2557
2558 bool should_unroll = policy_unroll(phase);
2559
2560 bool should_rce = policy_range_check(phase);
2561
2562 bool should_align = policy_align(phase);
2563
2564 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2565 // need a pre-loop. We may still need to peel an initial iteration but
2566 // we will not be needing an unknown number of pre-iterations.
2567 //
2568 // Basically, if may_rce_align reports FALSE first time through,
2569 // we will not be able to later do RCE or Aligning on this loop.
2570 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2571
2572 // If we have any of these conditions (RCE, alignment, unrolling) met, then
2573 // we switch to the pre-/main-/post-loop model. This model also covers
2574 // peeling.
2575 if (should_rce || should_align || should_unroll) {
2576 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
2577 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2578
2579 // Adjust the pre- and main-loop limits to let the pre and post loops run
2580 // with full checks, but the main-loop with no checks. Remove said
2581 // checks from the main body.
2582 if (should_rce)
2583 phase->do_range_check(this,old_new);
2584
2585 // Double loop body for unrolling. Adjust the minimum-trip test (will do
2586 // twice as many iterations as before) and the main body limit (only do
2587 // an even number of trips). If we are peeling, we might enable some RCE
2588 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2589 // peeling.
2590 if (should_unroll && !should_peel) {
2591 phase->do_unroll(this, old_new, true);
2592 }
2593
2594 // Adjust the pre-loop limits to align the main body
2595 // iterations.
2596 if (should_align)
2597 Unimplemented();
2598
2599 } else { // Else we have an unchanged counted loop
2600 if (should_peel) // Might want to peel but do nothing else
2601 phase->do_peeling(this,old_new);
2602 }
2603 return true;
2604 }
2605
2606
2607 //=============================================================================
2608 //------------------------------iteration_split--------------------------------
2609 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2610 // Recursively iteration split nested loops
2611 if (_child && !_child->iteration_split(phase, old_new))
2612 return false;
2613
2614 // Clean out prior deadwood
2615 DCE_loop_body();
2616
2617
2618 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2619 // Replace with a 1-in-10 exit guess.
2620 if (_parent /*not the root loop*/ &&
2621 !_irreducible &&
2622 // Also ignore the occasional dead backedge
2623 !tail()->is_top()) {
2624 adjust_loop_exit_prob(phase);
2625 }
2626
2627 // Gate unrolling, RCE and peeling efforts.
2628 if (!_child && // If not an inner loop, do not split
2629 !_irreducible &&
2630 _allow_optimizations &&
2631 !tail()->is_top()) { // Also ignore the occasional dead backedge
2632 if (!_has_call) {
2633 if (!iteration_split_impl(phase, old_new)) {
2634 return false;
2635 }
2636 } else if (policy_unswitching(phase)) {
2637 phase->do_unswitching(this, old_new);
2638 }
2639 }
2640
2641 // Minor offset re-organization to remove loop-fallout uses of
2642 // trip counter when there was no major reshaping.
2643 phase->reorg_offsets(this);
2644
2645 if (_next && !_next->iteration_split(phase, old_new))
2646 return false;
2647 return true;
2648 }
2649
2650
2651 //=============================================================================
2652 // Process all the loops in the loop tree and replace any fill
2653 // patterns with an intrisc version.
2654 bool PhaseIdealLoop::do_intrinsify_fill() {
2655 bool changed = false;
2656 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2657 IdealLoopTree* lpt = iter.current();
2658 changed |= intrinsify_fill(lpt);
2659 }
2660 return changed;
2661 }
2662
2663
2664 // Examine an inner loop looking for a a single store of an invariant
2665 // value in a unit stride loop,
2666 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2667 Node*& shift, Node*& con) {
2668 const char* msg = NULL;
2669 Node* msg_node = NULL;
2670
2671 store_value = NULL;
2672 con = NULL;
2673 shift = NULL;
2674
2675 // Process the loop looking for stores. If there are multiple
2676 // stores or extra control flow give at this point.
2677 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2678 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2679 Node* n = lpt->_body.at(i);
2680 if (n->outcnt() == 0) continue; // Ignore dead
2681 if (n->is_Store()) {
2682 if (store != NULL) {
2683 msg = "multiple stores";
2684 break;
2685 }
2686 int opc = n->Opcode();
2687 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
2688 msg = "oop fills not handled";
2689 break;
2690 }
2691 Node* value = n->in(MemNode::ValueIn);
2692 if (!lpt->is_invariant(value)) {
2693 msg = "variant store value";
2694 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2695 msg = "not array address";
2696 }
2697 store = n;
2698 store_value = value;
2699 } else if (n->is_If() && n != head->loopexit()) {
2700 msg = "extra control flow";
2701 msg_node = n;
2702 }
2703 }
2704
2705 if (store == NULL) {
2706 // No store in loop
2707 return false;
2708 }
2709
2710 if (msg == NULL && head->stride_con() != 1) {
2711 // could handle negative strides too
2712 if (head->stride_con() < 0) {
2713 msg = "negative stride";
2714 } else {
2715 msg = "non-unit stride";
2716 }
2717 }
2718
2719 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2720 msg = "can't handle store address";
2721 msg_node = store->in(MemNode::Address);
2722 }
2723
2724 if (msg == NULL &&
2725 (!store->in(MemNode::Memory)->is_Phi() ||
2726 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2727 msg = "store memory isn't proper phi";
2728 msg_node = store->in(MemNode::Memory);
2729 }
2730
2731 // Make sure there is an appropriate fill routine
2732 BasicType t = store->as_Mem()->memory_type();
2733 const char* fill_name;
2734 if (msg == NULL &&
2735 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2736 msg = "unsupported store";
2737 msg_node = store;
2738 }
2739
2740 if (msg != NULL) {
2741 #ifndef PRODUCT
2742 if (TraceOptimizeFill) {
2743 tty->print_cr("not fill intrinsic candidate: %s", msg);
2744 if (msg_node != NULL) msg_node->dump();
2745 }
2746 #endif
2747 return false;
2748 }
2749
2750 // Make sure the address expression can be handled. It should be
2751 // head->phi * elsize + con. head->phi might have a ConvI2L.
2752 Node* elements[4];
2753 Node* conv = NULL;
2754 bool found_index = false;
2755 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2756 for (int e = 0; e < count; e++) {
2757 Node* n = elements[e];
2758 if (n->is_Con() && con == NULL) {
2759 con = n;
2760 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2761 Node* value = n->in(1);
2762 #ifdef _LP64
2763 if (value->Opcode() == Op_ConvI2L) {
2764 conv = value;
2765 value = value->in(1);
2766 }
2767 #endif
2768 if (value != head->phi()) {
2769 msg = "unhandled shift in address";
2770 } else {
2771 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2772 msg = "scale doesn't match";
2773 } else {
2774 found_index = true;
2775 shift = n;
2776 }
2777 }
2778 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2779 if (n->in(1) == head->phi()) {
2780 found_index = true;
2781 conv = n;
2782 } else {
2783 msg = "unhandled input to ConvI2L";
2784 }
2785 } else if (n == head->phi()) {
2786 // no shift, check below for allowed cases
2787 found_index = true;
2788 } else {
2789 msg = "unhandled node in address";
2790 msg_node = n;
2791 }
2792 }
2793
2794 if (count == -1) {
2795 msg = "malformed address expression";
2796 msg_node = store;
2797 }
2798
2799 if (!found_index) {
2800 msg = "missing use of index";
2801 }
2802
2803 // byte sized items won't have a shift
2804 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2805 msg = "can't find shift";
2806 msg_node = store;
2807 }
2808
2809 if (msg != NULL) {
2810 #ifndef PRODUCT
2811 if (TraceOptimizeFill) {
2812 tty->print_cr("not fill intrinsic: %s", msg);
2813 if (msg_node != NULL) msg_node->dump();
2814 }
2815 #endif
2816 return false;
2817 }
2818
2819 // No make sure all the other nodes in the loop can be handled
2820 VectorSet ok(Thread::current()->resource_area());
2821
2822 // store related values are ok
2823 ok.set(store->_idx);
2824 ok.set(store->in(MemNode::Memory)->_idx);
2825
2826 CountedLoopEndNode* loop_exit = head->loopexit();
2827 guarantee(loop_exit != NULL, "no loop exit node");
2828
2829 // Loop structure is ok
2830 ok.set(head->_idx);
2831 ok.set(loop_exit->_idx);
2832 ok.set(head->phi()->_idx);
2833 ok.set(head->incr()->_idx);
2834 ok.set(loop_exit->cmp_node()->_idx);
2835 ok.set(loop_exit->in(1)->_idx);
2836
2837 // Address elements are ok
2838 if (con) ok.set(con->_idx);
2839 if (shift) ok.set(shift->_idx);
2840 if (conv) ok.set(conv->_idx);
2841
2842 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2843 Node* n = lpt->_body.at(i);
2844 if (n->outcnt() == 0) continue; // Ignore dead
2845 if (ok.test(n->_idx)) continue;
2846 // Backedge projection is ok
2847 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
2848 if (!n->is_AddP()) {
2849 msg = "unhandled node";
2850 msg_node = n;
2851 break;
2852 }
2853 }
2854
2855 // Make sure no unexpected values are used outside the loop
2856 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2857 Node* n = lpt->_body.at(i);
2858 // These values can be replaced with other nodes if they are used
2859 // outside the loop.
2860 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2861 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2862 Node* use = iter.get();
2863 if (!lpt->_body.contains(use)) {
2864 msg = "node is used outside loop";
2865 // lpt->_body.dump();
2866 msg_node = n;
2867 break;
2868 }
2869 }
2870 }
2871
2872 #ifdef ASSERT
2873 if (TraceOptimizeFill) {
2874 if (msg != NULL) {
2875 tty->print_cr("no fill intrinsic: %s", msg);
2876 if (msg_node != NULL) msg_node->dump();
2877 } else {
2878 tty->print_cr("fill intrinsic for:");
2879 }
2880 store->dump();
2881 if (Verbose) {
2882 lpt->_body.dump();
2883 }
2884 }
2885 #endif
2886
2887 return msg == NULL;
2888 }
2889
2890
2891
2892 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2893 // Only for counted inner loops
2894 if (!lpt->is_counted() || !lpt->is_inner()) {
2895 return false;
2896 }
2897
2898 // Must have constant stride
2899 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2900 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
2901 return false;
2902 }
2903
2904 // Check that the body only contains a store of a loop invariant
2905 // value that is indexed by the loop phi.
2906 Node* store = NULL;
2907 Node* store_value = NULL;
2908 Node* shift = NULL;
2909 Node* offset = NULL;
2910 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2911 return false;
2912 }
2913
2914 #ifndef PRODUCT
2915 if (TraceLoopOpts) {
2916 tty->print("ArrayFill ");
2917 lpt->dump_head();
2918 }
2919 #endif
2920
2921 // Now replace the whole loop body by a call to a fill routine that
2922 // covers the same region as the loop.
2923 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2924
2925 // Build an expression for the beginning of the copy region
2926 Node* index = head->init_trip();
2927 #ifdef _LP64
2928 index = new ConvI2LNode(index);
2929 _igvn.register_new_node_with_optimizer(index);
2930 #endif
2931 if (shift != NULL) {
2932 // byte arrays don't require a shift but others do.
2933 index = new LShiftXNode(index, shift->in(2));
2934 _igvn.register_new_node_with_optimizer(index);
2935 }
2936 index = new AddPNode(base, base, index);
2937 _igvn.register_new_node_with_optimizer(index);
2938 Node* from = new AddPNode(base, index, offset);
2939 _igvn.register_new_node_with_optimizer(from);
2940 // Compute the number of elements to copy
2941 Node* len = new SubINode(head->limit(), head->init_trip());
2942 _igvn.register_new_node_with_optimizer(len);
2943
2944 BasicType t = store->as_Mem()->memory_type();
2945 bool aligned = false;
2946 if (offset != NULL && head->init_trip()->is_Con()) {
2947 int element_size = type2aelembytes(t);
2948 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2949 }
2950
2951 // Build a call to the fill routine
2952 const char* fill_name;
2953 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2954 assert(fill != NULL, "what?");
2955
2956 // Convert float/double to int/long for fill routines
2957 if (t == T_FLOAT) {
2958 store_value = new MoveF2INode(store_value);
2959 _igvn.register_new_node_with_optimizer(store_value);
2960 } else if (t == T_DOUBLE) {
2961 store_value = new MoveD2LNode(store_value);
2962 _igvn.register_new_node_with_optimizer(store_value);
2963 }
2964
2965 Node* mem_phi = store->in(MemNode::Memory);
2966 Node* result_ctrl;
2967 Node* result_mem;
2968 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2969 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill,
2970 fill_name, TypeAryPtr::get_array_body_type(t));
2971 uint cnt = 0;
2972 call->init_req(TypeFunc::Parms + cnt++, from);
2973 call->init_req(TypeFunc::Parms + cnt++, store_value);
2974 #ifdef _LP64
2975 len = new ConvI2LNode(len);
2976 _igvn.register_new_node_with_optimizer(len);
2977 #endif
2978 call->init_req(TypeFunc::Parms + cnt++, len);
2979 #ifdef _LP64
2980 call->init_req(TypeFunc::Parms + cnt++, C->top());
2981 #endif
2982 call->init_req(TypeFunc::Control, head->init_control());
2983 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O.
2984 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl));
2985 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr));
2986 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
2987 _igvn.register_new_node_with_optimizer(call);
2988 result_ctrl = new ProjNode(call,TypeFunc::Control);
2989 _igvn.register_new_node_with_optimizer(result_ctrl);
2990 result_mem = new ProjNode(call,TypeFunc::Memory);
2991 _igvn.register_new_node_with_optimizer(result_mem);
2992
2993 /* Disable following optimization until proper fix (add missing checks).
2994
2995 // If this fill is tightly coupled to an allocation and overwrites
2996 // the whole body, allow it to take over the zeroing.
2997 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2998 if (alloc != NULL && alloc->is_AllocateArray()) {
2999 Node* length = alloc->as_AllocateArray()->Ideal_length();
3000 if (head->limit() == length &&
3001 head->init_trip() == _igvn.intcon(0)) {
3002 if (TraceOptimizeFill) {
3003 tty->print_cr("Eliminated zeroing in allocation");
3004 }
3005 alloc->maybe_set_complete(&_igvn);
3006 } else {
3007 #ifdef ASSERT
3008 if (TraceOptimizeFill) {
3009 tty->print_cr("filling array but bounds don't match");
3010 alloc->dump();
3011 head->init_trip()->dump();
3012 head->limit()->dump();
3013 length->dump();
3014 }
3015 #endif
3016 }
3017 }
3018 */
3019
3020 // Redirect the old control and memory edges that are outside the loop.
3021 Node* exit = head->loopexit()->proj_out(0);
3022 // Sometimes the memory phi of the head is used as the outgoing
3023 // state of the loop. It's safe in this case to replace it with the
3024 // result_mem.
3025 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
3026 _igvn.replace_node(exit, result_ctrl);
3027 _igvn.replace_node(store, result_mem);
3028 // Any uses the increment outside of the loop become the loop limit.
3029 _igvn.replace_node(head->incr(), head->limit());
3030
3031 // Disconnect the head from the loop.
3032 for (uint i = 0; i < lpt->_body.size(); i++) {
3033 Node* n = lpt->_body.at(i);
3034 _igvn.replace_node(n, C->top());
3035 }
3036
3037 return true;
3038 }