1 /*
2 * Copyright (c) 2000, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/convertnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/opaquenode.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/superword.hpp"
42 #include "opto/vectornode.hpp"
43
44 //------------------------------is_loop_exit-----------------------------------
45 // Given an IfNode, return the loop-exiting projection or NULL if both
46 // arms remain in the loop.
47 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
48 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
49 PhaseIdealLoop *phase = _phase;
50 // Test is an IfNode, has 2 projections. If BOTH are in the loop
51 // we need loop unswitching instead of peeling.
52 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
53 return iff->raw_out(0);
54 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
55 return iff->raw_out(1);
56 return NULL;
57 }
58
59
60 //=============================================================================
61
62
63 //------------------------------record_for_igvn----------------------------
64 // Put loop body on igvn work list
65 void IdealLoopTree::record_for_igvn() {
66 for( uint i = 0; i < _body.size(); i++ ) {
67 Node *n = _body.at(i);
68 _phase->_igvn._worklist.push(n);
69 }
70 }
71
72 //------------------------------compute_exact_trip_count-----------------------
73 // Compute loop 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_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 && limit_n != NULL) {
98 // Use longs to avoid integer overflow.
99 int stride_con = cl->stride_con();
100 jlong init_con = phase->_igvn.type(init_n)->is_int()->_lo;
101 jlong limit_con = phase->_igvn.type(limit_n)->is_int()->_hi;
102 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
103 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
104 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
105 if (init_n->is_Con() && limit_n->is_Con()) {
106 // Set exact trip count.
107 cl->set_exact_trip_count((uint)trip_count);
108 } else if (cl->unrolled_count() == 1) {
109 // Set maximum trip count before unrolling.
110 cl->set_trip_count((uint)trip_count);
111 }
112 }
113 }
114 }
115
116 //------------------------------compute_profile_trip_cnt----------------------------
117 // Compute loop trip count from profile data as
118 // (backedge_count + loop_exit_count) / loop_exit_count
119 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
120 if (!_head->is_CountedLoop()) {
121 return;
122 }
123 CountedLoopNode* head = _head->as_CountedLoop();
124 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
125 return; // Already computed
126 }
127 float trip_cnt = (float)max_jint; // default is big
128
129 Node* back = head->in(LoopNode::LoopBackControl);
130 while (back != head) {
131 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
132 back->in(0) &&
133 back->in(0)->is_If() &&
134 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
135 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
136 break;
137 }
138 back = phase->idom(back);
139 }
140 if (back != head) {
141 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
142 back->in(0), "if-projection exists");
143 IfNode* back_if = back->in(0)->as_If();
144 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
145
146 // Now compute a loop exit count
147 float loop_exit_cnt = 0.0f;
148 for( uint i = 0; i < _body.size(); i++ ) {
149 Node *n = _body[i];
150 if( n->is_If() ) {
151 IfNode *iff = n->as_If();
152 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
153 Node *exit = is_loop_exit(iff);
154 if( exit ) {
155 float exit_prob = iff->_prob;
156 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
157 if (exit_prob > PROB_MIN) {
158 float exit_cnt = iff->_fcnt * exit_prob;
159 loop_exit_cnt += exit_cnt;
160 }
161 }
162 }
163 }
164 }
165 if (loop_exit_cnt > 0.0f) {
166 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
167 } else {
168 // No exit count so use
169 trip_cnt = loop_back_cnt;
170 }
171 }
172 #ifndef PRODUCT
173 if (TraceProfileTripCount) {
174 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
175 }
176 #endif
177 head->set_profile_trip_cnt(trip_cnt);
178 }
179
180 //---------------------is_invariant_addition-----------------------------
181 // Return nonzero index of invariant operand for an Add or Sub
182 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
183 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
184 int op = n->Opcode();
185 if (op == Op_AddI || op == Op_SubI) {
186 bool in1_invar = this->is_invariant(n->in(1));
187 bool in2_invar = this->is_invariant(n->in(2));
188 if (in1_invar && !in2_invar) return 1;
189 if (!in1_invar && in2_invar) return 2;
190 }
191 return 0;
192 }
193
194 //---------------------reassociate_add_sub-----------------------------
195 // Reassociate invariant add and subtract expressions:
196 //
197 // inv1 + (x + inv2) => ( inv1 + inv2) + x
198 // (x + inv2) + inv1 => ( inv1 + inv2) + x
199 // inv1 + (x - inv2) => ( inv1 - inv2) + x
200 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
201 // (x + inv2) - inv1 => (-inv1 + inv2) + x
202 // (x - inv2) + inv1 => ( inv1 - inv2) + x
203 // (x - inv2) - inv1 => (-inv1 - inv2) + x
204 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
205 // inv1 - (x - inv2) => ( inv1 + inv2) - x
206 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
207 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
208 // inv1 - (x + inv2) => ( inv1 - inv2) - x
209 //
210 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
211 if ((!n1->is_Add() && !n1->is_Sub()) || n1->outcnt() == 0) return NULL;
212 if (is_invariant(n1)) return NULL;
213 int inv1_idx = is_invariant_addition(n1, phase);
214 if (!inv1_idx) return NULL;
215 // Don't mess with add of constant (igvn moves them to expression tree root.)
216 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
217 Node* inv1 = n1->in(inv1_idx);
218 Node* n2 = n1->in(3 - inv1_idx);
219 int inv2_idx = is_invariant_addition(n2, phase);
220 if (!inv2_idx) return NULL;
221 Node* x = n2->in(3 - inv2_idx);
222 Node* inv2 = n2->in(inv2_idx);
223
224 bool neg_x = n2->is_Sub() && inv2_idx == 1;
225 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
226 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
227 if (n1->is_Sub() && inv1_idx == 1) {
228 neg_x = !neg_x;
229 neg_inv2 = !neg_inv2;
230 }
231 Node* inv1_c = phase->get_ctrl(inv1);
232 Node* inv2_c = phase->get_ctrl(inv2);
233 Node* n_inv1;
234 if (neg_inv1) {
235 Node *zero = phase->_igvn.intcon(0);
236 phase->set_ctrl(zero, phase->C->root());
237 n_inv1 = new SubINode(zero, inv1);
238 phase->register_new_node(n_inv1, inv1_c);
239 } else {
240 n_inv1 = inv1;
241 }
242 Node* inv;
243 if (neg_inv2) {
244 inv = new SubINode(n_inv1, inv2);
245 } else {
246 inv = new AddINode(n_inv1, inv2);
247 }
248 phase->register_new_node(inv, phase->get_early_ctrl(inv));
249
250 Node* addx;
251 if (neg_x) {
252 addx = new SubINode(inv, x);
253 } else {
254 addx = new AddINode(x, inv);
255 }
256 phase->register_new_node(addx, phase->get_ctrl(x));
257 phase->_igvn.replace_node(n1, addx);
258 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
259 _body.yank(n1);
260 return addx;
261 }
262
263 //---------------------reassociate_invariants-----------------------------
264 // Reassociate invariant expressions:
265 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
266 for (int i = _body.size() - 1; i >= 0; i--) {
267 Node *n = _body.at(i);
268 for (int j = 0; j < 5; j++) {
269 Node* nn = reassociate_add_sub(n, phase);
270 if (nn == NULL) break;
271 n = nn; // again
272 };
273 }
274 }
275
276 //------------------------------policy_peeling---------------------------------
277 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
278 // make some loop-invariant test (usually a null-check) happen before the loop.
279 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
280 Node *test = ((IdealLoopTree*)this)->tail();
281 int body_size = ((IdealLoopTree*)this)->_body.size();
282 // Peeling does loop cloning which can result in O(N^2) node construction
283 if( body_size > 255 /* Prevent overflow for large body_size */
284 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) {
285 return false; // too large to safely clone
286 }
287
288 // check for vectorized loops, any peeling done was already applied
289 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false;
290
291 while( test != _head ) { // Scan till run off top of loop
292 if( test->is_If() ) { // Test?
293 Node *ctrl = phase->get_ctrl(test->in(1));
294 if (ctrl->is_top())
295 return false; // Found dead test on live IF? No peeling!
296 // Standard IF only has one input value to check for loop invariance
297 assert(test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd || test->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
298 // Condition is not a member of this loop?
299 if( !is_member(phase->get_loop(ctrl)) &&
300 is_loop_exit(test) )
301 return true; // Found reason to peel!
302 }
303 // Walk up dominators to loop _head looking for test which is
304 // executed on every path thru loop.
305 test = phase->idom(test);
306 }
307 return false;
308 }
309
310 //------------------------------peeled_dom_test_elim---------------------------
311 // If we got the effect of peeling, either by actually peeling or by making
312 // a pre-loop which must execute at least once, we can remove all
313 // loop-invariant dominated tests in the main body.
314 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
315 bool progress = true;
316 while( progress ) {
317 progress = false; // Reset for next iteration
318 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
319 Node *test = prev->in(0);
320 while( test != loop->_head ) { // Scan till run off top of loop
321
322 int p_op = prev->Opcode();
323 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
324 test->is_If() && // Test?
325 !test->in(1)->is_Con() && // And not already obvious?
326 // Condition is not a member of this loop?
327 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
328 // Walk loop body looking for instances of this test
329 for( uint i = 0; i < loop->_body.size(); i++ ) {
330 Node *n = loop->_body.at(i);
331 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
332 // IfNode was dominated by version in peeled loop body
333 progress = true;
334 dominated_by( old_new[prev->_idx], n );
335 }
336 }
337 }
338 prev = test;
339 test = idom(test);
340 } // End of scan tests in loop
341
342 } // End of while( progress )
343 }
344
345 //------------------------------do_peeling-------------------------------------
346 // Peel the first iteration of the given loop.
347 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
348 // The pre-loop illegally has 2 control users (old & new loops).
349 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
350 // Do this by making the old-loop fall-in edges act as if they came
351 // around the loopback from the prior iteration (follow the old-loop
352 // backedges) and then map to the new peeled iteration. This leaves
353 // the pre-loop with only 1 user (the new peeled iteration), but the
354 // peeled-loop backedge has 2 users.
355 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
356 // extra backedge user.
357 //
358 // orig
359 //
360 // stmt1
361 // |
362 // v
363 // loop predicate
364 // |
365 // v
366 // loop<----+
367 // | |
368 // stmt2 |
369 // | |
370 // v |
371 // if ^
372 // / \ |
373 // / \ |
374 // v v |
375 // false true |
376 // / \ |
377 // / ----+
378 // |
379 // v
380 // exit
381 //
382 //
383 // after clone loop
384 //
385 // stmt1
386 // |
387 // v
388 // loop predicate
389 // / \
390 // clone / \ orig
391 // / \
392 // / \
393 // v v
394 // +---->loop clone loop<----+
395 // | | | |
396 // | stmt2 clone stmt2 |
397 // | | | |
398 // | v v |
399 // ^ if clone If ^
400 // | / \ / \ |
401 // | / \ / \ |
402 // | v v v v |
403 // | true false false true |
404 // | / \ / \ |
405 // +---- \ / ----+
406 // \ /
407 // 1v v2
408 // region
409 // |
410 // v
411 // exit
412 //
413 //
414 // after peel and predicate move
415 //
416 // stmt1
417 // /
418 // /
419 // clone / orig
420 // /
421 // / +----------+
422 // / | |
423 // / loop predicate |
424 // / | |
425 // v v |
426 // TOP-->loop clone loop<----+ |
427 // | | | |
428 // stmt2 clone stmt2 | |
429 // | | | ^
430 // v v | |
431 // if clone If ^ |
432 // / \ / \ | |
433 // / \ / \ | |
434 // v v v v | |
435 // true false false true | |
436 // | \ / \ | |
437 // | \ / ----+ ^
438 // | \ / |
439 // | 1v v2 |
440 // v region |
441 // | | |
442 // | v |
443 // | exit |
444 // | |
445 // +--------------->-----------------+
446 //
447 //
448 // final graph
449 //
450 // stmt1
451 // |
452 // v
453 // stmt2 clone
454 // |
455 // v
456 // if clone
457 // / |
458 // / |
459 // v v
460 // false true
461 // | |
462 // | v
463 // | loop predicate
464 // | |
465 // | v
466 // | loop<----+
467 // | | |
468 // | stmt2 |
469 // | | |
470 // | v |
471 // v if ^
472 // | / \ |
473 // | / \ |
474 // | v v |
475 // | false true |
476 // | | \ |
477 // v v --+
478 // region
479 // |
480 // v
481 // exit
482 //
483 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
484
485 C->set_major_progress();
486 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
487 // 'pre' loop from the main and the 'pre' can no longer have its
488 // iterations adjusted. Therefore, we need to declare this loop as
489 // no longer a 'main' loop; it will need new pre and post loops before
490 // we can do further RCE.
491 #ifndef PRODUCT
492 if (TraceLoopOpts) {
493 tty->print("Peel ");
494 loop->dump_head();
495 }
496 #endif
497 Node* head = loop->_head;
498 bool counted_loop = head->is_CountedLoop();
499 if (counted_loop) {
500 CountedLoopNode *cl = head->as_CountedLoop();
501 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
502 cl->set_trip_count(cl->trip_count() - 1);
503 if (cl->is_main_loop()) {
504 cl->set_normal_loop();
505 #ifndef PRODUCT
506 if (PrintOpto && VerifyLoopOptimizations) {
507 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
508 loop->dump_head();
509 }
510 #endif
511 }
512 }
513 Node* entry = head->in(LoopNode::EntryControl);
514
515 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
516 // The pre-loop illegally has 2 control users (old & new loops).
517 clone_loop( loop, old_new, dom_depth(head) );
518
519 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
520 // Do this by making the old-loop fall-in edges act as if they came
521 // around the loopback from the prior iteration (follow the old-loop
522 // backedges) and then map to the new peeled iteration. This leaves
523 // the pre-loop with only 1 user (the new peeled iteration), but the
524 // peeled-loop backedge has 2 users.
525 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
526 _igvn.hash_delete(head);
527 head->set_req(LoopNode::EntryControl, new_entry);
528 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
529 Node* old = head->fast_out(j);
530 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
531 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
532 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
533 // Then loop body backedge value remains the same.
534 new_exit_value = old->in(LoopNode::LoopBackControl);
535 _igvn.hash_delete(old);
536 old->set_req(LoopNode::EntryControl, new_exit_value);
537 }
538 }
539
540
541 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
542 // extra backedge user.
543 Node* new_head = old_new[head->_idx];
544 _igvn.hash_delete(new_head);
545 new_head->set_req(LoopNode::LoopBackControl, C->top());
546 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
547 Node* use = new_head->fast_out(j2);
548 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
549 _igvn.hash_delete(use);
550 use->set_req(LoopNode::LoopBackControl, C->top());
551 }
552 }
553
554
555 // Step 4: Correct dom-depth info. Set to loop-head depth.
556 int dd = dom_depth(head);
557 set_idom(head, head->in(1), dd);
558 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
559 Node *old = loop->_body.at(j3);
560 Node *nnn = old_new[old->_idx];
561 if (!has_ctrl(nnn))
562 set_idom(nnn, idom(nnn), dd-1);
563 }
564
565 // Now force out all loop-invariant dominating tests. The optimizer
566 // finds some, but we _know_ they are all useless.
567 peeled_dom_test_elim(loop,old_new);
568
569 loop->record_for_igvn();
570 }
571
572 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
573
574 //------------------------------policy_maximally_unroll------------------------
575 // Calculate exact loop trip count and return true if loop can be maximally
576 // unrolled.
577 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
578 CountedLoopNode *cl = _head->as_CountedLoop();
579 assert(cl->is_normal_loop(), "");
580 if (!cl->is_valid_counted_loop())
581 return false; // Malformed counted loop
582
583 if (!cl->has_exact_trip_count()) {
584 // Trip count is not exact.
585 return false;
586 }
587
588 uint trip_count = cl->trip_count();
589 // Note, max_juint is used to indicate unknown trip count.
590 assert(trip_count > 1, "one iteration loop should be optimized out already");
591 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
592
593 // Real policy: if we maximally unroll, does it get too big?
594 // Allow the unrolled mess to get larger than standard loop
595 // size. After all, it will no longer be a loop.
596 uint body_size = _body.size();
597 uint unroll_limit = (uint)LoopUnrollLimit * 4;
598 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
599 if (trip_count > unroll_limit || body_size > unroll_limit) {
600 return false;
601 }
602
603 // Fully unroll a loop with few iterations regardless next
604 // conditions since following loop optimizations will split
605 // such loop anyway (pre-main-post).
606 if (trip_count <= 3)
607 return true;
608
609 // Take into account that after unroll conjoined heads and tails will fold,
610 // otherwise policy_unroll() may allow more unrolling than max unrolling.
611 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
612 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
613 if (body_size != tst_body_size) // Check for int overflow
614 return false;
615 if (new_body_size > unroll_limit ||
616 // Unrolling can result in a large amount of node construction
617 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) {
618 return false;
619 }
620
621 // Do not unroll a loop with String intrinsics code.
622 // String intrinsics are large and have loops.
623 for (uint k = 0; k < _body.size(); k++) {
624 Node* n = _body.at(k);
625 switch (n->Opcode()) {
626 case Op_StrComp:
627 case Op_StrEquals:
628 case Op_StrIndexOf:
629 case Op_StrIndexOfChar:
630 case Op_EncodeISOArray:
631 case Op_AryEq:
632 case Op_HasNegatives: {
633 return false;
634 }
635 #if INCLUDE_RTM_OPT
636 case Op_FastLock:
637 case Op_FastUnlock: {
638 // Don't unroll RTM locking code because it is large.
639 if (UseRTMLocking) {
640 return false;
641 }
642 }
643 #endif
644 } // switch
645 }
646
647 return true; // Do maximally unroll
648 }
649
650
651 //------------------------------policy_unroll----------------------------------
652 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
653 // the loop is a CountedLoop and the body is small enough.
654 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
655
656 CountedLoopNode *cl = _head->as_CountedLoop();
657 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
658
659 if (!cl->is_valid_counted_loop())
660 return false; // Malformed counted loop
661
662 // Protect against over-unrolling.
663 // After split at least one iteration will be executed in pre-loop.
664 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
665
666 _local_loop_unroll_limit = LoopUnrollLimit;
667 _local_loop_unroll_factor = 4;
668 int future_unroll_ct = cl->unrolled_count() * 2;
669 if (!cl->is_vectorized_loop()) {
670 if (future_unroll_ct > LoopMaxUnroll) return false;
671 } else {
672 // obey user constraints on vector mapped loops with additional unrolling applied
673 int unroll_constraint = (cl->slp_max_unroll()) ? cl->slp_max_unroll() : 1;
674 if ((future_unroll_ct / unroll_constraint) > LoopMaxUnroll) return false;
675 }
676
677 // Check for initial stride being a small enough constant
678 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
679
680 // Don't unroll if the next round of unrolling would push us
681 // over the expected trip count of the loop. One is subtracted
682 // from the expected trip count because the pre-loop normally
683 // executes 1 iteration.
684 if (UnrollLimitForProfileCheck > 0 &&
685 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
686 future_unroll_ct > UnrollLimitForProfileCheck &&
687 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
688 return false;
689 }
690
691 // When unroll count is greater than LoopUnrollMin, don't unroll if:
692 // the residual iterations are more than 10% of the trip count
693 // and rounds of "unroll,optimize" are not making significant progress
694 // Progress defined as current size less than 20% larger than previous size.
695 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
696 future_unroll_ct > LoopUnrollMin &&
697 (future_unroll_ct - 1) * (100 / LoopPercentProfileLimit) > cl->profile_trip_cnt() &&
698 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
699 return false;
700 }
701
702 Node *init_n = cl->init_trip();
703 Node *limit_n = cl->limit();
704 int stride_con = cl->stride_con();
705 // Non-constant bounds.
706 // Protect against over-unrolling when init or/and limit are not constant
707 // (so that trip_count's init value is maxint) but iv range is known.
708 if (init_n == NULL || !init_n->is_Con() ||
709 limit_n == NULL || !limit_n->is_Con()) {
710 Node* phi = cl->phi();
711 if (phi != NULL) {
712 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
713 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
714 int next_stride = stride_con * 2; // stride after this unroll
715 if (next_stride > 0) {
716 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
717 iv_type->_lo + next_stride > iv_type->_hi) {
718 return false; // over-unrolling
719 }
720 } else if (next_stride < 0) {
721 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
722 iv_type->_hi + next_stride < iv_type->_lo) {
723 return false; // over-unrolling
724 }
725 }
726 }
727 }
728
729 // After unroll limit will be adjusted: new_limit = limit-stride.
730 // Bailout if adjustment overflow.
731 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
732 if ((stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi)) ||
733 (stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)))
734 return false; // overflow
735
736 // Adjust body_size to determine if we unroll or not
737 uint body_size = _body.size();
738 // Key test to unroll loop in CRC32 java code
739 int xors_in_loop = 0;
740 // Also count ModL, DivL and MulL which expand mightly
741 for (uint k = 0; k < _body.size(); k++) {
742 Node* n = _body.at(k);
743 switch (n->Opcode()) {
744 case Op_XorI: xors_in_loop++; break; // CRC32 java code
745 case Op_ModL: body_size += 30; break;
746 case Op_DivL: body_size += 30; break;
747 case Op_MulL: body_size += 10; break;
748 case Op_StrComp:
749 case Op_StrEquals:
750 case Op_StrIndexOf:
751 case Op_StrIndexOfChar:
752 case Op_EncodeISOArray:
753 case Op_AryEq:
754 case Op_HasNegatives: {
755 // Do not unroll a loop with String intrinsics code.
756 // String intrinsics are large and have loops.
757 return false;
758 }
759 #if INCLUDE_RTM_OPT
760 case Op_FastLock:
761 case Op_FastUnlock: {
762 // Don't unroll RTM locking code because it is large.
763 if (UseRTMLocking) {
764 return false;
765 }
766 }
767 #endif
768 } // switch
769 }
770
771 if (UseSuperWord) {
772 if (!cl->is_reduction_loop()) {
773 phase->mark_reductions(this);
774 }
775
776 // Only attempt slp analysis when user controls do not prohibit it
777 if (LoopMaxUnroll > _local_loop_unroll_factor) {
778 // Once policy_slp_analysis succeeds, mark the loop with the
779 // maximal unroll factor so that we minimize analysis passes
780 if (future_unroll_ct >= _local_loop_unroll_factor) {
781 policy_unroll_slp_analysis(cl, phase, future_unroll_ct);
782 }
783 }
784 }
785
786 int slp_max_unroll_factor = cl->slp_max_unroll();
787 if ((LoopMaxUnroll < slp_max_unroll_factor) && FLAG_IS_DEFAULT(LoopMaxUnroll) && UseSubwordForMaxVector) {
788 LoopMaxUnroll = slp_max_unroll_factor;
789 }
790 if (cl->has_passed_slp()) {
791 if (slp_max_unroll_factor >= future_unroll_ct) return true;
792 // Normal case: loop too big
793 return false;
794 }
795
796 // Check for being too big
797 if (body_size > (uint)_local_loop_unroll_limit) {
798 if ((UseSubwordForMaxVector || xors_in_loop >= 4) && body_size < (uint)LoopUnrollLimit * 4) return true;
799 // Normal case: loop too big
800 return false;
801 }
802
803 if (cl->do_unroll_only()) {
804 if (TraceSuperWordLoopUnrollAnalysis) {
805 tty->print_cr("policy_unroll passed vector loop(vlen=%d,factor = %d)\n", slp_max_unroll_factor, future_unroll_ct);
806 }
807 }
808
809 // Unroll once! (Each trip will soon do double iterations)
810 return true;
811 }
812
813 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) {
814 // Enable this functionality target by target as needed
815 if (SuperWordLoopUnrollAnalysis) {
816 if (!cl->was_slp_analyzed()) {
817 SuperWord sw(phase);
818 sw.transform_loop(this, false);
819
820 // If the loop is slp canonical analyze it
821 if (sw.early_return() == false) {
822 sw.unrolling_analysis(_local_loop_unroll_factor);
823 }
824 }
825
826 if (cl->has_passed_slp()) {
827 int slp_max_unroll_factor = cl->slp_max_unroll();
828 if (slp_max_unroll_factor >= future_unroll_ct) {
829 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor;
830 if (new_limit > LoopUnrollLimit) {
831 if (TraceSuperWordLoopUnrollAnalysis) {
832 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit);
833 }
834 _local_loop_unroll_limit = new_limit;
835 }
836 }
837 }
838 }
839 }
840
841 //------------------------------policy_align-----------------------------------
842 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
843 // expression that does the alignment. Note that only one array base can be
844 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
845 // if we vectorize short memory ops into longer memory ops, we may want to
846 // increase alignment.
847 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
848 return false;
849 }
850
851 //------------------------------policy_range_check-----------------------------
852 // Return TRUE or FALSE if the loop should be range-check-eliminated.
853 // Actually we do iteration-splitting, a more powerful form of RCE.
854 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
855 if (!RangeCheckElimination) return false;
856
857 CountedLoopNode *cl = _head->as_CountedLoop();
858 // If we unrolled with no intention of doing RCE and we later
859 // changed our minds, we got no pre-loop. Either we need to
860 // make a new pre-loop, or we gotta disallow RCE.
861 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
862 Node *trip_counter = cl->phi();
863
864 // check for vectorized loops, some opts are no longer needed
865 if (cl->do_unroll_only()) return false;
866
867 // Check loop body for tests of trip-counter plus loop-invariant vs
868 // loop-invariant.
869 for (uint i = 0; i < _body.size(); i++) {
870 Node *iff = _body[i];
871 if (iff->Opcode() == Op_If ||
872 iff->Opcode() == Op_RangeCheck) { // Test?
873
874 // Comparing trip+off vs limit
875 Node *bol = iff->in(1);
876 if (bol->req() != 2) continue; // dead constant test
877 if (!bol->is_Bool()) {
878 assert(bol->Opcode() == Op_Conv2B, "predicate check only");
879 continue;
880 }
881 if (bol->as_Bool()->_test._test == BoolTest::ne)
882 continue; // not RC
883
884 Node *cmp = bol->in(1);
885 Node *rc_exp = cmp->in(1);
886 Node *limit = cmp->in(2);
887
888 Node *limit_c = phase->get_ctrl(limit);
889 if( limit_c == phase->C->top() )
890 return false; // Found dead test on live IF? No RCE!
891 if( is_member(phase->get_loop(limit_c) ) ) {
892 // Compare might have operands swapped; commute them
893 rc_exp = cmp->in(2);
894 limit = cmp->in(1);
895 limit_c = phase->get_ctrl(limit);
896 if( is_member(phase->get_loop(limit_c) ) )
897 continue; // Both inputs are loop varying; cannot RCE
898 }
899
900 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
901 continue;
902 }
903 // Yeah! Found a test like 'trip+off vs limit'
904 // Test is an IfNode, has 2 projections. If BOTH are in the loop
905 // we need loop unswitching instead of iteration splitting.
906 if( is_loop_exit(iff) )
907 return true; // Found reason to split iterations
908 } // End of is IF
909 }
910
911 return false;
912 }
913
914 //------------------------------policy_peel_only-------------------------------
915 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
916 // for unrolling loops with NO array accesses.
917 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
918 // check for vectorized loops, any peeling done was already applied
919 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false;
920
921 for( uint i = 0; i < _body.size(); i++ )
922 if( _body[i]->is_Mem() )
923 return false;
924
925 // No memory accesses at all!
926 return true;
927 }
928
929 //------------------------------clone_up_backedge_goo--------------------------
930 // If Node n lives in the back_ctrl block and cannot float, we clone a private
931 // version of n in preheader_ctrl block and return that, otherwise return n.
932 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
933 if( get_ctrl(n) != back_ctrl ) return n;
934
935 // Only visit once
936 if (visited.test_set(n->_idx)) {
937 Node *x = clones.find(n->_idx);
938 if (x != NULL)
939 return x;
940 return n;
941 }
942
943 Node *x = NULL; // If required, a clone of 'n'
944 // Check for 'n' being pinned in the backedge.
945 if( n->in(0) && n->in(0) == back_ctrl ) {
946 assert(clones.find(n->_idx) == NULL, "dead loop");
947 x = n->clone(); // Clone a copy of 'n' to preheader
948 clones.push(x, n->_idx);
949 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
950 }
951
952 // Recursive fixup any other input edges into x.
953 // If there are no changes we can just return 'n', otherwise
954 // we need to clone a private copy and change it.
955 for( uint i = 1; i < n->req(); i++ ) {
956 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
957 if( g != n->in(i) ) {
958 if( !x ) {
959 assert(clones.find(n->_idx) == NULL, "dead loop");
960 x = n->clone();
961 clones.push(x, n->_idx);
962 }
963 x->set_req(i, g);
964 }
965 }
966 if( x ) { // x can legally float to pre-header location
967 register_new_node( x, preheader_ctrl );
968 return x;
969 } else { // raise n to cover LCA of uses
970 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
971 }
972 return n;
973 }
974
975 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) {
976 Node* castii = new CastIINode(incr, TypeInt::INT, true);
977 castii->set_req(0, ctrl);
978 register_new_node(castii, ctrl);
979 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) {
980 Node* n = incr->fast_out(i);
981 if (n->is_Phi() && n->in(0) == loop) {
982 int nrep = n->replace_edge(incr, castii);
983 return true;
984 }
985 }
986 return false;
987 }
988
989 //------------------------------insert_pre_post_loops--------------------------
990 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
991 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
992 // alignment. Useful to unroll loops that do no array accesses.
993 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
994
995 #ifndef PRODUCT
996 if (TraceLoopOpts) {
997 if (peel_only)
998 tty->print("PeelMainPost ");
999 else
1000 tty->print("PreMainPost ");
1001 loop->dump_head();
1002 }
1003 #endif
1004 C->set_major_progress();
1005
1006 // Find common pieces of the loop being guarded with pre & post loops
1007 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1008 assert( main_head->is_normal_loop(), "" );
1009 CountedLoopEndNode *main_end = main_head->loopexit();
1010 guarantee(main_end != NULL, "no loop exit node");
1011 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
1012 uint dd_main_head = dom_depth(main_head);
1013 uint max = main_head->outcnt();
1014
1015 Node *pre_header= main_head->in(LoopNode::EntryControl);
1016 Node *init = main_head->init_trip();
1017 Node *incr = main_end ->incr();
1018 Node *limit = main_end ->limit();
1019 Node *stride = main_end ->stride();
1020 Node *cmp = main_end ->cmp_node();
1021 BoolTest::mask b_test = main_end->test_trip();
1022
1023 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
1024 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
1025 if( bol->outcnt() != 1 ) {
1026 bol = bol->clone();
1027 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
1028 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol);
1029 }
1030 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
1031 if( cmp->outcnt() != 1 ) {
1032 cmp = cmp->clone();
1033 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
1034 _igvn.replace_input_of(bol, 1, cmp);
1035 }
1036
1037 // Add the post loop
1038 CountedLoopNode *post_head = NULL;
1039 Node *main_exit = insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1040
1041 //------------------------------
1042 // Step B: Create Pre-Loop.
1043
1044 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1045 // loop pre-header illegally has 2 control users (old & new loops).
1046 clone_loop( loop, old_new, dd_main_head );
1047 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1048 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1049 pre_head->set_pre_loop(main_head);
1050 Node *pre_incr = old_new[incr->_idx];
1051
1052 // Reduce the pre-loop trip count.
1053 pre_end->_prob = PROB_FAIR;
1054
1055 // Find the pre-loop normal exit.
1056 Node* pre_exit = pre_end->proj_out(false);
1057 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1058 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end);
1059 _igvn.register_new_node_with_optimizer( new_pre_exit );
1060 set_idom(new_pre_exit, pre_end, dd_main_head);
1061 set_loop(new_pre_exit, loop->_parent);
1062
1063 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1064 // pre-loop, the main-loop may not execute at all. Later in life this
1065 // zero-trip guard will become the minimum-trip guard when we unroll
1066 // the main-loop.
1067 Node *min_opaq = new Opaque1Node(C, limit);
1068 Node *min_cmp = new CmpINode( pre_incr, min_opaq );
1069 Node *min_bol = new BoolNode( min_cmp, b_test );
1070 register_new_node( min_opaq, new_pre_exit );
1071 register_new_node( min_cmp , new_pre_exit );
1072 register_new_node( min_bol , new_pre_exit );
1073
1074 // Build the IfNode (assume the main-loop is executed always).
1075 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1076 _igvn.register_new_node_with_optimizer( min_iff );
1077 set_idom(min_iff, new_pre_exit, dd_main_head);
1078 set_loop(min_iff, loop->_parent);
1079
1080 // Plug in the false-path, taken if we need to skip main-loop
1081 _igvn.hash_delete( pre_exit );
1082 pre_exit->set_req(0, min_iff);
1083 set_idom(pre_exit, min_iff, dd_main_head);
1084 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1085 // Make the true-path, must enter the main loop
1086 Node *min_taken = new IfTrueNode( min_iff );
1087 _igvn.register_new_node_with_optimizer( min_taken );
1088 set_idom(min_taken, min_iff, dd_main_head);
1089 set_loop(min_taken, loop->_parent);
1090 // Plug in the true path
1091 _igvn.hash_delete( main_head );
1092 main_head->set_req(LoopNode::EntryControl, min_taken);
1093 set_idom(main_head, min_taken, dd_main_head);
1094
1095 Arena *a = Thread::current()->resource_area();
1096 VectorSet visited(a);
1097 Node_Stack clones(a, main_head->back_control()->outcnt());
1098 // Step B3: Make the fall-in values to the main-loop come from the
1099 // fall-out values of the pre-loop.
1100 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1101 Node* main_phi = main_head->fast_out(i2);
1102 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1103 Node *pre_phi = old_new[main_phi->_idx];
1104 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1105 main_head->init_control(),
1106 pre_phi->in(LoopNode::LoopBackControl),
1107 visited, clones);
1108 _igvn.hash_delete(main_phi);
1109 main_phi->set_req( LoopNode::EntryControl, fallpre );
1110 }
1111 }
1112
1113 // Nodes inside the loop may be control dependent on a predicate
1114 // that was moved before the preloop. If the back branch of the main
1115 // or post loops becomes dead, those nodes won't be dependent on the
1116 // test that guards that loop nest anymore which could lead to an
1117 // incorrect array access because it executes independently of the
1118 // test that was guarding the loop nest. We add a special CastII on
1119 // the if branch that enters the loop, between the input induction
1120 // variable value and the induction variable Phi to preserve correct
1121 // dependencies.
1122
1123 // CastII for the main loop:
1124 bool inserted = cast_incr_before_loop( pre_incr, min_taken, main_head );
1125 assert(inserted, "no castII inserted");
1126
1127 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1128 // RCE and alignment may change this later.
1129 Node *cmp_end = pre_end->cmp_node();
1130 assert( cmp_end->in(2) == limit, "" );
1131 Node *pre_limit = new AddINode( init, stride );
1132
1133 // Save the original loop limit in this Opaque1 node for
1134 // use by range check elimination.
1135 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit);
1136
1137 register_new_node( pre_limit, pre_head->in(0) );
1138 register_new_node( pre_opaq , pre_head->in(0) );
1139
1140 // Since no other users of pre-loop compare, I can hack limit directly
1141 assert( cmp_end->outcnt() == 1, "no other users" );
1142 _igvn.hash_delete(cmp_end);
1143 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1144
1145 // Special case for not-equal loop bounds:
1146 // Change pre loop test, main loop test, and the
1147 // main loop guard test to use lt or gt depending on stride
1148 // direction:
1149 // positive stride use <
1150 // negative stride use >
1151 //
1152 // not-equal test is kept for post loop to handle case
1153 // when init > limit when stride > 0 (and reverse).
1154
1155 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1156
1157 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1158 // Modify pre loop end condition
1159 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1160 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test);
1161 register_new_node( new_bol0, pre_head->in(0) );
1162 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0);
1163 // Modify main loop guard condition
1164 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1165 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test);
1166 register_new_node( new_bol1, new_pre_exit );
1167 _igvn.hash_delete(min_iff);
1168 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1169 // Modify main loop end condition
1170 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1171 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test);
1172 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1173 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2);
1174 }
1175
1176 // Flag main loop
1177 main_head->set_main_loop();
1178 if( peel_only ) main_head->set_main_no_pre_loop();
1179
1180 // Subtract a trip count for the pre-loop.
1181 main_head->set_trip_count(main_head->trip_count() - 1);
1182
1183 // It's difficult to be precise about the trip-counts
1184 // for the pre/post loops. They are usually very short,
1185 // so guess that 4 trips is a reasonable value.
1186 post_head->set_profile_trip_cnt(4.0);
1187 pre_head->set_profile_trip_cnt(4.0);
1188
1189 // Now force out all loop-invariant dominating tests. The optimizer
1190 // finds some, but we _know_ they are all useless.
1191 peeled_dom_test_elim(loop,old_new);
1192 loop->record_for_igvn();
1193 }
1194
1195 //------------------------------insert_vector_post_loop------------------------
1196 // Insert a copy of the atomic unrolled vectorized main loop as a post loop,
1197 // unroll_policy has already informed us that more unrolling is about to happen to
1198 // the main loop. The resultant post loop will serve as a vectorized drain loop.
1199 void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1200 if (!loop->_head->is_CountedLoop()) return;
1201
1202 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1203
1204 // only process vectorized main loops
1205 if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return;
1206
1207 int slp_max_unroll_factor = cl->slp_max_unroll();
1208 int cur_unroll = cl->unrolled_count();
1209
1210 if (slp_max_unroll_factor == 0) return;
1211
1212 // only process atomic unroll vector loops (not super unrolled after vectorization)
1213 if (cur_unroll != slp_max_unroll_factor) return;
1214
1215 // we only ever process this one time
1216 if (cl->has_atomic_post_loop()) return;
1217
1218 #ifndef PRODUCT
1219 if (TraceLoopOpts) {
1220 tty->print("PostVector ");
1221 loop->dump_head();
1222 }
1223 #endif
1224 C->set_major_progress();
1225
1226 // Find common pieces of the loop being guarded with pre & post loops
1227 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1228 CountedLoopEndNode *main_end = main_head->loopexit();
1229 guarantee(main_end != NULL, "no loop exit node");
1230 // diagnostic to show loop end is not properly formed
1231 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1232
1233 // mark this loop as processed
1234 main_head->mark_has_atomic_post_loop();
1235
1236 Node *incr = main_end->incr();
1237 Node *limit = main_end->limit();
1238
1239 // In this case we throw away the result as we are not using it to connect anything else.
1240 CountedLoopNode *post_head = NULL;
1241 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1242
1243 // It's difficult to be precise about the trip-counts
1244 // for post loops. They are usually very short,
1245 // so guess that unit vector trips is a reasonable value.
1246 post_head->set_profile_trip_cnt(cur_unroll);
1247
1248 // Now force out all loop-invariant dominating tests. The optimizer
1249 // finds some, but we _know_ they are all useless.
1250 peeled_dom_test_elim(loop, old_new);
1251 loop->record_for_igvn();
1252 }
1253
1254
1255 //-------------------------insert_scalar_rced_post_loop------------------------
1256 // Insert a copy of the rce'd main loop as a post loop,
1257 // We have not unrolled the main loop, so this is the right time to inject this.
1258 // Later we will examine the partner of this post loop pair which still has range checks
1259 // to see inject code which tests at runtime if the range checks are applicable.
1260 void PhaseIdealLoop::insert_scalar_rced_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1261 if (!loop->_head->is_CountedLoop()) return;
1262
1263 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1264
1265 // only process RCE'd main loops
1266 if (!cl->is_main_loop() || cl->range_checks_present()) return;
1267
1268 #ifndef PRODUCT
1269 if (TraceLoopOpts) {
1270 tty->print("PostScalarRce ");
1271 loop->dump_head();
1272 }
1273 #endif
1274 C->set_major_progress();
1275
1276 // Find common pieces of the loop being guarded with pre & post loops
1277 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1278 CountedLoopEndNode *main_end = main_head->loopexit();
1279 guarantee(main_end != NULL, "no loop exit node");
1280 // diagnostic to show loop end is not properly formed
1281 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1282
1283 Node *incr = main_end->incr();
1284 Node *limit = main_end->limit();
1285
1286 // In this case we throw away the result as we are not using it to connect anything else.
1287 CountedLoopNode *post_head = NULL;
1288 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1289
1290 // It's difficult to be precise about the trip-counts
1291 // for post loops. They are usually very short,
1292 // so guess that unit vector trips is a reasonable value.
1293 post_head->set_profile_trip_cnt(4.0);
1294 post_head->set_is_rce_post_loop();
1295
1296 // Now force out all loop-invariant dominating tests. The optimizer
1297 // finds some, but we _know_ they are all useless.
1298 peeled_dom_test_elim(loop, old_new);
1299 loop->record_for_igvn();
1300 }
1301
1302
1303 //------------------------------insert_post_loop-------------------------------
1304 // Insert post loops. Add a post loop to the given loop passed.
1305 Node *PhaseIdealLoop::insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
1306 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
1307 Node *incr, Node *limit, CountedLoopNode *&post_head) {
1308
1309 //------------------------------
1310 // Step A: Create a new post-Loop.
1311 Node* main_exit = main_end->proj_out(false);
1312 assert(main_exit->Opcode() == Op_IfFalse, "");
1313 int dd_main_exit = dom_depth(main_exit);
1314
1315 // Step A1: Clone the loop body of main. The clone becomes the post-loop.
1316 // The main loop pre-header illegally has 2 control users (old & new loops).
1317 clone_loop(loop, old_new, dd_main_exit);
1318 assert(old_new[main_end->_idx]->Opcode() == Op_CountedLoopEnd, "");
1319 post_head = old_new[main_head->_idx]->as_CountedLoop();
1320 post_head->set_normal_loop();
1321 post_head->set_post_loop(main_head);
1322
1323 // Reduce the post-loop trip count.
1324 CountedLoopEndNode* post_end = old_new[main_end->_idx]->as_CountedLoopEnd();
1325 post_end->_prob = PROB_FAIR;
1326
1327 // Build the main-loop normal exit.
1328 IfFalseNode *new_main_exit = new IfFalseNode(main_end);
1329 _igvn.register_new_node_with_optimizer(new_main_exit);
1330 set_idom(new_main_exit, main_end, dd_main_exit);
1331 set_loop(new_main_exit, loop->_parent);
1332
1333 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
1334 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
1335 // (the previous loop trip-counter exit value) because we will be changing
1336 // the exit value (via additional unrolling) so we cannot constant-fold away the zero
1337 // trip guard until all unrolling is done.
1338 Node *zer_opaq = new Opaque1Node(C, incr);
1339 Node *zer_cmp = new CmpINode(zer_opaq, limit);
1340 Node *zer_bol = new BoolNode(zer_cmp, main_end->test_trip());
1341 register_new_node(zer_opaq, new_main_exit);
1342 register_new_node(zer_cmp, new_main_exit);
1343 register_new_node(zer_bol, new_main_exit);
1344
1345 // Build the IfNode
1346 IfNode *zer_iff = new IfNode(new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN);
1347 _igvn.register_new_node_with_optimizer(zer_iff);
1348 set_idom(zer_iff, new_main_exit, dd_main_exit);
1349 set_loop(zer_iff, loop->_parent);
1350
1351 // Plug in the false-path, taken if we need to skip this post-loop
1352 _igvn.replace_input_of(main_exit, 0, zer_iff);
1353 set_idom(main_exit, zer_iff, dd_main_exit);
1354 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
1355 // Make the true-path, must enter this post loop
1356 Node *zer_taken = new IfTrueNode(zer_iff);
1357 _igvn.register_new_node_with_optimizer(zer_taken);
1358 set_idom(zer_taken, zer_iff, dd_main_exit);
1359 set_loop(zer_taken, loop->_parent);
1360 // Plug in the true path
1361 _igvn.hash_delete(post_head);
1362 post_head->set_req(LoopNode::EntryControl, zer_taken);
1363 set_idom(post_head, zer_taken, dd_main_exit);
1364
1365 Arena *a = Thread::current()->resource_area();
1366 VectorSet visited(a);
1367 Node_Stack clones(a, main_head->back_control()->outcnt());
1368 // Step A3: Make the fall-in values to the post-loop come from the
1369 // fall-out values of the main-loop.
1370 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
1371 Node* main_phi = main_head->fast_out(i);
1372 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0) {
1373 Node *cur_phi = old_new[main_phi->_idx];
1374 Node *fallnew = clone_up_backedge_goo(main_head->back_control(),
1375 post_head->init_control(),
1376 main_phi->in(LoopNode::LoopBackControl),
1377 visited, clones);
1378 _igvn.hash_delete(cur_phi);
1379 cur_phi->set_req(LoopNode::EntryControl, fallnew);
1380 }
1381 }
1382
1383 // CastII for the new post loop:
1384 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1385 assert(inserted, "no castII inserted");
1386
1387 return new_main_exit;
1388 }
1389
1390 //------------------------------is_invariant-----------------------------
1391 // Return true if n is invariant
1392 bool IdealLoopTree::is_invariant(Node* n) const {
1393 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1394 if (n_c->is_top()) return false;
1395 return !is_member(_phase->get_loop(n_c));
1396 }
1397
1398
1399 //------------------------------do_unroll--------------------------------------
1400 // Unroll the loop body one step - make each trip do 2 iterations.
1401 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1402 assert(LoopUnrollLimit, "");
1403 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1404 CountedLoopEndNode *loop_end = loop_head->loopexit();
1405 assert(loop_end, "");
1406 #ifndef PRODUCT
1407 if (PrintOpto && VerifyLoopOptimizations) {
1408 tty->print("Unrolling ");
1409 loop->dump_head();
1410 } else if (TraceLoopOpts) {
1411 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1412 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1413 } else {
1414 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1415 }
1416 loop->dump_head();
1417 }
1418
1419 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
1420 Arena* arena = Thread::current()->resource_area();
1421 Node_Stack stack(arena, C->live_nodes() >> 2);
1422 Node_List rpo_list;
1423 VectorSet visited(arena);
1424 visited.set(loop_head->_idx);
1425 rpo( loop_head, stack, visited, rpo_list );
1426 dump(loop, rpo_list.size(), rpo_list );
1427 }
1428 #endif
1429
1430 // Remember loop node count before unrolling to detect
1431 // if rounds of unroll,optimize are making progress
1432 loop_head->set_node_count_before_unroll(loop->_body.size());
1433
1434 Node *ctrl = loop_head->in(LoopNode::EntryControl);
1435 Node *limit = loop_head->limit();
1436 Node *init = loop_head->init_trip();
1437 Node *stride = loop_head->stride();
1438
1439 Node *opaq = NULL;
1440 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1441 // Search for zero-trip guard.
1442
1443 // Check the shape of the graph at the loop entry. If an inappropriate
1444 // graph shape is encountered, the compiler bails out loop unrolling;
1445 // compilation of the method will still succeed.
1446 if (!is_canonical_loop_entry(loop_head)) {
1447 return;
1448 }
1449 opaq = ctrl->in(0)->in(1)->in(1)->in(2);
1450 // Zero-trip test uses an 'opaque' node which is not shared.
1451 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1452 }
1453
1454 C->set_major_progress();
1455
1456 Node* new_limit = NULL;
1457 int stride_con = stride->get_int();
1458 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1459 uint old_trip_count = loop_head->trip_count();
1460 // Verify that unroll policy result is still valid.
1461 assert(old_trip_count > 1 &&
1462 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1463
1464 // Adjust loop limit to keep valid iterations number after unroll.
1465 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1466 // which may overflow.
1467 if (!adjust_min_trip) {
1468 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1469 "odd trip count for maximally unroll");
1470 // Don't need to adjust limit for maximally unroll since trip count is even.
1471 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1472 // Loop's limit is constant. Loop's init could be constant when pre-loop
1473 // become peeled iteration.
1474 jlong init_con = init->get_int();
1475 // We can keep old loop limit if iterations count stays the same:
1476 // old_trip_count == new_trip_count * 2
1477 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1478 // so we also don't need to adjust zero trip test.
1479 jlong limit_con = limit->get_int();
1480 // (stride_con*2) not overflow since stride_con <= 8.
1481 int new_stride_con = stride_con * 2;
1482 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1483 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1484 // New trip count should satisfy next conditions.
1485 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1486 uint new_trip_count = (uint)trip_count;
1487 adjust_min_trip = (old_trip_count != new_trip_count*2);
1488 }
1489
1490 if (adjust_min_trip) {
1491 // Step 2: Adjust the trip limit if it is called for.
1492 // The adjustment amount is -stride. Need to make sure if the
1493 // adjustment underflows or overflows, then the main loop is skipped.
1494 Node* cmp = loop_end->cmp_node();
1495 assert(cmp->in(2) == limit, "sanity");
1496 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1497
1498 // Verify that policy_unroll result is still valid.
1499 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1500 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1501 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1502
1503 if (limit->is_Con()) {
1504 // The check in policy_unroll and the assert above guarantee
1505 // no underflow if limit is constant.
1506 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1507 set_ctrl(new_limit, C->root());
1508 } else {
1509 // Limit is not constant.
1510 if (loop_head->unrolled_count() == 1) { // only for first unroll
1511 // Separate limit by Opaque node in case it is an incremented
1512 // variable from previous loop to avoid using pre-incremented
1513 // value which could increase register pressure.
1514 // Otherwise reorg_offsets() optimization will create a separate
1515 // Opaque node for each use of trip-counter and as result
1516 // zero trip guard limit will be different from loop limit.
1517 assert(has_ctrl(opaq), "should have it");
1518 Node* opaq_ctrl = get_ctrl(opaq);
1519 limit = new Opaque2Node( C, limit );
1520 register_new_node( limit, opaq_ctrl );
1521 }
1522 if ((stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo)) ||
1523 (stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi))) {
1524 // No underflow.
1525 new_limit = new SubINode(limit, stride);
1526 } else {
1527 // (limit - stride) may underflow.
1528 // Clamp the adjustment value with MININT or MAXINT:
1529 //
1530 // new_limit = limit-stride
1531 // if (stride > 0)
1532 // new_limit = (limit < new_limit) ? MININT : new_limit;
1533 // else
1534 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1535 //
1536 BoolTest::mask bt = loop_end->test_trip();
1537 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1538 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1539 set_ctrl(adj_max, C->root());
1540 Node* old_limit = NULL;
1541 Node* adj_limit = NULL;
1542 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1543 if (loop_head->unrolled_count() > 1 &&
1544 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1545 limit->in(CMoveNode::IfTrue) == adj_max &&
1546 bol->as_Bool()->_test._test == bt &&
1547 bol->in(1)->Opcode() == Op_CmpI &&
1548 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1549 // Loop was unrolled before.
1550 // Optimize the limit to avoid nested CMove:
1551 // use original limit as old limit.
1552 old_limit = bol->in(1)->in(1);
1553 // Adjust previous adjusted limit.
1554 adj_limit = limit->in(CMoveNode::IfFalse);
1555 adj_limit = new SubINode(adj_limit, stride);
1556 } else {
1557 old_limit = limit;
1558 adj_limit = new SubINode(limit, stride);
1559 }
1560 assert(old_limit != NULL && adj_limit != NULL, "");
1561 register_new_node( adj_limit, ctrl ); // adjust amount
1562 Node* adj_cmp = new CmpINode(old_limit, adj_limit);
1563 register_new_node( adj_cmp, ctrl );
1564 Node* adj_bool = new BoolNode(adj_cmp, bt);
1565 register_new_node( adj_bool, ctrl );
1566 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1567 }
1568 register_new_node(new_limit, ctrl);
1569 }
1570 assert(new_limit != NULL, "");
1571 // Replace in loop test.
1572 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1573 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1574 // Don't need to create new test since only one user.
1575 _igvn.hash_delete(cmp);
1576 cmp->set_req(2, new_limit);
1577 } else {
1578 // Create new test since it is shared.
1579 Node* ctrl2 = loop_end->in(0);
1580 Node* cmp2 = cmp->clone();
1581 cmp2->set_req(2, new_limit);
1582 register_new_node(cmp2, ctrl2);
1583 Node* bol2 = loop_end->in(1)->clone();
1584 bol2->set_req(1, cmp2);
1585 register_new_node(bol2, ctrl2);
1586 _igvn.replace_input_of(loop_end, 1, bol2);
1587 }
1588 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1589 // Make it a 1-trip test (means at least 2 trips).
1590
1591 // Guard test uses an 'opaque' node which is not shared. Hence I
1592 // can edit it's inputs directly. Hammer in the new limit for the
1593 // minimum-trip guard.
1594 assert(opaq->outcnt() == 1, "");
1595 _igvn.replace_input_of(opaq, 1, new_limit);
1596 }
1597
1598 // Adjust max trip count. The trip count is intentionally rounded
1599 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1600 // the main, unrolled, part of the loop will never execute as it is protected
1601 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1602 // and later determined that part of the unrolled loop was dead.
1603 loop_head->set_trip_count(old_trip_count / 2);
1604
1605 // Double the count of original iterations in the unrolled loop body.
1606 loop_head->double_unrolled_count();
1607
1608 // ---------
1609 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1610 // represents the odd iterations; since the loop trips an even number of
1611 // times its backedge is never taken. Kill the backedge.
1612 uint dd = dom_depth(loop_head);
1613 clone_loop( loop, old_new, dd );
1614
1615 // Make backedges of the clone equal to backedges of the original.
1616 // Make the fall-in from the original come from the fall-out of the clone.
1617 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1618 Node* phi = loop_head->fast_out(j);
1619 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1620 Node *newphi = old_new[phi->_idx];
1621 _igvn.hash_delete( phi );
1622 _igvn.hash_delete( newphi );
1623
1624 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1625 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1626 phi ->set_req(LoopNode::LoopBackControl, C->top());
1627 }
1628 }
1629 Node *clone_head = old_new[loop_head->_idx];
1630 _igvn.hash_delete( clone_head );
1631 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1632 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1633 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1634 loop->_head = clone_head; // New loop header
1635
1636 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1637 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1638
1639 // Kill the clone's backedge
1640 Node *newcle = old_new[loop_end->_idx];
1641 _igvn.hash_delete( newcle );
1642 Node *one = _igvn.intcon(1);
1643 set_ctrl(one, C->root());
1644 newcle->set_req(1, one);
1645 // Force clone into same loop body
1646 uint max = loop->_body.size();
1647 for( uint k = 0; k < max; k++ ) {
1648 Node *old = loop->_body.at(k);
1649 Node *nnn = old_new[old->_idx];
1650 loop->_body.push(nnn);
1651 if (!has_ctrl(old))
1652 set_loop(nnn, loop);
1653 }
1654
1655 loop->record_for_igvn();
1656
1657 #ifndef PRODUCT
1658 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
1659 tty->print("\nnew loop after unroll\n"); loop->dump_head();
1660 for (uint i = 0; i < loop->_body.size(); i++) {
1661 loop->_body.at(i)->dump();
1662 }
1663 if(C->clone_map().is_debug()) {
1664 tty->print("\nCloneMap\n");
1665 Dict* dict = C->clone_map().dict();
1666 DictI i(dict);
1667 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size());
1668 for (int ii = 0; i.test(); ++i, ++ii) {
1669 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key));
1670 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen());
1671 if (ii % 10 == 9) {
1672 tty->print_cr(" ");
1673 }
1674 }
1675 tty->print_cr(" ");
1676 }
1677 }
1678 #endif
1679
1680 }
1681
1682 //------------------------------do_maximally_unroll----------------------------
1683
1684 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1685 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1686 assert(cl->has_exact_trip_count(), "trip count is not exact");
1687 assert(cl->trip_count() > 0, "");
1688 #ifndef PRODUCT
1689 if (TraceLoopOpts) {
1690 tty->print("MaxUnroll %d ", cl->trip_count());
1691 loop->dump_head();
1692 }
1693 #endif
1694
1695 // If loop is tripping an odd number of times, peel odd iteration
1696 if ((cl->trip_count() & 1) == 1) {
1697 do_peeling(loop, old_new);
1698 }
1699
1700 // Now its tripping an even number of times remaining. Double loop body.
1701 // Do not adjust pre-guards; they are not needed and do not exist.
1702 if (cl->trip_count() > 0) {
1703 assert((cl->trip_count() & 1) == 0, "missed peeling");
1704 do_unroll(loop, old_new, false);
1705 }
1706 }
1707
1708 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) {
1709 if (SuperWordReductions == false) return;
1710
1711 CountedLoopNode* loop_head = loop->_head->as_CountedLoop();
1712 if (loop_head->unrolled_count() > 1) {
1713 return;
1714 }
1715
1716 Node* trip_phi = loop_head->phi();
1717 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) {
1718 Node* phi = loop_head->fast_out(i);
1719 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) {
1720 // For definitions which are loop inclusive and not tripcounts.
1721 Node* def_node = phi->in(LoopNode::LoopBackControl);
1722
1723 if (def_node != NULL) {
1724 Node* n_ctrl = get_ctrl(def_node);
1725 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) {
1726 // Now test it to see if it fits the standard pattern for a reduction operator.
1727 int opc = def_node->Opcode();
1728 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) {
1729 if (!def_node->is_reduction()) { // Not marked yet
1730 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it
1731 bool ok = false;
1732 for (unsigned j = 1; j < def_node->req(); j++) {
1733 Node* in = def_node->in(j);
1734 if (in == phi) {
1735 ok = true;
1736 break;
1737 }
1738 }
1739
1740 // do nothing if we did not match the initial criteria
1741 if (ok == false) {
1742 continue;
1743 }
1744
1745 // The result of the reduction must not be used in the loop
1746 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) {
1747 Node* u = def_node->fast_out(i);
1748 if (!loop->is_member(get_loop(ctrl_or_self(u)))) {
1749 continue;
1750 }
1751 if (u == phi) {
1752 continue;
1753 }
1754 ok = false;
1755 }
1756
1757 // iff the uses conform
1758 if (ok) {
1759 def_node->add_flag(Node::Flag_is_reduction);
1760 loop_head->mark_has_reductions();
1761 }
1762 }
1763 }
1764 }
1765 }
1766 }
1767 }
1768 }
1769
1770 //------------------------------adjust_limit-----------------------------------
1771 // Helper function for add_constraint().
1772 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1773 // Compute "I :: (limit-offset)/scale"
1774 Node *con = new SubINode(rc_limit, offset);
1775 register_new_node(con, pre_ctrl);
1776 Node *X = new DivINode(0, con, scale);
1777 register_new_node(X, pre_ctrl);
1778
1779 // Adjust loop limit
1780 loop_limit = (stride_con > 0)
1781 ? (Node*)(new MinINode(loop_limit, X))
1782 : (Node*)(new MaxINode(loop_limit, X));
1783 register_new_node(loop_limit, pre_ctrl);
1784 return loop_limit;
1785 }
1786
1787 //------------------------------add_constraint---------------------------------
1788 // Constrain the main loop iterations so the conditions:
1789 // low_limit <= scale_con * I + offset < upper_limit
1790 // always holds true. That is, either increase the number of iterations in
1791 // the pre-loop or the post-loop until the condition holds true in the main
1792 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1793 // stride and scale are constants (offset and limit often are).
1794 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 ) {
1795 // For positive stride, the pre-loop limit always uses a MAX function
1796 // and the main loop a MIN function. For negative stride these are
1797 // reversed.
1798
1799 // Also for positive stride*scale the affine function is increasing, so the
1800 // pre-loop must check for underflow and the post-loop for overflow.
1801 // Negative stride*scale reverses this; pre-loop checks for overflow and
1802 // post-loop for underflow.
1803
1804 Node *scale = _igvn.intcon(scale_con);
1805 set_ctrl(scale, C->root());
1806
1807 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1808 // The overflow limit: scale*I+offset < upper_limit
1809 // For main-loop compute
1810 // ( if (scale > 0) /* and stride > 0 */
1811 // I < (upper_limit-offset)/scale
1812 // else /* scale < 0 and stride < 0 */
1813 // I > (upper_limit-offset)/scale
1814 // )
1815 //
1816 // (upper_limit-offset) may overflow or underflow.
1817 // But it is fine since main loop will either have
1818 // less iterations or will be skipped in such case.
1819 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1820
1821 // The underflow limit: low_limit <= scale*I+offset.
1822 // For pre-loop compute
1823 // NOT(scale*I+offset >= low_limit)
1824 // scale*I+offset < low_limit
1825 // ( if (scale > 0) /* and stride > 0 */
1826 // I < (low_limit-offset)/scale
1827 // else /* scale < 0 and stride < 0 */
1828 // I > (low_limit-offset)/scale
1829 // )
1830
1831 if (low_limit->get_int() == -max_jint) {
1832 // We need this guard when scale*pre_limit+offset >= limit
1833 // due to underflow. So we need execute pre-loop until
1834 // scale*I+offset >= min_int. But (min_int-offset) will
1835 // underflow when offset > 0 and X will be > original_limit
1836 // when stride > 0. To avoid it we replace positive offset with 0.
1837 //
1838 // Also (min_int+1 == -max_int) is used instead of min_int here
1839 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1840 Node* shift = _igvn.intcon(31);
1841 set_ctrl(shift, C->root());
1842 Node* sign = new RShiftINode(offset, shift);
1843 register_new_node(sign, pre_ctrl);
1844 offset = new AndINode(offset, sign);
1845 register_new_node(offset, pre_ctrl);
1846 } else {
1847 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1848 // The only problem we have here when offset == min_int
1849 // since (0-min_int) == min_int. It may be fine for stride > 0
1850 // but for stride < 0 X will be < original_limit. To avoid it
1851 // max(pre_limit, original_limit) is used in do_range_check().
1852 }
1853 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1854 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1855
1856 } else { // stride_con*scale_con < 0
1857 // For negative stride*scale pre-loop checks for overflow and
1858 // post-loop for underflow.
1859 //
1860 // The overflow limit: scale*I+offset < upper_limit
1861 // For pre-loop compute
1862 // NOT(scale*I+offset < upper_limit)
1863 // scale*I+offset >= upper_limit
1864 // scale*I+offset+1 > upper_limit
1865 // ( if (scale < 0) /* and stride > 0 */
1866 // I < (upper_limit-(offset+1))/scale
1867 // else /* scale > 0 and stride < 0 */
1868 // I > (upper_limit-(offset+1))/scale
1869 // )
1870 //
1871 // (upper_limit-offset-1) may underflow or overflow.
1872 // To avoid it min(pre_limit, original_limit) is used
1873 // in do_range_check() for stride > 0 and max() for < 0.
1874 Node *one = _igvn.intcon(1);
1875 set_ctrl(one, C->root());
1876
1877 Node *plus_one = new AddINode(offset, one);
1878 register_new_node( plus_one, pre_ctrl );
1879 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1880 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1881
1882 if (low_limit->get_int() == -max_jint) {
1883 // We need this guard when scale*main_limit+offset >= limit
1884 // due to underflow. So we need execute main-loop while
1885 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1886 // underflow when (offset+1) > 0 and X will be < main_limit
1887 // when scale < 0 (and stride > 0). To avoid it we replace
1888 // positive (offset+1) with 0.
1889 //
1890 // Also (min_int+1 == -max_int) is used instead of min_int here
1891 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1892 Node* shift = _igvn.intcon(31);
1893 set_ctrl(shift, C->root());
1894 Node* sign = new RShiftINode(plus_one, shift);
1895 register_new_node(sign, pre_ctrl);
1896 plus_one = new AndINode(plus_one, sign);
1897 register_new_node(plus_one, pre_ctrl);
1898 } else {
1899 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1900 // The only problem we have here when offset == max_int
1901 // since (max_int+1) == min_int and (0-min_int) == min_int.
1902 // But it is fine since main loop will either have
1903 // less iterations or will be skipped in such case.
1904 }
1905 // The underflow limit: low_limit <= scale*I+offset.
1906 // For main-loop compute
1907 // scale*I+offset+1 > low_limit
1908 // ( if (scale < 0) /* and stride > 0 */
1909 // I < (low_limit-(offset+1))/scale
1910 // else /* scale > 0 and stride < 0 */
1911 // I > (low_limit-(offset+1))/scale
1912 // )
1913
1914 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1915 }
1916 }
1917
1918
1919 //------------------------------is_scaled_iv---------------------------------
1920 // Return true if exp is a constant times an induction var
1921 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1922 if (exp == iv) {
1923 if (p_scale != NULL) {
1924 *p_scale = 1;
1925 }
1926 return true;
1927 }
1928 int opc = exp->Opcode();
1929 if (opc == Op_MulI) {
1930 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1931 if (p_scale != NULL) {
1932 *p_scale = exp->in(2)->get_int();
1933 }
1934 return true;
1935 }
1936 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1937 if (p_scale != NULL) {
1938 *p_scale = exp->in(1)->get_int();
1939 }
1940 return true;
1941 }
1942 } else if (opc == Op_LShiftI) {
1943 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1944 if (p_scale != NULL) {
1945 *p_scale = 1 << exp->in(2)->get_int();
1946 }
1947 return true;
1948 }
1949 }
1950 return false;
1951 }
1952
1953 //-----------------------------is_scaled_iv_plus_offset------------------------------
1954 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1955 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1956 if (is_scaled_iv(exp, iv, p_scale)) {
1957 if (p_offset != NULL) {
1958 Node *zero = _igvn.intcon(0);
1959 set_ctrl(zero, C->root());
1960 *p_offset = zero;
1961 }
1962 return true;
1963 }
1964 int opc = exp->Opcode();
1965 if (opc == Op_AddI) {
1966 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1967 if (p_offset != NULL) {
1968 *p_offset = exp->in(2);
1969 }
1970 return true;
1971 }
1972 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1973 if (p_offset != NULL) {
1974 *p_offset = exp->in(1);
1975 }
1976 return true;
1977 }
1978 if (exp->in(2)->is_Con()) {
1979 Node* offset2 = NULL;
1980 if (depth < 2 &&
1981 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1982 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1983 if (p_offset != NULL) {
1984 Node *ctrl_off2 = get_ctrl(offset2);
1985 Node* offset = new AddINode(offset2, exp->in(2));
1986 register_new_node(offset, ctrl_off2);
1987 *p_offset = offset;
1988 }
1989 return true;
1990 }
1991 }
1992 } else if (opc == Op_SubI) {
1993 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1994 if (p_offset != NULL) {
1995 Node *zero = _igvn.intcon(0);
1996 set_ctrl(zero, C->root());
1997 Node *ctrl_off = get_ctrl(exp->in(2));
1998 Node* offset = new SubINode(zero, exp->in(2));
1999 register_new_node(offset, ctrl_off);
2000 *p_offset = offset;
2001 }
2002 return true;
2003 }
2004 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
2005 if (p_offset != NULL) {
2006 *p_scale *= -1;
2007 *p_offset = exp->in(1);
2008 }
2009 return true;
2010 }
2011 }
2012 return false;
2013 }
2014
2015 //------------------------------do_range_check---------------------------------
2016 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
2017 int PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
2018 #ifndef PRODUCT
2019 if (PrintOpto && VerifyLoopOptimizations) {
2020 tty->print("Range Check Elimination ");
2021 loop->dump_head();
2022 } else if (TraceLoopOpts) {
2023 tty->print("RangeCheck ");
2024 loop->dump_head();
2025 }
2026 #endif
2027 assert(RangeCheckElimination, "");
2028 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2029 // If we fail before trying to eliminate range checks, set multiversion state
2030 int closed_range_checks = 1;
2031
2032 // protect against stride not being a constant
2033 if (!cl->stride_is_con())
2034 return closed_range_checks;
2035
2036 // Find the trip counter; we are iteration splitting based on it
2037 Node *trip_counter = cl->phi();
2038 // Find the main loop limit; we will trim it's iterations
2039 // to not ever trip end tests
2040 Node *main_limit = cl->limit();
2041
2042 // Check graph shape. Cannot optimize a loop if zero-trip
2043 // Opaque1 node is optimized away and then another round
2044 // of loop opts attempted.
2045 if (!is_canonical_loop_entry(cl)) {
2046 return closed_range_checks;
2047 }
2048
2049 // Need to find the main-loop zero-trip guard
2050 Node *ctrl = cl->in(LoopNode::EntryControl);
2051 Node *iffm = ctrl->in(0);
2052 Node *opqzm = iffm->in(1)->in(1)->in(2);
2053 assert(opqzm->in(1) == main_limit, "do not understand situation");
2054
2055 // Find the pre-loop limit; we will expand its iterations to
2056 // not ever trip low tests.
2057 Node *p_f = iffm->in(0);
2058 // pre loop may have been optimized out
2059 if (p_f->Opcode() != Op_IfFalse) {
2060 return closed_range_checks;
2061 }
2062 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2063 assert(pre_end->loopnode()->is_pre_loop(), "");
2064 Node *pre_opaq1 = pre_end->limit();
2065 // Occasionally it's possible for a pre-loop Opaque1 node to be
2066 // optimized away and then another round of loop opts attempted.
2067 // We can not optimize this particular loop in that case.
2068 if (pre_opaq1->Opcode() != Op_Opaque1)
2069 return closed_range_checks;
2070 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2071 Node *pre_limit = pre_opaq->in(1);
2072
2073 // Where do we put new limit calculations
2074 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2075
2076 // Ensure the original loop limit is available from the
2077 // pre-loop Opaque1 node.
2078 Node *orig_limit = pre_opaq->original_loop_limit();
2079 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
2080 return closed_range_checks;
2081
2082 // Must know if its a count-up or count-down loop
2083
2084 int stride_con = cl->stride_con();
2085 Node *zero = _igvn.intcon(0);
2086 Node *one = _igvn.intcon(1);
2087 // Use symmetrical int range [-max_jint,max_jint]
2088 Node *mini = _igvn.intcon(-max_jint);
2089 set_ctrl(zero, C->root());
2090 set_ctrl(one, C->root());
2091 set_ctrl(mini, C->root());
2092
2093 // Range checks that do not dominate the loop backedge (ie.
2094 // conditionally executed) can lengthen the pre loop limit beyond
2095 // the original loop limit. To prevent this, the pre limit is
2096 // (for stride > 0) MINed with the original loop limit (MAXed
2097 // stride < 0) when some range_check (rc) is conditionally
2098 // executed.
2099 bool conditional_rc = false;
2100
2101 // Count number of range checks and reduce by load range limits, if zero,
2102 // the loop is in canonical form to multiversion.
2103 closed_range_checks = 0;
2104
2105 // Check loop body for tests of trip-counter plus loop-invariant vs loop-variant.
2106 for( uint i = 0; i < loop->_body.size(); i++ ) {
2107 Node *iff = loop->_body[i];
2108 if (iff->Opcode() == Op_If ||
2109 iff->Opcode() == Op_RangeCheck) { // Test?
2110 // Test is an IfNode, has 2 projections. If BOTH are in the loop
2111 // we need loop unswitching instead of iteration splitting.
2112 closed_range_checks++;
2113 Node *exit = loop->is_loop_exit(iff);
2114 if( !exit ) continue;
2115 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
2116
2117 // Get boolean condition to test
2118 Node *i1 = iff->in(1);
2119 if( !i1->is_Bool() ) continue;
2120 BoolNode *bol = i1->as_Bool();
2121 BoolTest b_test = bol->_test;
2122 // Flip sense of test if exit condition is flipped
2123 if( flip )
2124 b_test = b_test.negate();
2125
2126 // Get compare
2127 Node *cmp = bol->in(1);
2128
2129 // Look for trip_counter + offset vs limit
2130 Node *rc_exp = cmp->in(1);
2131 Node *limit = cmp->in(2);
2132 int scale_con= 1; // Assume trip counter not scaled
2133
2134 Node *limit_c = get_ctrl(limit);
2135 if( loop->is_member(get_loop(limit_c) ) ) {
2136 // Compare might have operands swapped; commute them
2137 b_test = b_test.commute();
2138 rc_exp = cmp->in(2);
2139 limit = cmp->in(1);
2140 limit_c = get_ctrl(limit);
2141 if( loop->is_member(get_loop(limit_c) ) )
2142 continue; // Both inputs are loop varying; cannot RCE
2143 }
2144 // Here we know 'limit' is loop invariant
2145
2146 // 'limit' maybe pinned below the zero trip test (probably from a
2147 // previous round of rce), in which case, it can't be used in the
2148 // zero trip test expression which must occur before the zero test's if.
2149 if( limit_c == ctrl ) {
2150 continue; // Don't rce this check but continue looking for other candidates.
2151 }
2152
2153 // Check for scaled induction variable plus an offset
2154 Node *offset = NULL;
2155
2156 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
2157 continue;
2158 }
2159
2160 Node *offset_c = get_ctrl(offset);
2161 if( loop->is_member( get_loop(offset_c) ) )
2162 continue; // Offset is not really loop invariant
2163 // Here we know 'offset' is loop invariant.
2164
2165 // As above for the 'limit', the 'offset' maybe pinned below the
2166 // zero trip test.
2167 if( offset_c == ctrl ) {
2168 continue; // Don't rce this check but continue looking for other candidates.
2169 }
2170 #ifdef ASSERT
2171 if (TraceRangeLimitCheck) {
2172 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
2173 bol->dump(2);
2174 }
2175 #endif
2176 // At this point we have the expression as:
2177 // scale_con * trip_counter + offset :: limit
2178 // where scale_con, offset and limit are loop invariant. Trip_counter
2179 // monotonically increases by stride_con, a constant. Both (or either)
2180 // stride_con and scale_con can be negative which will flip about the
2181 // sense of the test.
2182
2183 // Adjust pre and main loop limits to guard the correct iteration set
2184 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
2185 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
2186 // The underflow and overflow limits: 0 <= scale*I+offset < limit
2187 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
2188 // (0-offset)/scale could be outside of loop iterations range.
2189 conditional_rc = true;
2190 } else {
2191 if (PrintOpto) {
2192 tty->print_cr("missed RCE opportunity");
2193 }
2194 continue; // In release mode, ignore it
2195 }
2196 } else { // Otherwise work on normal compares
2197 switch( b_test._test ) {
2198 case BoolTest::gt:
2199 // Fall into GE case
2200 case BoolTest::ge:
2201 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
2202 scale_con = -scale_con;
2203 offset = new SubINode( zero, offset );
2204 register_new_node( offset, pre_ctrl );
2205 limit = new SubINode( zero, limit );
2206 register_new_node( limit, pre_ctrl );
2207 // Fall into LE case
2208 case BoolTest::le:
2209 if (b_test._test != BoolTest::gt) {
2210 // Convert X <= Y to X < Y+1
2211 limit = new AddINode( limit, one );
2212 register_new_node( limit, pre_ctrl );
2213 }
2214 // Fall into LT case
2215 case BoolTest::lt:
2216 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
2217 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
2218 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
2219 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
2220 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
2221 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
2222 // still be outside of loop range.
2223 conditional_rc = true;
2224 break;
2225 default:
2226 if (PrintOpto) {
2227 tty->print_cr("missed RCE opportunity");
2228 }
2229 continue; // Unhandled case
2230 }
2231 }
2232
2233 // Kill the eliminated test
2234 C->set_major_progress();
2235 Node *kill_con = _igvn.intcon( 1-flip );
2236 set_ctrl(kill_con, C->root());
2237 _igvn.replace_input_of(iff, 1, kill_con);
2238 // Find surviving projection
2239 assert(iff->is_If(), "");
2240 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2241 // Find loads off the surviving projection; remove their control edge
2242 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2243 Node* cd = dp->fast_out(i); // Control-dependent node
2244 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop
2245 // Allow the load to float around in the loop, or before it
2246 // but NOT before the pre-loop.
2247 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
2248 --i;
2249 --imax;
2250 }
2251 }
2252 if (limit->Opcode() == Op_LoadRange) {
2253 closed_range_checks--;
2254 }
2255
2256 } // End of is IF
2257
2258 }
2259
2260 // Update loop limits
2261 if (conditional_rc) {
2262 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit)
2263 : (Node*)new MaxINode(pre_limit, orig_limit);
2264 register_new_node(pre_limit, pre_ctrl);
2265 }
2266 _igvn.replace_input_of(pre_opaq, 1, pre_limit);
2267
2268 // Note:: we are making the main loop limit no longer precise;
2269 // need to round up based on stride.
2270 cl->set_nonexact_trip_count();
2271 Node *main_cle = cl->loopexit();
2272 Node *main_bol = main_cle->in(1);
2273 // Hacking loop bounds; need private copies of exit test
2274 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2275 main_bol = main_bol->clone();// Clone a private BoolNode
2276 register_new_node( main_bol, main_cle->in(0) );
2277 _igvn.replace_input_of(main_cle, 1, main_bol);
2278 }
2279 Node *main_cmp = main_bol->in(1);
2280 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2281 main_cmp = main_cmp->clone();// Clone a private CmpNode
2282 register_new_node( main_cmp, main_cle->in(0) );
2283 _igvn.replace_input_of(main_bol, 1, main_cmp);
2284 }
2285 // Hack the now-private loop bounds
2286 _igvn.replace_input_of(main_cmp, 2, main_limit);
2287 // The OpaqueNode is unshared by design
2288 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2289 _igvn.replace_input_of(opqzm, 1, main_limit);
2290
2291 return closed_range_checks;
2292 }
2293
2294 //------------------------------has_range_checks-------------------------------
2295 // Check to see if RCE cleaned the current loop of range-checks.
2296 void PhaseIdealLoop::has_range_checks(IdealLoopTree *loop) {
2297 assert(RangeCheckElimination, "");
2298
2299 // skip if not a counted loop
2300 if (!loop->is_counted()) return;
2301
2302 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2303
2304 // skip this loop if it is already checked
2305 if (cl->has_been_range_checked()) return;
2306
2307 // Now check for existence of range checks
2308 for (uint i = 0; i < loop->_body.size(); i++) {
2309 Node *iff = loop->_body[i];
2310 int iff_opc = iff->Opcode();
2311 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
2312 cl->mark_has_range_checks();
2313 break;
2314 }
2315 }
2316 cl->set_has_been_range_checked();
2317 }
2318
2319 //-------------------------multi_version_post_loops----------------------------
2320 // Check the range checks that remain, if simple, use the bounds to guard
2321 // which version to a post loop we execute, one with range checks or one without
2322 bool PhaseIdealLoop::multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop) {
2323 bool multi_version_succeeded = false;
2324 assert(RangeCheckElimination, "");
2325 CountedLoopNode *legacy_cl = legacy_loop->_head->as_CountedLoop();
2326 assert(legacy_cl->is_post_loop(), "");
2327
2328 // Check for existence of range checks using the unique instance to make a guard with
2329 Unique_Node_List worklist;
2330 for (uint i = 0; i < legacy_loop->_body.size(); i++) {
2331 Node *iff = legacy_loop->_body[i];
2332 int iff_opc = iff->Opcode();
2333 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
2334 worklist.push(iff);
2335 }
2336 }
2337
2338 // Find RCE'd post loop so that we can stage its guard.
2339 if (!is_canonical_loop_entry(legacy_cl)) return multi_version_succeeded;
2340 Node* ctrl = legacy_cl->in(LoopNode::EntryControl);
2341 Node* iffm = ctrl->in(0);
2342
2343 // Now we test that both the post loops are connected
2344 Node* post_loop_region = iffm->in(0);
2345 if (post_loop_region == NULL) return multi_version_succeeded;
2346 if (!post_loop_region->is_Region()) return multi_version_succeeded;
2347 Node* covering_region = post_loop_region->in(RegionNode::Control+1);
2348 if (covering_region == NULL) return multi_version_succeeded;
2349 if (!covering_region->is_Region()) return multi_version_succeeded;
2350 Node* p_f = covering_region->in(RegionNode::Control);
2351 if (p_f == NULL) return multi_version_succeeded;
2352 if (!p_f->is_IfFalse()) return multi_version_succeeded;
2353 if (!p_f->in(0)->is_CountedLoopEnd()) return multi_version_succeeded;
2354 CountedLoopEndNode* rce_loop_end = p_f->in(0)->as_CountedLoopEnd();
2355 if (rce_loop_end == NULL) return multi_version_succeeded;
2356 CountedLoopNode* rce_cl = rce_loop_end->loopnode();
2357 if (rce_cl == NULL || !rce_cl->is_post_loop()) return multi_version_succeeded;
2358 CountedLoopNode *known_rce_cl = rce_loop->_head->as_CountedLoop();
2359 if (rce_cl != known_rce_cl) return multi_version_succeeded;
2360
2361 // Then we fetch the cover entry test
2362 ctrl = rce_cl->in(LoopNode::EntryControl);
2363 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return multi_version_succeeded;
2364
2365 #ifndef PRODUCT
2366 if (TraceLoopOpts) {
2367 tty->print("PostMultiVersion\n");
2368 rce_loop->dump_head();
2369 legacy_loop->dump_head();
2370 }
2371 #endif
2372
2373 // Now fetch the limit we want to compare against
2374 Node *limit = rce_cl->limit();
2375 bool first_time = true;
2376
2377 // If we got this far, we identified the post loop which has been RCE'd and
2378 // we have a work list. Now we will try to transform the if guard to cause
2379 // the loop pair to be multi version executed with the determination left to runtime
2380 // or the optimizer if full information is known about the given arrays at compile time.
2381 Node *last_min = NULL;
2382 multi_version_succeeded = true;
2383 while (worklist.size()) {
2384 Node* rc_iffm = worklist.pop();
2385 if (rc_iffm->is_If()) {
2386 Node *rc_bolzm = rc_iffm->in(1);
2387 if (rc_bolzm->is_Bool()) {
2388 Node *rc_cmpzm = rc_bolzm->in(1);
2389 if (rc_cmpzm->is_Cmp()) {
2390 Node *rc_left = rc_cmpzm->in(2);
2391 if (rc_left->Opcode() != Op_LoadRange) {
2392 multi_version_succeeded = false;
2393 break;
2394 }
2395 if (first_time) {
2396 last_min = rc_left;
2397 first_time = false;
2398 } else {
2399 Node *cur_min = new MinINode(last_min, rc_left);
2400 last_min = cur_min;
2401 _igvn.register_new_node_with_optimizer(last_min);
2402 }
2403 }
2404 }
2405 }
2406 }
2407
2408 // All we have to do is update the limit of the rce loop
2409 // with the min of our expression and the current limit.
2410 // We will use this expression to replace the current limit.
2411 if (last_min && multi_version_succeeded) {
2412 Node *cur_min = new MinINode(last_min, limit);
2413 _igvn.register_new_node_with_optimizer(cur_min);
2414 Node *cmp_node = rce_loop_end->cmp_node();
2415 _igvn.replace_input_of(cmp_node, 2, cur_min);
2416 set_idom(cmp_node, cur_min, dom_depth(ctrl));
2417 set_ctrl(cur_min, ctrl);
2418 set_loop(cur_min, rce_loop->_parent);
2419
2420 legacy_cl->mark_is_multiversioned();
2421 rce_cl->mark_is_multiversioned();
2422 multi_version_succeeded = true;
2423
2424 C->set_major_progress();
2425 }
2426
2427 return multi_version_succeeded;
2428 }
2429
2430 //-------------------------poison_rce_post_loop--------------------------------
2431 // Causes the rce'd post loop to be optimized away if multiversioning fails
2432 void PhaseIdealLoop::poison_rce_post_loop(IdealLoopTree *rce_loop) {
2433 CountedLoopNode *rce_cl = rce_loop->_head->as_CountedLoop();
2434 Node* ctrl = rce_cl->in(LoopNode::EntryControl);
2435 if (ctrl->is_IfTrue() || ctrl->is_IfFalse()) {
2436 Node* iffm = ctrl->in(0);
2437 if (iffm->is_If()) {
2438 Node* cur_bool = iffm->in(1);
2439 if (cur_bool->is_Bool()) {
2440 Node* cur_cmp = cur_bool->in(1);
2441 if (cur_cmp->is_Cmp()) {
2442 BoolTest::mask new_test = BoolTest::gt;
2443 BoolNode *new_bool = new BoolNode(cur_cmp, new_test);
2444 _igvn.replace_node(cur_bool, new_bool);
2445 _igvn._worklist.push(new_bool);
2446 Node* left_op = cur_cmp->in(1);
2447 _igvn.replace_input_of(cur_cmp, 2, left_op);
2448 C->set_major_progress();
2449 }
2450 }
2451 }
2452 }
2453 }
2454
2455 //------------------------------DCE_loop_body----------------------------------
2456 // Remove simplistic dead code from loop body
2457 void IdealLoopTree::DCE_loop_body() {
2458 for( uint i = 0; i < _body.size(); i++ )
2459 if( _body.at(i)->outcnt() == 0 )
2460 _body.map( i--, _body.pop() );
2461 }
2462
2463
2464 //------------------------------adjust_loop_exit_prob--------------------------
2465 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2466 // Replace with a 1-in-10 exit guess.
2467 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2468 Node *test = tail();
2469 while( test != _head ) {
2470 uint top = test->Opcode();
2471 if( top == Op_IfTrue || top == Op_IfFalse ) {
2472 int test_con = ((ProjNode*)test)->_con;
2473 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2474 IfNode *iff = test->in(0)->as_If();
2475 if( iff->outcnt() == 2 ) { // Ignore dead tests
2476 Node *bol = iff->in(1);
2477 if( bol && bol->req() > 1 && bol->in(1) &&
2478 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2479 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2480 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2481 (bol->in(1)->Opcode() == Op_CompareAndExchangeB ) ||
2482 (bol->in(1)->Opcode() == Op_CompareAndExchangeS ) ||
2483 (bol->in(1)->Opcode() == Op_CompareAndExchangeI ) ||
2484 (bol->in(1)->Opcode() == Op_CompareAndExchangeL ) ||
2485 (bol->in(1)->Opcode() == Op_CompareAndExchangeP ) ||
2486 (bol->in(1)->Opcode() == Op_CompareAndExchangeN ) ||
2487 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapB ) ||
2488 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapS ) ||
2489 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapI ) ||
2490 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapL ) ||
2491 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapP ) ||
2492 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapN ) ||
2493 (bol->in(1)->Opcode() == Op_CompareAndSwapB ) ||
2494 (bol->in(1)->Opcode() == Op_CompareAndSwapS ) ||
2495 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2496 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2497 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2498 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2499 return; // Allocation loops RARELY take backedge
2500 // Find the OTHER exit path from the IF
2501 Node* ex = iff->proj_out(1-test_con);
2502 float p = iff->_prob;
2503 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2504 if( top == Op_IfTrue ) {
2505 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2506 iff->_prob = PROB_STATIC_FREQUENT;
2507 }
2508 } else {
2509 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2510 iff->_prob = PROB_STATIC_INFREQUENT;
2511 }
2512 }
2513 }
2514 }
2515 }
2516 test = phase->idom(test);
2517 }
2518 }
2519
2520 #ifdef ASSERT
2521 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) {
2522 Node *ctrl = cl->in(LoopNode::EntryControl);
2523 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
2524 Node *iffm = ctrl->in(0);
2525 assert(iffm->Opcode() == Op_If, "");
2526 Node *p_f = iffm->in(0);
2527 assert(p_f->Opcode() == Op_IfFalse, "");
2528 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2529 assert(pre_end->loopnode()->is_pre_loop(), "");
2530 return pre_end->loopnode();
2531 }
2532 #endif
2533
2534 // Remove the main and post loops and make the pre loop execute all
2535 // iterations. Useful when the pre loop is found empty.
2536 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) {
2537 CountedLoopEndNode* pre_end = cl->loopexit();
2538 Node* pre_cmp = pre_end->cmp_node();
2539 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) {
2540 // Only safe to remove the main loop if the compiler optimized it
2541 // out based on an unknown number of iterations
2542 return;
2543 }
2544
2545 // Can we find the main loop?
2546 if (_next == NULL) {
2547 return;
2548 }
2549
2550 Node* next_head = _next->_head;
2551 if (!next_head->is_CountedLoop()) {
2552 return;
2553 }
2554
2555 CountedLoopNode* main_head = next_head->as_CountedLoop();
2556 if (!main_head->is_main_loop()) {
2557 return;
2558 }
2559
2560 assert(locate_pre_from_main(main_head) == cl, "bad main loop");
2561 Node* main_iff = main_head->in(LoopNode::EntryControl)->in(0);
2562
2563 // Remove the Opaque1Node of the pre loop and make it execute all iterations
2564 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2));
2565 // Remove the Opaque1Node of the main loop so it can be optimized out
2566 Node* main_cmp = main_iff->in(1)->in(1);
2567 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?");
2568 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1));
2569 }
2570
2571 //------------------------------policy_do_remove_empty_loop--------------------
2572 // Micro-benchmark spamming. Policy is to always remove empty loops.
2573 // The 'DO' part is to replace the trip counter with the value it will
2574 // have on the last iteration. This will break the loop.
2575 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2576 // Minimum size must be empty loop
2577 if (_body.size() > EMPTY_LOOP_SIZE)
2578 return false;
2579
2580 if (!_head->is_CountedLoop())
2581 return false; // Dead loop
2582 CountedLoopNode *cl = _head->as_CountedLoop();
2583 if (!cl->is_valid_counted_loop())
2584 return false; // Malformed loop
2585 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2586 return false; // Infinite loop
2587
2588 if (cl->is_pre_loop()) {
2589 // If the loop we are removing is a pre-loop then the main and
2590 // post loop can be removed as well
2591 remove_main_post_loops(cl, phase);
2592 }
2593
2594 #ifdef ASSERT
2595 // Ensure only one phi which is the iv.
2596 Node* iv = NULL;
2597 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2598 Node* n = cl->fast_out(i);
2599 if (n->Opcode() == Op_Phi) {
2600 assert(iv == NULL, "Too many phis" );
2601 iv = n;
2602 }
2603 }
2604 assert(iv == cl->phi(), "Wrong phi" );
2605 #endif
2606
2607 // main and post loops have explicitly created zero trip guard
2608 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2609 if (needs_guard) {
2610 // Skip guard if values not overlap.
2611 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2612 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2613 int stride_con = cl->stride_con();
2614 if (stride_con > 0) {
2615 needs_guard = (init_t->_hi >= limit_t->_lo);
2616 } else {
2617 needs_guard = (init_t->_lo <= limit_t->_hi);
2618 }
2619 }
2620 if (needs_guard) {
2621 // Check for an obvious zero trip guard.
2622 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2623 if (inctrl->Opcode() == Op_IfTrue || inctrl->Opcode() == Op_IfFalse) {
2624 bool maybe_swapped = (inctrl->Opcode() == Op_IfFalse);
2625 // The test should look like just the backedge of a CountedLoop
2626 Node* iff = inctrl->in(0);
2627 if (iff->is_If()) {
2628 Node* bol = iff->in(1);
2629 if (bol->is_Bool()) {
2630 BoolTest test = bol->as_Bool()->_test;
2631 if (maybe_swapped) {
2632 test._test = test.commute();
2633 test._test = test.negate();
2634 }
2635 if (test._test == cl->loopexit()->test_trip()) {
2636 Node* cmp = bol->in(1);
2637 int init_idx = maybe_swapped ? 2 : 1;
2638 int limit_idx = maybe_swapped ? 1 : 2;
2639 if (cmp->is_Cmp() && cmp->in(init_idx) == cl->init_trip() && cmp->in(limit_idx) == cl->limit()) {
2640 needs_guard = false;
2641 }
2642 }
2643 }
2644 }
2645 }
2646 }
2647
2648 #ifndef PRODUCT
2649 if (PrintOpto) {
2650 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2651 this->dump_head();
2652 } else if (TraceLoopOpts) {
2653 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2654 this->dump_head();
2655 }
2656 #endif
2657
2658 if (needs_guard) {
2659 // Peel the loop to ensure there's a zero trip guard
2660 Node_List old_new;
2661 phase->do_peeling(this, old_new);
2662 }
2663
2664 // Replace the phi at loop head with the final value of the last
2665 // iteration. Then the CountedLoopEnd will collapse (backedge never
2666 // taken) and all loop-invariant uses of the exit values will be correct.
2667 Node *phi = cl->phi();
2668 Node *exact_limit = phase->exact_limit(this);
2669 if (exact_limit != cl->limit()) {
2670 // We also need to replace the original limit to collapse loop exit.
2671 Node* cmp = cl->loopexit()->cmp_node();
2672 assert(cl->limit() == cmp->in(2), "sanity");
2673 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2674 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2675 }
2676 // Note: the final value after increment should not overflow since
2677 // counted loop has limit check predicate.
2678 Node *final = new SubINode( exact_limit, cl->stride() );
2679 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2680 phase->_igvn.replace_node(phi,final);
2681 phase->C->set_major_progress();
2682 return true;
2683 }
2684
2685 //------------------------------policy_do_one_iteration_loop-------------------
2686 // Convert one iteration loop into normal code.
2687 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2688 if (!_head->as_Loop()->is_valid_counted_loop())
2689 return false; // Only for counted loop
2690
2691 CountedLoopNode *cl = _head->as_CountedLoop();
2692 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2693 return false;
2694 }
2695
2696 #ifndef PRODUCT
2697 if(TraceLoopOpts) {
2698 tty->print("OneIteration ");
2699 this->dump_head();
2700 }
2701 #endif
2702
2703 Node *init_n = cl->init_trip();
2704 #ifdef ASSERT
2705 // Loop boundaries should be constant since trip count is exact.
2706 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2707 #endif
2708 // Replace the phi at loop head with the value of the init_trip.
2709 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2710 // and all loop-invariant uses of the exit values will be correct.
2711 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2712 phase->C->set_major_progress();
2713 return true;
2714 }
2715
2716 //=============================================================================
2717 //------------------------------iteration_split_impl---------------------------
2718 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2719 // Compute loop trip count if possible.
2720 compute_trip_count(phase);
2721
2722 // Convert one iteration loop into normal code.
2723 if (policy_do_one_iteration_loop(phase))
2724 return true;
2725
2726 // Check and remove empty loops (spam micro-benchmarks)
2727 if (policy_do_remove_empty_loop(phase))
2728 return true; // Here we removed an empty loop
2729
2730 bool should_peel = policy_peeling(phase); // Should we peel?
2731
2732 bool should_unswitch = policy_unswitching(phase);
2733
2734 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2735 // This removes loop-invariant tests (usually null checks).
2736 if (!_head->is_CountedLoop()) { // Non-counted loop
2737 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2738 // Partial peel succeeded so terminate this round of loop opts
2739 return false;
2740 }
2741 if (should_peel) { // Should we peel?
2742 if (PrintOpto) { tty->print_cr("should_peel"); }
2743 phase->do_peeling(this,old_new);
2744 } else if (should_unswitch) {
2745 phase->do_unswitching(this, old_new);
2746 }
2747 return true;
2748 }
2749 CountedLoopNode *cl = _head->as_CountedLoop();
2750
2751 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2752
2753 // Do nothing special to pre- and post- loops
2754 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2755
2756 // Compute loop trip count from profile data
2757 compute_profile_trip_cnt(phase);
2758
2759 // Before attempting fancy unrolling, RCE or alignment, see if we want
2760 // to completely unroll this loop or do loop unswitching.
2761 if (cl->is_normal_loop()) {
2762 if (should_unswitch) {
2763 phase->do_unswitching(this, old_new);
2764 return true;
2765 }
2766 bool should_maximally_unroll = policy_maximally_unroll(phase);
2767 if (should_maximally_unroll) {
2768 // Here we did some unrolling and peeling. Eventually we will
2769 // completely unroll this loop and it will no longer be a loop.
2770 phase->do_maximally_unroll(this,old_new);
2771 return true;
2772 }
2773 }
2774
2775 // Skip next optimizations if running low on nodes. Note that
2776 // policy_unswitching and policy_maximally_unroll have this check.
2777 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes();
2778 if ((int)(2 * _body.size()) > nodes_left) {
2779 return true;
2780 }
2781
2782 // Counted loops may be peeled, may need some iterations run up
2783 // front for RCE, and may want to align loop refs to a cache
2784 // line. Thus we clone a full loop up front whose trip count is
2785 // at least 1 (if peeling), but may be several more.
2786
2787 // The main loop will start cache-line aligned with at least 1
2788 // iteration of the unrolled body (zero-trip test required) and
2789 // will have some range checks removed.
2790
2791 // A post-loop will finish any odd iterations (leftover after
2792 // unrolling), plus any needed for RCE purposes.
2793
2794 bool should_unroll = policy_unroll(phase);
2795
2796 bool should_rce = policy_range_check(phase);
2797
2798 bool should_align = policy_align(phase);
2799
2800 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2801 // need a pre-loop. We may still need to peel an initial iteration but
2802 // we will not be needing an unknown number of pre-iterations.
2803 //
2804 // Basically, if may_rce_align reports FALSE first time through,
2805 // we will not be able to later do RCE or Aligning on this loop.
2806 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2807
2808 // If we have any of these conditions (RCE, alignment, unrolling) met, then
2809 // we switch to the pre-/main-/post-loop model. This model also covers
2810 // peeling.
2811 if (should_rce || should_align || should_unroll) {
2812 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
2813 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2814
2815 // Adjust the pre- and main-loop limits to let the pre and post loops run
2816 // with full checks, but the main-loop with no checks. Remove said
2817 // checks from the main body.
2818 if (should_rce) {
2819 if (phase->do_range_check(this, old_new) != 0) {
2820 cl->mark_has_range_checks();
2821 }
2822 } else if (PostLoopMultiversioning) {
2823 phase->has_range_checks(this);
2824 }
2825
2826 if (should_unroll && !should_peel && PostLoopMultiversioning) {
2827 // Try to setup multiversioning on main loops before they are unrolled
2828 if (cl->is_main_loop() && (cl->unrolled_count() == 1)) {
2829 phase->insert_scalar_rced_post_loop(this, old_new);
2830 }
2831 }
2832
2833 // Double loop body for unrolling. Adjust the minimum-trip test (will do
2834 // twice as many iterations as before) and the main body limit (only do
2835 // an even number of trips). If we are peeling, we might enable some RCE
2836 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2837 // peeling.
2838 if (should_unroll && !should_peel) {
2839 if (SuperWordLoopUnrollAnalysis) {
2840 phase->insert_vector_post_loop(this, old_new);
2841 }
2842 phase->do_unroll(this, old_new, true);
2843 }
2844
2845 // Adjust the pre-loop limits to align the main body
2846 // iterations.
2847 if (should_align)
2848 Unimplemented();
2849
2850 } else { // Else we have an unchanged counted loop
2851 if (should_peel) // Might want to peel but do nothing else
2852 phase->do_peeling(this,old_new);
2853 }
2854 return true;
2855 }
2856
2857
2858 //=============================================================================
2859 //------------------------------iteration_split--------------------------------
2860 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2861 // Recursively iteration split nested loops
2862 if (_child && !_child->iteration_split(phase, old_new))
2863 return false;
2864
2865 // Clean out prior deadwood
2866 DCE_loop_body();
2867
2868
2869 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2870 // Replace with a 1-in-10 exit guess.
2871 if (_parent /*not the root loop*/ &&
2872 !_irreducible &&
2873 // Also ignore the occasional dead backedge
2874 !tail()->is_top()) {
2875 adjust_loop_exit_prob(phase);
2876 }
2877
2878 // Gate unrolling, RCE and peeling efforts.
2879 if (!_child && // If not an inner loop, do not split
2880 !_irreducible &&
2881 _allow_optimizations &&
2882 !tail()->is_top()) { // Also ignore the occasional dead backedge
2883 if (!_has_call) {
2884 if (!iteration_split_impl(phase, old_new)) {
2885 return false;
2886 }
2887 } else if (policy_unswitching(phase)) {
2888 phase->do_unswitching(this, old_new);
2889 }
2890 }
2891
2892 // Minor offset re-organization to remove loop-fallout uses of
2893 // trip counter when there was no major reshaping.
2894 phase->reorg_offsets(this);
2895
2896 if (_next && !_next->iteration_split(phase, old_new))
2897 return false;
2898 return true;
2899 }
2900
2901
2902 //=============================================================================
2903 // Process all the loops in the loop tree and replace any fill
2904 // patterns with an intrinsic version.
2905 bool PhaseIdealLoop::do_intrinsify_fill() {
2906 bool changed = false;
2907 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2908 IdealLoopTree* lpt = iter.current();
2909 changed |= intrinsify_fill(lpt);
2910 }
2911 return changed;
2912 }
2913
2914
2915 // Examine an inner loop looking for a a single store of an invariant
2916 // value in a unit stride loop,
2917 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2918 Node*& shift, Node*& con) {
2919 const char* msg = NULL;
2920 Node* msg_node = NULL;
2921
2922 store_value = NULL;
2923 con = NULL;
2924 shift = NULL;
2925
2926 // Process the loop looking for stores. If there are multiple
2927 // stores or extra control flow give at this point.
2928 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2929 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2930 Node* n = lpt->_body.at(i);
2931 if (n->outcnt() == 0) continue; // Ignore dead
2932 if (n->is_Store()) {
2933 if (store != NULL) {
2934 msg = "multiple stores";
2935 break;
2936 }
2937 int opc = n->Opcode();
2938 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
2939 msg = "oop fills not handled";
2940 break;
2941 }
2942 Node* value = n->in(MemNode::ValueIn);
2943 if (!lpt->is_invariant(value)) {
2944 msg = "variant store value";
2945 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2946 msg = "not array address";
2947 }
2948 store = n;
2949 store_value = value;
2950 } else if (n->is_If() && n != head->loopexit()) {
2951 msg = "extra control flow";
2952 msg_node = n;
2953 }
2954 }
2955
2956 if (store == NULL) {
2957 // No store in loop
2958 return false;
2959 }
2960
2961 if (msg == NULL && head->stride_con() != 1) {
2962 // could handle negative strides too
2963 if (head->stride_con() < 0) {
2964 msg = "negative stride";
2965 } else {
2966 msg = "non-unit stride";
2967 }
2968 }
2969
2970 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2971 msg = "can't handle store address";
2972 msg_node = store->in(MemNode::Address);
2973 }
2974
2975 if (msg == NULL &&
2976 (!store->in(MemNode::Memory)->is_Phi() ||
2977 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2978 msg = "store memory isn't proper phi";
2979 msg_node = store->in(MemNode::Memory);
2980 }
2981
2982 // Make sure there is an appropriate fill routine
2983 BasicType t = store->as_Mem()->memory_type();
2984 const char* fill_name;
2985 if (msg == NULL &&
2986 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2987 msg = "unsupported store";
2988 msg_node = store;
2989 }
2990
2991 if (msg != NULL) {
2992 #ifndef PRODUCT
2993 if (TraceOptimizeFill) {
2994 tty->print_cr("not fill intrinsic candidate: %s", msg);
2995 if (msg_node != NULL) msg_node->dump();
2996 }
2997 #endif
2998 return false;
2999 }
3000
3001 // Make sure the address expression can be handled. It should be
3002 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()).
3003 Node* elements[4];
3004 Node* cast = NULL;
3005 Node* conv = NULL;
3006 bool found_index = false;
3007 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
3008 for (int e = 0; e < count; e++) {
3009 Node* n = elements[e];
3010 if (n->is_Con() && con == NULL) {
3011 con = n;
3012 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
3013 Node* value = n->in(1);
3014 #ifdef _LP64
3015 if (value->Opcode() == Op_ConvI2L) {
3016 conv = value;
3017 value = value->in(1);
3018 }
3019 if (value->Opcode() == Op_CastII &&
3020 value->as_CastII()->has_range_check()) {
3021 // Skip range check dependent CastII nodes
3022 cast = value;
3023 value = value->in(1);
3024 }
3025 #endif
3026 if (value != head->phi()) {
3027 msg = "unhandled shift in address";
3028 } else {
3029 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
3030 msg = "scale doesn't match";
3031 } else {
3032 found_index = true;
3033 shift = n;
3034 }
3035 }
3036 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
3037 conv = n;
3038 n = n->in(1);
3039 if (n->Opcode() == Op_CastII &&
3040 n->as_CastII()->has_range_check()) {
3041 // Skip range check dependent CastII nodes
3042 cast = n;
3043 n = n->in(1);
3044 }
3045 if (n == head->phi()) {
3046 found_index = true;
3047 } else {
3048 msg = "unhandled input to ConvI2L";
3049 }
3050 } else if (n == head->phi()) {
3051 // no shift, check below for allowed cases
3052 found_index = true;
3053 } else {
3054 msg = "unhandled node in address";
3055 msg_node = n;
3056 }
3057 }
3058
3059 if (count == -1) {
3060 msg = "malformed address expression";
3061 msg_node = store;
3062 }
3063
3064 if (!found_index) {
3065 msg = "missing use of index";
3066 }
3067
3068 // byte sized items won't have a shift
3069 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
3070 msg = "can't find shift";
3071 msg_node = store;
3072 }
3073
3074 if (msg != NULL) {
3075 #ifndef PRODUCT
3076 if (TraceOptimizeFill) {
3077 tty->print_cr("not fill intrinsic: %s", msg);
3078 if (msg_node != NULL) msg_node->dump();
3079 }
3080 #endif
3081 return false;
3082 }
3083
3084 // No make sure all the other nodes in the loop can be handled
3085 VectorSet ok(Thread::current()->resource_area());
3086
3087 // store related values are ok
3088 ok.set(store->_idx);
3089 ok.set(store->in(MemNode::Memory)->_idx);
3090
3091 CountedLoopEndNode* loop_exit = head->loopexit();
3092 guarantee(loop_exit != NULL, "no loop exit node");
3093
3094 // Loop structure is ok
3095 ok.set(head->_idx);
3096 ok.set(loop_exit->_idx);
3097 ok.set(head->phi()->_idx);
3098 ok.set(head->incr()->_idx);
3099 ok.set(loop_exit->cmp_node()->_idx);
3100 ok.set(loop_exit->in(1)->_idx);
3101
3102 // Address elements are ok
3103 if (con) ok.set(con->_idx);
3104 if (shift) ok.set(shift->_idx);
3105 if (cast) ok.set(cast->_idx);
3106 if (conv) ok.set(conv->_idx);
3107
3108 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3109 Node* n = lpt->_body.at(i);
3110 if (n->outcnt() == 0) continue; // Ignore dead
3111 if (ok.test(n->_idx)) continue;
3112 // Backedge projection is ok
3113 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
3114 if (!n->is_AddP()) {
3115 msg = "unhandled node";
3116 msg_node = n;
3117 break;
3118 }
3119 }
3120
3121 // Make sure no unexpected values are used outside the loop
3122 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3123 Node* n = lpt->_body.at(i);
3124 // These values can be replaced with other nodes if they are used
3125 // outside the loop.
3126 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
3127 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
3128 Node* use = iter.get();
3129 if (!lpt->_body.contains(use)) {
3130 msg = "node is used outside loop";
3131 // lpt->_body.dump();
3132 msg_node = n;
3133 break;
3134 }
3135 }
3136 }
3137
3138 #ifdef ASSERT
3139 if (TraceOptimizeFill) {
3140 if (msg != NULL) {
3141 tty->print_cr("no fill intrinsic: %s", msg);
3142 if (msg_node != NULL) msg_node->dump();
3143 } else {
3144 tty->print_cr("fill intrinsic for:");
3145 }
3146 store->dump();
3147 if (Verbose) {
3148 lpt->_body.dump();
3149 }
3150 }
3151 #endif
3152
3153 return msg == NULL;
3154 }
3155
3156
3157
3158 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
3159 // Only for counted inner loops
3160 if (!lpt->is_counted() || !lpt->is_inner()) {
3161 return false;
3162 }
3163
3164 // Must have constant stride
3165 CountedLoopNode* head = lpt->_head->as_CountedLoop();
3166 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
3167 return false;
3168 }
3169
3170 // Check that the body only contains a store of a loop invariant
3171 // value that is indexed by the loop phi.
3172 Node* store = NULL;
3173 Node* store_value = NULL;
3174 Node* shift = NULL;
3175 Node* offset = NULL;
3176 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
3177 return false;
3178 }
3179
3180 Node* exit = head->loopexit()->proj_out(0);
3181 if (exit == NULL) {
3182 return false;
3183 }
3184
3185 #ifndef PRODUCT
3186 if (TraceLoopOpts) {
3187 tty->print("ArrayFill ");
3188 lpt->dump_head();
3189 }
3190 #endif
3191
3192 // Now replace the whole loop body by a call to a fill routine that
3193 // covers the same region as the loop.
3194 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
3195
3196 // Build an expression for the beginning of the copy region
3197 Node* index = head->init_trip();
3198 #ifdef _LP64
3199 index = new ConvI2LNode(index);
3200 _igvn.register_new_node_with_optimizer(index);
3201 #endif
3202 if (shift != NULL) {
3203 // byte arrays don't require a shift but others do.
3204 index = new LShiftXNode(index, shift->in(2));
3205 _igvn.register_new_node_with_optimizer(index);
3206 }
3207 index = new AddPNode(base, base, index);
3208 _igvn.register_new_node_with_optimizer(index);
3209 Node* from = new AddPNode(base, index, offset);
3210 _igvn.register_new_node_with_optimizer(from);
3211 // Compute the number of elements to copy
3212 Node* len = new SubINode(head->limit(), head->init_trip());
3213 _igvn.register_new_node_with_optimizer(len);
3214
3215 BasicType t = store->as_Mem()->memory_type();
3216 bool aligned = false;
3217 if (offset != NULL && head->init_trip()->is_Con()) {
3218 int element_size = type2aelembytes(t);
3219 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
3220 }
3221
3222 // Build a call to the fill routine
3223 const char* fill_name;
3224 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
3225 assert(fill != NULL, "what?");
3226
3227 // Convert float/double to int/long for fill routines
3228 if (t == T_FLOAT) {
3229 store_value = new MoveF2INode(store_value);
3230 _igvn.register_new_node_with_optimizer(store_value);
3231 } else if (t == T_DOUBLE) {
3232 store_value = new MoveD2LNode(store_value);
3233 _igvn.register_new_node_with_optimizer(store_value);
3234 }
3235
3236 Node* mem_phi = store->in(MemNode::Memory);
3237 Node* result_ctrl;
3238 Node* result_mem;
3239 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
3240 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill,
3241 fill_name, TypeAryPtr::get_array_body_type(t));
3242 uint cnt = 0;
3243 call->init_req(TypeFunc::Parms + cnt++, from);
3244 call->init_req(TypeFunc::Parms + cnt++, store_value);
3245 #ifdef _LP64
3246 len = new ConvI2LNode(len);
3247 _igvn.register_new_node_with_optimizer(len);
3248 #endif
3249 call->init_req(TypeFunc::Parms + cnt++, len);
3250 #ifdef _LP64
3251 call->init_req(TypeFunc::Parms + cnt++, C->top());
3252 #endif
3253 call->init_req(TypeFunc::Control, head->init_control());
3254 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O.
3255 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl));
3256 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr));
3257 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
3258 _igvn.register_new_node_with_optimizer(call);
3259 result_ctrl = new ProjNode(call,TypeFunc::Control);
3260 _igvn.register_new_node_with_optimizer(result_ctrl);
3261 result_mem = new ProjNode(call,TypeFunc::Memory);
3262 _igvn.register_new_node_with_optimizer(result_mem);
3263
3264 /* Disable following optimization until proper fix (add missing checks).
3265
3266 // If this fill is tightly coupled to an allocation and overwrites
3267 // the whole body, allow it to take over the zeroing.
3268 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
3269 if (alloc != NULL && alloc->is_AllocateArray()) {
3270 Node* length = alloc->as_AllocateArray()->Ideal_length();
3271 if (head->limit() == length &&
3272 head->init_trip() == _igvn.intcon(0)) {
3273 if (TraceOptimizeFill) {
3274 tty->print_cr("Eliminated zeroing in allocation");
3275 }
3276 alloc->maybe_set_complete(&_igvn);
3277 } else {
3278 #ifdef ASSERT
3279 if (TraceOptimizeFill) {
3280 tty->print_cr("filling array but bounds don't match");
3281 alloc->dump();
3282 head->init_trip()->dump();
3283 head->limit()->dump();
3284 length->dump();
3285 }
3286 #endif
3287 }
3288 }
3289 */
3290
3291 // Redirect the old control and memory edges that are outside the loop.
3292 // Sometimes the memory phi of the head is used as the outgoing
3293 // state of the loop. It's safe in this case to replace it with the
3294 // result_mem.
3295 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
3296 lazy_replace(exit, result_ctrl);
3297 _igvn.replace_node(store, result_mem);
3298 // Any uses the increment outside of the loop become the loop limit.
3299 _igvn.replace_node(head->incr(), head->limit());
3300
3301 // Disconnect the head from the loop.
3302 for (uint i = 0; i < lpt->_body.size(); i++) {
3303 Node* n = lpt->_body.at(i);
3304 _igvn.replace_node(n, C->top());
3305 }
3306
3307 return true;
3308 }