1 /*
2 * Copyright (c) 2000, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/convertnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/opaquenode.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/superword.hpp"
42 #include "opto/vectornode.hpp"
43
44 //------------------------------is_loop_exit-----------------------------------
45 // Given an IfNode, return the loop-exiting projection or NULL if both
46 // arms remain in the loop.
47 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
48 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
49 PhaseIdealLoop *phase = _phase;
50 // Test is an IfNode, has 2 projections. If BOTH are in the loop
51 // we need loop unswitching instead of peeling.
52 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
53 return iff->raw_out(0);
54 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
55 return iff->raw_out(1);
56 return NULL;
57 }
58
59
60 //=============================================================================
61
62
63 //------------------------------record_for_igvn----------------------------
64 // Put loop body on igvn work list
65 void IdealLoopTree::record_for_igvn() {
66 for( uint i = 0; i < _body.size(); i++ ) {
67 Node *n = _body.at(i);
68 _phase->_igvn._worklist.push(n);
69 }
70 }
71
72 //------------------------------compute_exact_trip_count-----------------------
73 // Compute loop exact trip count if possible. Do not recalculate trip count for
74 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
75 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) {
76 if (!_head->as_Loop()->is_valid_counted_loop()) {
77 return;
78 }
79 CountedLoopNode* cl = _head->as_CountedLoop();
80 // Trip count may become nonexact for iteration split loops since
81 // RCE modifies limits. Note, _trip_count value is not reset since
82 // it is used to limit unrolling of main loop.
83 cl->set_nonexact_trip_count();
84
85 // Loop's test should be part of loop.
86 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
87 return; // Infinite loop
88
89 #ifdef ASSERT
90 BoolTest::mask bt = cl->loopexit()->test_trip();
91 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
92 bt == BoolTest::ne, "canonical test is expected");
93 #endif
94
95 Node* init_n = cl->init_trip();
96 Node* limit_n = cl->limit();
97 if (init_n != NULL && init_n->is_Con() &&
98 limit_n != NULL && limit_n->is_Con()) {
99 // Use longs to avoid integer overflow.
100 int stride_con = cl->stride_con();
101 jlong init_con = cl->init_trip()->get_int();
102 jlong limit_con = cl->limit()->get_int();
103 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
104 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
105 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
106 // Set exact trip count.
107 cl->set_exact_trip_count((uint)trip_count);
108 }
109 }
110 }
111
112 //------------------------------compute_profile_trip_cnt----------------------------
113 // Compute loop trip count from profile data as
114 // (backedge_count + loop_exit_count) / loop_exit_count
115 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
116 if (!_head->is_CountedLoop()) {
117 return;
118 }
119 CountedLoopNode* head = _head->as_CountedLoop();
120 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
121 return; // Already computed
122 }
123 float trip_cnt = (float)max_jint; // default is big
124
125 Node* back = head->in(LoopNode::LoopBackControl);
126 while (back != head) {
127 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
128 back->in(0) &&
129 back->in(0)->is_If() &&
130 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
131 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
132 break;
133 }
134 back = phase->idom(back);
135 }
136 if (back != head) {
137 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
138 back->in(0), "if-projection exists");
139 IfNode* back_if = back->in(0)->as_If();
140 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
141
142 // Now compute a loop exit count
143 float loop_exit_cnt = 0.0f;
144 for( uint i = 0; i < _body.size(); i++ ) {
145 Node *n = _body[i];
146 if( n->is_If() ) {
147 IfNode *iff = n->as_If();
148 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
149 Node *exit = is_loop_exit(iff);
150 if( exit ) {
151 float exit_prob = iff->_prob;
152 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
153 if (exit_prob > PROB_MIN) {
154 float exit_cnt = iff->_fcnt * exit_prob;
155 loop_exit_cnt += exit_cnt;
156 }
157 }
158 }
159 }
160 }
161 if (loop_exit_cnt > 0.0f) {
162 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
163 } else {
164 // No exit count so use
165 trip_cnt = loop_back_cnt;
166 }
167 }
168 #ifndef PRODUCT
169 if (TraceProfileTripCount) {
170 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
171 }
172 #endif
173 head->set_profile_trip_cnt(trip_cnt);
174 }
175
176 //---------------------is_invariant_addition-----------------------------
177 // Return nonzero index of invariant operand for an Add or Sub
178 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
179 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
180 int op = n->Opcode();
181 if (op == Op_AddI || op == Op_SubI) {
182 bool in1_invar = this->is_invariant(n->in(1));
183 bool in2_invar = this->is_invariant(n->in(2));
184 if (in1_invar && !in2_invar) return 1;
185 if (!in1_invar && in2_invar) return 2;
186 }
187 return 0;
188 }
189
190 //---------------------reassociate_add_sub-----------------------------
191 // Reassociate invariant add and subtract expressions:
192 //
193 // inv1 + (x + inv2) => ( inv1 + inv2) + x
194 // (x + inv2) + inv1 => ( inv1 + inv2) + x
195 // inv1 + (x - inv2) => ( inv1 - inv2) + x
196 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
197 // (x + inv2) - inv1 => (-inv1 + inv2) + x
198 // (x - inv2) + inv1 => ( inv1 - inv2) + x
199 // (x - inv2) - inv1 => (-inv1 - inv2) + x
200 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
201 // inv1 - (x - inv2) => ( inv1 + inv2) - x
202 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
203 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
204 // inv1 - (x + inv2) => ( inv1 - inv2) - x
205 //
206 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
207 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
208 if (is_invariant(n1)) return NULL;
209 int inv1_idx = is_invariant_addition(n1, phase);
210 if (!inv1_idx) return NULL;
211 // Don't mess with add of constant (igvn moves them to expression tree root.)
212 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
213 Node* inv1 = n1->in(inv1_idx);
214 Node* n2 = n1->in(3 - inv1_idx);
215 int inv2_idx = is_invariant_addition(n2, phase);
216 if (!inv2_idx) return NULL;
217 Node* x = n2->in(3 - inv2_idx);
218 Node* inv2 = n2->in(inv2_idx);
219
220 bool neg_x = n2->is_Sub() && inv2_idx == 1;
221 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
222 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
223 if (n1->is_Sub() && inv1_idx == 1) {
224 neg_x = !neg_x;
225 neg_inv2 = !neg_inv2;
226 }
227 Node* inv1_c = phase->get_ctrl(inv1);
228 Node* inv2_c = phase->get_ctrl(inv2);
229 Node* n_inv1;
230 if (neg_inv1) {
231 Node *zero = phase->_igvn.intcon(0);
232 phase->set_ctrl(zero, phase->C->root());
233 n_inv1 = new SubINode(zero, inv1);
234 phase->register_new_node(n_inv1, inv1_c);
235 } else {
236 n_inv1 = inv1;
237 }
238 Node* inv;
239 if (neg_inv2) {
240 inv = new SubINode(n_inv1, inv2);
241 } else {
242 inv = new AddINode(n_inv1, inv2);
243 }
244 phase->register_new_node(inv, phase->get_early_ctrl(inv));
245
246 Node* addx;
247 if (neg_x) {
248 addx = new SubINode(inv, x);
249 } else {
250 addx = new AddINode(x, inv);
251 }
252 phase->register_new_node(addx, phase->get_ctrl(x));
253 phase->_igvn.replace_node(n1, addx);
254 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
255 _body.yank(n1);
256 return addx;
257 }
258
259 //---------------------reassociate_invariants-----------------------------
260 // Reassociate invariant expressions:
261 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
262 for (int i = _body.size() - 1; i >= 0; i--) {
263 Node *n = _body.at(i);
264 for (int j = 0; j < 5; j++) {
265 Node* nn = reassociate_add_sub(n, phase);
266 if (nn == NULL) break;
267 n = nn; // again
268 };
269 }
270 }
271
272 //------------------------------policy_peeling---------------------------------
273 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
274 // make some loop-invariant test (usually a null-check) happen before the loop.
275 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
276 Node *test = ((IdealLoopTree*)this)->tail();
277 int body_size = ((IdealLoopTree*)this)->_body.size();
278 // Peeling does loop cloning which can result in O(N^2) node construction
279 if( body_size > 255 /* Prevent overflow for large body_size */
280 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) {
281 return false; // too large to safely clone
282 }
283 while( test != _head ) { // Scan till run off top of loop
284 if( test->is_If() ) { // Test?
285 Node *ctrl = phase->get_ctrl(test->in(1));
286 if (ctrl->is_top())
287 return false; // Found dead test on live IF? No peeling!
288 // Standard IF only has one input value to check for loop invariance
289 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
290 // Condition is not a member of this loop?
291 if( !is_member(phase->get_loop(ctrl)) &&
292 is_loop_exit(test) )
293 return true; // Found reason to peel!
294 }
295 // Walk up dominators to loop _head looking for test which is
296 // executed on every path thru loop.
297 test = phase->idom(test);
298 }
299 return false;
300 }
301
302 //------------------------------peeled_dom_test_elim---------------------------
303 // If we got the effect of peeling, either by actually peeling or by making
304 // a pre-loop which must execute at least once, we can remove all
305 // loop-invariant dominated tests in the main body.
306 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
307 bool progress = true;
308 while( progress ) {
309 progress = false; // Reset for next iteration
310 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
311 Node *test = prev->in(0);
312 while( test != loop->_head ) { // Scan till run off top of loop
313
314 int p_op = prev->Opcode();
315 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
316 test->is_If() && // Test?
317 !test->in(1)->is_Con() && // And not already obvious?
318 // Condition is not a member of this loop?
319 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
320 // Walk loop body looking for instances of this test
321 for( uint i = 0; i < loop->_body.size(); i++ ) {
322 Node *n = loop->_body.at(i);
323 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
324 // IfNode was dominated by version in peeled loop body
325 progress = true;
326 dominated_by( old_new[prev->_idx], n );
327 }
328 }
329 }
330 prev = test;
331 test = idom(test);
332 } // End of scan tests in loop
333
334 } // End of while( progress )
335 }
336
337 //------------------------------do_peeling-------------------------------------
338 // Peel the first iteration of the given loop.
339 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
340 // The pre-loop illegally has 2 control users (old & new loops).
341 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
342 // Do this by making the old-loop fall-in edges act as if they came
343 // around the loopback from the prior iteration (follow the old-loop
344 // backedges) and then map to the new peeled iteration. This leaves
345 // the pre-loop with only 1 user (the new peeled iteration), but the
346 // peeled-loop backedge has 2 users.
347 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
348 // extra backedge user.
349 //
350 // orig
351 //
352 // stmt1
353 // |
354 // v
355 // loop predicate
356 // |
357 // v
358 // loop<----+
359 // | |
360 // stmt2 |
361 // | |
362 // v |
363 // if ^
364 // / \ |
365 // / \ |
366 // v v |
367 // false true |
368 // / \ |
369 // / ----+
370 // |
371 // v
372 // exit
373 //
374 //
375 // after clone loop
376 //
377 // stmt1
378 // |
379 // v
380 // loop predicate
381 // / \
382 // clone / \ orig
383 // / \
384 // / \
385 // v v
386 // +---->loop clone loop<----+
387 // | | | |
388 // | stmt2 clone stmt2 |
389 // | | | |
390 // | v v |
391 // ^ if clone If ^
392 // | / \ / \ |
393 // | / \ / \ |
394 // | v v v v |
395 // | true false false true |
396 // | / \ / \ |
397 // +---- \ / ----+
398 // \ /
399 // 1v v2
400 // region
401 // |
402 // v
403 // exit
404 //
405 //
406 // after peel and predicate move
407 //
408 // stmt1
409 // /
410 // /
411 // clone / orig
412 // /
413 // / +----------+
414 // / | |
415 // / loop predicate |
416 // / | |
417 // v v |
418 // TOP-->loop clone loop<----+ |
419 // | | | |
420 // stmt2 clone stmt2 | |
421 // | | | ^
422 // v v | |
423 // if clone If ^ |
424 // / \ / \ | |
425 // / \ / \ | |
426 // v v v v | |
427 // true false false true | |
428 // | \ / \ | |
429 // | \ / ----+ ^
430 // | \ / |
431 // | 1v v2 |
432 // v region |
433 // | | |
434 // | v |
435 // | exit |
436 // | |
437 // +--------------->-----------------+
438 //
439 //
440 // final graph
441 //
442 // stmt1
443 // |
444 // v
445 // stmt2 clone
446 // |
447 // v
448 // if clone
449 // / |
450 // / |
451 // v v
452 // false true
453 // | |
454 // | v
455 // | loop predicate
456 // | |
457 // | v
458 // | loop<----+
459 // | | |
460 // | stmt2 |
461 // | | |
462 // | v |
463 // v if ^
464 // | / \ |
465 // | / \ |
466 // | v v |
467 // | false true |
468 // | | \ |
469 // v v --+
470 // region
471 // |
472 // v
473 // exit
474 //
475 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
476
477 C->set_major_progress();
478 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
479 // 'pre' loop from the main and the 'pre' can no longer have its
480 // iterations adjusted. Therefore, we need to declare this loop as
481 // no longer a 'main' loop; it will need new pre and post loops before
482 // we can do further RCE.
483 #ifndef PRODUCT
484 if (TraceLoopOpts) {
485 tty->print("Peel ");
486 loop->dump_head();
487 }
488 #endif
489 Node* head = loop->_head;
490 bool counted_loop = head->is_CountedLoop();
491 if (counted_loop) {
492 CountedLoopNode *cl = head->as_CountedLoop();
493 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
494 cl->set_trip_count(cl->trip_count() - 1);
495 if (cl->is_main_loop()) {
496 cl->set_normal_loop();
497 #ifndef PRODUCT
498 if (PrintOpto && VerifyLoopOptimizations) {
499 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
500 loop->dump_head();
501 }
502 #endif
503 }
504 }
505 Node* entry = head->in(LoopNode::EntryControl);
506
507 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
508 // The pre-loop illegally has 2 control users (old & new loops).
509 clone_loop( loop, old_new, dom_depth(head) );
510
511 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
512 // Do this by making the old-loop fall-in edges act as if they came
513 // around the loopback from the prior iteration (follow the old-loop
514 // backedges) and then map to the new peeled iteration. This leaves
515 // the pre-loop with only 1 user (the new peeled iteration), but the
516 // peeled-loop backedge has 2 users.
517 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
518 _igvn.hash_delete(head);
519 head->set_req(LoopNode::EntryControl, new_entry);
520 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
521 Node* old = head->fast_out(j);
522 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
523 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
524 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
525 // Then loop body backedge value remains the same.
526 new_exit_value = old->in(LoopNode::LoopBackControl);
527 _igvn.hash_delete(old);
528 old->set_req(LoopNode::EntryControl, new_exit_value);
529 }
530 }
531
532
533 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
534 // extra backedge user.
535 Node* new_head = old_new[head->_idx];
536 _igvn.hash_delete(new_head);
537 new_head->set_req(LoopNode::LoopBackControl, C->top());
538 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
539 Node* use = new_head->fast_out(j2);
540 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
541 _igvn.hash_delete(use);
542 use->set_req(LoopNode::LoopBackControl, C->top());
543 }
544 }
545
546
547 // Step 4: Correct dom-depth info. Set to loop-head depth.
548 int dd = dom_depth(head);
549 set_idom(head, head->in(1), dd);
550 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
551 Node *old = loop->_body.at(j3);
552 Node *nnn = old_new[old->_idx];
553 if (!has_ctrl(nnn))
554 set_idom(nnn, idom(nnn), dd-1);
555 }
556
557 // Now force out all loop-invariant dominating tests. The optimizer
558 // finds some, but we _know_ they are all useless.
559 peeled_dom_test_elim(loop,old_new);
560
561 loop->record_for_igvn();
562 }
563
564 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
565
566 //------------------------------policy_maximally_unroll------------------------
567 // Calculate exact loop trip count and return true if loop can be maximally
568 // unrolled.
569 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
570 CountedLoopNode *cl = _head->as_CountedLoop();
571 assert(cl->is_normal_loop(), "");
572 if (!cl->is_valid_counted_loop())
573 return false; // Malformed counted loop
574
575 if (!cl->has_exact_trip_count()) {
576 // Trip count is not exact.
577 return false;
578 }
579
580 uint trip_count = cl->trip_count();
581 // Note, max_juint is used to indicate unknown trip count.
582 assert(trip_count > 1, "one iteration loop should be optimized out already");
583 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
584
585 // Real policy: if we maximally unroll, does it get too big?
586 // Allow the unrolled mess to get larger than standard loop
587 // size. After all, it will no longer be a loop.
588 uint body_size = _body.size();
589 uint unroll_limit = (uint)LoopUnrollLimit * 4;
590 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
591 if (trip_count > unroll_limit || body_size > unroll_limit) {
592 return false;
593 }
594
595 // Fully unroll a loop with few iterations regardless next
596 // conditions since following loop optimizations will split
597 // such loop anyway (pre-main-post).
598 if (trip_count <= 3)
599 return true;
600
601 // Take into account that after unroll conjoined heads and tails will fold,
602 // otherwise policy_unroll() may allow more unrolling than max unrolling.
603 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
604 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
605 if (body_size != tst_body_size) // Check for int overflow
606 return false;
607 if (new_body_size > unroll_limit ||
608 // Unrolling can result in a large amount of node construction
609 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) {
610 return false;
611 }
612
613 // Do not unroll a loop with String intrinsics code.
614 // String intrinsics are large and have loops.
615 for (uint k = 0; k < _body.size(); k++) {
616 Node* n = _body.at(k);
617 switch (n->Opcode()) {
618 case Op_StrComp:
619 case Op_StrEquals:
620 case Op_StrIndexOf:
621 case Op_EncodeISOArray:
622 case Op_AryEq: {
623 return false;
624 }
625 #if INCLUDE_RTM_OPT
626 case Op_FastLock:
627 case Op_FastUnlock: {
628 // Don't unroll RTM locking code because it is large.
629 if (UseRTMLocking) {
630 return false;
631 }
632 }
633 #endif
634 } // switch
635 }
636
637 return true; // Do maximally unroll
638 }
639
640
641 //------------------------------policy_unroll----------------------------------
642 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
643 // the loop is a CountedLoop and the body is small enough.
644 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
645
646 CountedLoopNode *cl = _head->as_CountedLoop();
647 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
648
649 if (!cl->is_valid_counted_loop())
650 return false; // Malformed counted loop
651
652 // Protect against over-unrolling.
653 // After split at least one iteration will be executed in pre-loop.
654 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
655
656 _local_loop_unroll_limit = LoopUnrollLimit;
657 _local_loop_unroll_factor = 4;
658 int future_unroll_ct = cl->unrolled_count() * 2;
659 if (future_unroll_ct > LoopMaxUnroll) return false;
660
661 // Check for initial stride being a small enough constant
662 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
663
664 // Don't unroll if the next round of unrolling would push us
665 // over the expected trip count of the loop. One is subtracted
666 // from the expected trip count because the pre-loop normally
667 // executes 1 iteration.
668 if (UnrollLimitForProfileCheck > 0 &&
669 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
670 future_unroll_ct > UnrollLimitForProfileCheck &&
671 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
672 return false;
673 }
674
675 // When unroll count is greater than LoopUnrollMin, don't unroll if:
676 // the residual iterations are more than 10% of the trip count
677 // and rounds of "unroll,optimize" are not making significant progress
678 // Progress defined as current size less than 20% larger than previous size.
679 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
680 future_unroll_ct > LoopUnrollMin &&
681 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
682 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
683 return false;
684 }
685
686 Node *init_n = cl->init_trip();
687 Node *limit_n = cl->limit();
688 int stride_con = cl->stride_con();
689 // Non-constant bounds.
690 // Protect against over-unrolling when init or/and limit are not constant
691 // (so that trip_count's init value is maxint) but iv range is known.
692 if (init_n == NULL || !init_n->is_Con() ||
693 limit_n == NULL || !limit_n->is_Con()) {
694 Node* phi = cl->phi();
695 if (phi != NULL) {
696 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
697 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
698 int next_stride = stride_con * 2; // stride after this unroll
699 if (next_stride > 0) {
700 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
701 iv_type->_lo + next_stride > iv_type->_hi) {
702 return false; // over-unrolling
703 }
704 } else if (next_stride < 0) {
705 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
706 iv_type->_hi + next_stride < iv_type->_lo) {
707 return false; // over-unrolling
708 }
709 }
710 }
711 }
712
713 // After unroll limit will be adjusted: new_limit = limit-stride.
714 // Bailout if adjustment overflow.
715 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
716 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
717 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
718 return false; // overflow
719
720 // Adjust body_size to determine if we unroll or not
721 uint body_size = _body.size();
722 // Key test to unroll loop in CRC32 java code
723 int xors_in_loop = 0;
724 // Also count ModL, DivL and MulL which expand mightly
725 for (uint k = 0; k < _body.size(); k++) {
726 Node* n = _body.at(k);
727 switch (n->Opcode()) {
728 case Op_XorI: xors_in_loop++; break; // CRC32 java code
729 case Op_ModL: body_size += 30; break;
730 case Op_DivL: body_size += 30; break;
731 case Op_MulL: body_size += 10; break;
732 case Op_StrComp:
733 case Op_StrEquals:
734 case Op_StrIndexOf:
735 case Op_EncodeISOArray:
736 case Op_AryEq: {
737 // Do not unroll a loop with String intrinsics code.
738 // String intrinsics are large and have loops.
739 return false;
740 }
741 #if INCLUDE_RTM_OPT
742 case Op_FastLock:
743 case Op_FastUnlock: {
744 // Don't unroll RTM locking code because it is large.
745 if (UseRTMLocking) {
746 return false;
747 }
748 }
749 #endif
750 } // switch
751 }
752
753 if (UseSuperWord) {
754 if (!cl->is_reduction_loop()) {
755 phase->mark_reductions(this);
756 }
757
758 // Only attempt slp analysis when user controls do not prohibit it
759 if (LoopMaxUnroll > _local_loop_unroll_factor) {
760 // Once policy_slp_analysis succeeds, mark the loop with the
761 // maximal unroll factor so that we minimize analysis passes
762 if ((future_unroll_ct > _local_loop_unroll_factor) ||
763 (body_size > (uint)_local_loop_unroll_limit)) {
764 policy_unroll_slp_analysis(cl, phase, future_unroll_ct);
765 }
766 }
767 }
768
769 // Check for being too big
770 if (body_size > (uint)_local_loop_unroll_limit) {
771 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
772 // Normal case: loop too big
773 return false;
774 }
775
776 // Unroll once! (Each trip will soon do double iterations)
777 return true;
778 }
779
780 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) {
781 // Enable this functionality target by target as needed
782 if (SuperWordLoopUnrollAnalysis) {
783 if (!cl->has_passed_slp()) {
784 SuperWord sw(phase);
785 sw.transform_loop(this, false);
786
787 // If the loop is slp canonical analyze it
788 if (sw.early_return() == false) {
789 sw.unrolling_analysis(cl, _local_loop_unroll_factor);
790 }
791 }
792
793 int slp_max_unroll_factor = cl->slp_max_unroll();
794 if ((slp_max_unroll_factor > 4) &&
795 (slp_max_unroll_factor >= future_unroll_ct)) {
796 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor;
797 if (new_limit > LoopUnrollLimit) {
798 #ifndef PRODUCT
799 if (TraceSuperWordLoopUnrollAnalysis) {
800 tty->print_cr("slp analysis is applying unroll limit %d, the original limit was %d\n",
801 new_limit, _local_loop_unroll_limit);
802 }
803 #endif
804 _local_loop_unroll_limit = new_limit;
805 }
806 }
807 }
808 }
809
810 //------------------------------policy_align-----------------------------------
811 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
812 // expression that does the alignment. Note that only one array base can be
813 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
814 // if we vectorize short memory ops into longer memory ops, we may want to
815 // increase alignment.
816 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
817 return false;
818 }
819
820 //------------------------------policy_range_check-----------------------------
821 // Return TRUE or FALSE if the loop should be range-check-eliminated.
822 // Actually we do iteration-splitting, a more powerful form of RCE.
823 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
824 if (!RangeCheckElimination) return false;
825
826 CountedLoopNode *cl = _head->as_CountedLoop();
827 // If we unrolled with no intention of doing RCE and we later
828 // changed our minds, we got no pre-loop. Either we need to
829 // make a new pre-loop, or we gotta disallow RCE.
830 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
831 Node *trip_counter = cl->phi();
832
833 // Check loop body for tests of trip-counter plus loop-invariant vs
834 // loop-invariant.
835 for (uint i = 0; i < _body.size(); i++) {
836 Node *iff = _body[i];
837 if (iff->Opcode() == Op_If) { // Test?
838
839 // Comparing trip+off vs limit
840 Node *bol = iff->in(1);
841 if (bol->req() != 2) continue; // dead constant test
842 if (!bol->is_Bool()) {
843 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
844 continue;
845 }
846 if (bol->as_Bool()->_test._test == BoolTest::ne)
847 continue; // not RC
848
849 Node *cmp = bol->in(1);
850 Node *rc_exp = cmp->in(1);
851 Node *limit = cmp->in(2);
852
853 Node *limit_c = phase->get_ctrl(limit);
854 if( limit_c == phase->C->top() )
855 return false; // Found dead test on live IF? No RCE!
856 if( is_member(phase->get_loop(limit_c) ) ) {
857 // Compare might have operands swapped; commute them
858 rc_exp = cmp->in(2);
859 limit = cmp->in(1);
860 limit_c = phase->get_ctrl(limit);
861 if( is_member(phase->get_loop(limit_c) ) )
862 continue; // Both inputs are loop varying; cannot RCE
863 }
864
865 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
866 continue;
867 }
868 // Yeah! Found a test like 'trip+off vs limit'
869 // Test is an IfNode, has 2 projections. If BOTH are in the loop
870 // we need loop unswitching instead of iteration splitting.
871 if( is_loop_exit(iff) )
872 return true; // Found reason to split iterations
873 } // End of is IF
874 }
875
876 return false;
877 }
878
879 //------------------------------policy_peel_only-------------------------------
880 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
881 // for unrolling loops with NO array accesses.
882 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
883
884 for( uint i = 0; i < _body.size(); i++ )
885 if( _body[i]->is_Mem() )
886 return false;
887
888 // No memory accesses at all!
889 return true;
890 }
891
892 //------------------------------clone_up_backedge_goo--------------------------
893 // If Node n lives in the back_ctrl block and cannot float, we clone a private
894 // version of n in preheader_ctrl block and return that, otherwise return n.
895 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
896 if( get_ctrl(n) != back_ctrl ) return n;
897
898 // Only visit once
899 if (visited.test_set(n->_idx)) {
900 Node *x = clones.find(n->_idx);
901 if (x != NULL)
902 return x;
903 return n;
904 }
905
906 Node *x = NULL; // If required, a clone of 'n'
907 // Check for 'n' being pinned in the backedge.
908 if( n->in(0) && n->in(0) == back_ctrl ) {
909 assert(clones.find(n->_idx) == NULL, "dead loop");
910 x = n->clone(); // Clone a copy of 'n' to preheader
911 clones.push(x, n->_idx);
912 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
913 }
914
915 // Recursive fixup any other input edges into x.
916 // If there are no changes we can just return 'n', otherwise
917 // we need to clone a private copy and change it.
918 for( uint i = 1; i < n->req(); i++ ) {
919 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
920 if( g != n->in(i) ) {
921 if( !x ) {
922 assert(clones.find(n->_idx) == NULL, "dead loop");
923 x = n->clone();
924 clones.push(x, n->_idx);
925 }
926 x->set_req(i, g);
927 }
928 }
929 if( x ) { // x can legally float to pre-header location
930 register_new_node( x, preheader_ctrl );
931 return x;
932 } else { // raise n to cover LCA of uses
933 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
934 }
935 return n;
936 }
937
938 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) {
939 Node* castii = new CastIINode(incr, TypeInt::INT, true);
940 castii->set_req(0, ctrl);
941 register_new_node(castii, ctrl);
942 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) {
943 Node* n = incr->fast_out(i);
944 if (n->is_Phi() && n->in(0) == loop) {
945 int nrep = n->replace_edge(incr, castii);
946 return true;
947 }
948 }
949 return false;
950 }
951
952 //------------------------------insert_pre_post_loops--------------------------
953 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
954 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
955 // alignment. Useful to unroll loops that do no array accesses.
956 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
957
958 #ifndef PRODUCT
959 if (TraceLoopOpts) {
960 if (peel_only)
961 tty->print("PeelMainPost ");
962 else
963 tty->print("PreMainPost ");
964 loop->dump_head();
965 }
966 #endif
967 C->set_major_progress();
968
969 // Find common pieces of the loop being guarded with pre & post loops
970 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
971 assert( main_head->is_normal_loop(), "" );
972 CountedLoopEndNode *main_end = main_head->loopexit();
973 guarantee(main_end != NULL, "no loop exit node");
974 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
975 uint dd_main_head = dom_depth(main_head);
976 uint max = main_head->outcnt();
977
978 Node *pre_header= main_head->in(LoopNode::EntryControl);
979 Node *init = main_head->init_trip();
980 Node *incr = main_end ->incr();
981 Node *limit = main_end ->limit();
982 Node *stride = main_end ->stride();
983 Node *cmp = main_end ->cmp_node();
984 BoolTest::mask b_test = main_end->test_trip();
985
986 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
987 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
988 if( bol->outcnt() != 1 ) {
989 bol = bol->clone();
990 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
991 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol);
992 }
993 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
994 if( cmp->outcnt() != 1 ) {
995 cmp = cmp->clone();
996 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
997 _igvn.replace_input_of(bol, 1, cmp);
998 }
999
1000 //------------------------------
1001 // Step A: Create Post-Loop.
1002 Node* main_exit = main_end->proj_out(false);
1003 assert( main_exit->Opcode() == Op_IfFalse, "" );
1004 int dd_main_exit = dom_depth(main_exit);
1005
1006 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
1007 // loop pre-header illegally has 2 control users (old & new loops).
1008 clone_loop( loop, old_new, dd_main_exit );
1009 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
1010 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
1011 post_head->set_post_loop(main_head);
1012
1013 // Reduce the post-loop trip count.
1014 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1015 post_end->_prob = PROB_FAIR;
1016
1017 // Build the main-loop normal exit.
1018 IfFalseNode *new_main_exit = new IfFalseNode(main_end);
1019 _igvn.register_new_node_with_optimizer( new_main_exit );
1020 set_idom(new_main_exit, main_end, dd_main_exit );
1021 set_loop(new_main_exit, loop->_parent);
1022
1023 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
1024 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
1025 // (the main-loop trip-counter exit value) because we will be changing
1026 // the exit value (via unrolling) so we cannot constant-fold away the zero
1027 // trip guard until all unrolling is done.
1028 Node *zer_opaq = new Opaque1Node(C, incr);
1029 Node *zer_cmp = new CmpINode( zer_opaq, limit );
1030 Node *zer_bol = new BoolNode( zer_cmp, b_test );
1031 register_new_node( zer_opaq, new_main_exit );
1032 register_new_node( zer_cmp , new_main_exit );
1033 register_new_node( zer_bol , new_main_exit );
1034
1035 // Build the IfNode
1036 IfNode *zer_iff = new IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
1037 _igvn.register_new_node_with_optimizer( zer_iff );
1038 set_idom(zer_iff, new_main_exit, dd_main_exit);
1039 set_loop(zer_iff, loop->_parent);
1040
1041 // Plug in the false-path, taken if we need to skip post-loop
1042 _igvn.replace_input_of(main_exit, 0, zer_iff);
1043 set_idom(main_exit, zer_iff, dd_main_exit);
1044 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
1045 // Make the true-path, must enter the post loop
1046 Node *zer_taken = new IfTrueNode( zer_iff );
1047 _igvn.register_new_node_with_optimizer( zer_taken );
1048 set_idom(zer_taken, zer_iff, dd_main_exit);
1049 set_loop(zer_taken, loop->_parent);
1050 // Plug in the true path
1051 _igvn.hash_delete( post_head );
1052 post_head->set_req(LoopNode::EntryControl, zer_taken);
1053 set_idom(post_head, zer_taken, dd_main_exit);
1054
1055 Arena *a = Thread::current()->resource_area();
1056 VectorSet visited(a);
1057 Node_Stack clones(a, main_head->back_control()->outcnt());
1058 // Step A3: Make the fall-in values to the post-loop come from the
1059 // fall-out values of the main-loop.
1060 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
1061 Node* main_phi = main_head->fast_out(i);
1062 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
1063 Node *post_phi = old_new[main_phi->_idx];
1064 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
1065 post_head->init_control(),
1066 main_phi->in(LoopNode::LoopBackControl),
1067 visited, clones);
1068 _igvn.hash_delete(post_phi);
1069 post_phi->set_req( LoopNode::EntryControl, fallmain );
1070 }
1071 }
1072
1073 // Update local caches for next stanza
1074 main_exit = new_main_exit;
1075
1076
1077 //------------------------------
1078 // Step B: Create Pre-Loop.
1079
1080 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1081 // loop pre-header illegally has 2 control users (old & new loops).
1082 clone_loop( loop, old_new, dd_main_head );
1083 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1084 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1085 pre_head->set_pre_loop(main_head);
1086 Node *pre_incr = old_new[incr->_idx];
1087
1088 // Reduce the pre-loop trip count.
1089 pre_end->_prob = PROB_FAIR;
1090
1091 // Find the pre-loop normal exit.
1092 Node* pre_exit = pre_end->proj_out(false);
1093 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1094 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end);
1095 _igvn.register_new_node_with_optimizer( new_pre_exit );
1096 set_idom(new_pre_exit, pre_end, dd_main_head);
1097 set_loop(new_pre_exit, loop->_parent);
1098
1099 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1100 // pre-loop, the main-loop may not execute at all. Later in life this
1101 // zero-trip guard will become the minimum-trip guard when we unroll
1102 // the main-loop.
1103 Node *min_opaq = new Opaque1Node(C, limit);
1104 Node *min_cmp = new CmpINode( pre_incr, min_opaq );
1105 Node *min_bol = new BoolNode( min_cmp, b_test );
1106 register_new_node( min_opaq, new_pre_exit );
1107 register_new_node( min_cmp , new_pre_exit );
1108 register_new_node( min_bol , new_pre_exit );
1109
1110 // Build the IfNode (assume the main-loop is executed always).
1111 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1112 _igvn.register_new_node_with_optimizer( min_iff );
1113 set_idom(min_iff, new_pre_exit, dd_main_head);
1114 set_loop(min_iff, loop->_parent);
1115
1116 // Plug in the false-path, taken if we need to skip main-loop
1117 _igvn.hash_delete( pre_exit );
1118 pre_exit->set_req(0, min_iff);
1119 set_idom(pre_exit, min_iff, dd_main_head);
1120 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1121 // Make the true-path, must enter the main loop
1122 Node *min_taken = new IfTrueNode( min_iff );
1123 _igvn.register_new_node_with_optimizer( min_taken );
1124 set_idom(min_taken, min_iff, dd_main_head);
1125 set_loop(min_taken, loop->_parent);
1126 // Plug in the true path
1127 _igvn.hash_delete( main_head );
1128 main_head->set_req(LoopNode::EntryControl, min_taken);
1129 set_idom(main_head, min_taken, dd_main_head);
1130
1131 visited.Clear();
1132 clones.clear();
1133 // Step B3: Make the fall-in values to the main-loop come from the
1134 // fall-out values of the pre-loop.
1135 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1136 Node* main_phi = main_head->fast_out(i2);
1137 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1138 Node *pre_phi = old_new[main_phi->_idx];
1139 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1140 main_head->init_control(),
1141 pre_phi->in(LoopNode::LoopBackControl),
1142 visited, clones);
1143 _igvn.hash_delete(main_phi);
1144 main_phi->set_req( LoopNode::EntryControl, fallpre );
1145 }
1146 }
1147
1148 // Nodes inside the loop may be control dependent on a predicate
1149 // that was moved before the preloop. If the back branch of the main
1150 // or post loops becomes dead, those nodes won't be dependent on the
1151 // test that guards that loop nest anymore which could lead to an
1152 // incorrect array access because it executes independently of the
1153 // test that was guarding the loop nest. We add a special CastII on
1154 // the if branch that enters the loop, between the input induction
1155 // variable value and the induction variable Phi to preserve correct
1156 // dependencies.
1157
1158 // CastII for the post loop:
1159 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1160 assert(inserted, "no castII inserted");
1161
1162 // CastII for the main loop:
1163 inserted = cast_incr_before_loop(pre_incr, min_taken, main_head);
1164 assert(inserted, "no castII inserted");
1165
1166 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1167 // RCE and alignment may change this later.
1168 Node *cmp_end = pre_end->cmp_node();
1169 assert( cmp_end->in(2) == limit, "" );
1170 Node *pre_limit = new AddINode( init, stride );
1171
1172 // Save the original loop limit in this Opaque1 node for
1173 // use by range check elimination.
1174 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit);
1175
1176 register_new_node( pre_limit, pre_head->in(0) );
1177 register_new_node( pre_opaq , pre_head->in(0) );
1178
1179 // Since no other users of pre-loop compare, I can hack limit directly
1180 assert( cmp_end->outcnt() == 1, "no other users" );
1181 _igvn.hash_delete(cmp_end);
1182 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1183
1184 // Special case for not-equal loop bounds:
1185 // Change pre loop test, main loop test, and the
1186 // main loop guard test to use lt or gt depending on stride
1187 // direction:
1188 // positive stride use <
1189 // negative stride use >
1190 //
1191 // not-equal test is kept for post loop to handle case
1192 // when init > limit when stride > 0 (and reverse).
1193
1194 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1195
1196 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1197 // Modify pre loop end condition
1198 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1199 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test);
1200 register_new_node( new_bol0, pre_head->in(0) );
1201 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0);
1202 // Modify main loop guard condition
1203 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1204 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test);
1205 register_new_node( new_bol1, new_pre_exit );
1206 _igvn.hash_delete(min_iff);
1207 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1208 // Modify main loop end condition
1209 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1210 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test);
1211 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1212 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2);
1213 }
1214
1215 // Flag main loop
1216 main_head->set_main_loop();
1217 if( peel_only ) main_head->set_main_no_pre_loop();
1218
1219 // Subtract a trip count for the pre-loop.
1220 main_head->set_trip_count(main_head->trip_count() - 1);
1221
1222 // It's difficult to be precise about the trip-counts
1223 // for the pre/post loops. They are usually very short,
1224 // so guess that 4 trips is a reasonable value.
1225 post_head->set_profile_trip_cnt(4.0);
1226 pre_head->set_profile_trip_cnt(4.0);
1227
1228 // Now force out all loop-invariant dominating tests. The optimizer
1229 // finds some, but we _know_ they are all useless.
1230 peeled_dom_test_elim(loop,old_new);
1231 loop->record_for_igvn();
1232 }
1233
1234 //------------------------------is_invariant-----------------------------
1235 // Return true if n is invariant
1236 bool IdealLoopTree::is_invariant(Node* n) const {
1237 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1238 if (n_c->is_top()) return false;
1239 return !is_member(_phase->get_loop(n_c));
1240 }
1241
1242
1243 //------------------------------do_unroll--------------------------------------
1244 // Unroll the loop body one step - make each trip do 2 iterations.
1245 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1246 assert(LoopUnrollLimit, "");
1247 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1248 CountedLoopEndNode *loop_end = loop_head->loopexit();
1249 assert(loop_end, "");
1250 #ifndef PRODUCT
1251 if (PrintOpto && VerifyLoopOptimizations) {
1252 tty->print("Unrolling ");
1253 loop->dump_head();
1254 } else if (TraceLoopOpts) {
1255 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1256 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1257 } else {
1258 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1259 }
1260 loop->dump_head();
1261 }
1262
1263 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) {
1264 Arena* arena = Thread::current()->resource_area();
1265 Node_Stack stack(arena, C->unique() >> 2);
1266 Node_List rpo_list;
1267 VectorSet visited(arena);
1268 visited.set(loop_head->_idx);
1269 rpo( loop_head, stack, visited, rpo_list );
1270 dump(loop, rpo_list.size(), rpo_list );
1271 }
1272 #endif
1273
1274 // Remember loop node count before unrolling to detect
1275 // if rounds of unroll,optimize are making progress
1276 loop_head->set_node_count_before_unroll(loop->_body.size());
1277
1278 Node *ctrl = loop_head->in(LoopNode::EntryControl);
1279 Node *limit = loop_head->limit();
1280 Node *init = loop_head->init_trip();
1281 Node *stride = loop_head->stride();
1282
1283 Node *opaq = NULL;
1284 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1285 // Search for zero-trip guard.
1286 assert( loop_head->is_main_loop(), "" );
1287 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1288 Node *iff = ctrl->in(0);
1289 assert( iff->Opcode() == Op_If, "" );
1290 Node *bol = iff->in(1);
1291 assert( bol->Opcode() == Op_Bool, "" );
1292 Node *cmp = bol->in(1);
1293 assert( cmp->Opcode() == Op_CmpI, "" );
1294 opaq = cmp->in(2);
1295 // Occasionally it's possible for a zero-trip guard Opaque1 node to be
1296 // optimized away and then another round of loop opts attempted.
1297 // We can not optimize this particular loop in that case.
1298 if (opaq->Opcode() != Op_Opaque1)
1299 return; // Cannot find zero-trip guard! Bail out!
1300 // Zero-trip test uses an 'opaque' node which is not shared.
1301 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1302 }
1303
1304 C->set_major_progress();
1305
1306 Node* new_limit = NULL;
1307 if (UnrollLimitCheck) {
1308 int stride_con = stride->get_int();
1309 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1310 uint old_trip_count = loop_head->trip_count();
1311 // Verify that unroll policy result is still valid.
1312 assert(old_trip_count > 1 &&
1313 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1314
1315 // Adjust loop limit to keep valid iterations number after unroll.
1316 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1317 // which may overflow.
1318 if (!adjust_min_trip) {
1319 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1320 "odd trip count for maximally unroll");
1321 // Don't need to adjust limit for maximally unroll since trip count is even.
1322 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1323 // Loop's limit is constant. Loop's init could be constant when pre-loop
1324 // become peeled iteration.
1325 jlong init_con = init->get_int();
1326 // We can keep old loop limit if iterations count stays the same:
1327 // old_trip_count == new_trip_count * 2
1328 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1329 // so we also don't need to adjust zero trip test.
1330 jlong limit_con = limit->get_int();
1331 // (stride_con*2) not overflow since stride_con <= 8.
1332 int new_stride_con = stride_con * 2;
1333 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1334 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1335 // New trip count should satisfy next conditions.
1336 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1337 uint new_trip_count = (uint)trip_count;
1338 adjust_min_trip = (old_trip_count != new_trip_count*2);
1339 }
1340
1341 if (adjust_min_trip) {
1342 // Step 2: Adjust the trip limit if it is called for.
1343 // The adjustment amount is -stride. Need to make sure if the
1344 // adjustment underflows or overflows, then the main loop is skipped.
1345 Node* cmp = loop_end->cmp_node();
1346 assert(cmp->in(2) == limit, "sanity");
1347 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1348
1349 // Verify that policy_unroll result is still valid.
1350 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1351 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1352 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1353
1354 if (limit->is_Con()) {
1355 // The check in policy_unroll and the assert above guarantee
1356 // no underflow if limit is constant.
1357 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1358 set_ctrl(new_limit, C->root());
1359 } else {
1360 // Limit is not constant.
1361 if (loop_head->unrolled_count() == 1) { // only for first unroll
1362 // Separate limit by Opaque node in case it is an incremented
1363 // variable from previous loop to avoid using pre-incremented
1364 // value which could increase register pressure.
1365 // Otherwise reorg_offsets() optimization will create a separate
1366 // Opaque node for each use of trip-counter and as result
1367 // zero trip guard limit will be different from loop limit.
1368 assert(has_ctrl(opaq), "should have it");
1369 Node* opaq_ctrl = get_ctrl(opaq);
1370 limit = new Opaque2Node( C, limit );
1371 register_new_node( limit, opaq_ctrl );
1372 }
1373 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) ||
1374 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) {
1375 // No underflow.
1376 new_limit = new SubINode(limit, stride);
1377 } else {
1378 // (limit - stride) may underflow.
1379 // Clamp the adjustment value with MININT or MAXINT:
1380 //
1381 // new_limit = limit-stride
1382 // if (stride > 0)
1383 // new_limit = (limit < new_limit) ? MININT : new_limit;
1384 // else
1385 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1386 //
1387 BoolTest::mask bt = loop_end->test_trip();
1388 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1389 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1390 set_ctrl(adj_max, C->root());
1391 Node* old_limit = NULL;
1392 Node* adj_limit = NULL;
1393 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1394 if (loop_head->unrolled_count() > 1 &&
1395 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1396 limit->in(CMoveNode::IfTrue) == adj_max &&
1397 bol->as_Bool()->_test._test == bt &&
1398 bol->in(1)->Opcode() == Op_CmpI &&
1399 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1400 // Loop was unrolled before.
1401 // Optimize the limit to avoid nested CMove:
1402 // use original limit as old limit.
1403 old_limit = bol->in(1)->in(1);
1404 // Adjust previous adjusted limit.
1405 adj_limit = limit->in(CMoveNode::IfFalse);
1406 adj_limit = new SubINode(adj_limit, stride);
1407 } else {
1408 old_limit = limit;
1409 adj_limit = new SubINode(limit, stride);
1410 }
1411 assert(old_limit != NULL && adj_limit != NULL, "");
1412 register_new_node( adj_limit, ctrl ); // adjust amount
1413 Node* adj_cmp = new CmpINode(old_limit, adj_limit);
1414 register_new_node( adj_cmp, ctrl );
1415 Node* adj_bool = new BoolNode(adj_cmp, bt);
1416 register_new_node( adj_bool, ctrl );
1417 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1418 }
1419 register_new_node(new_limit, ctrl);
1420 }
1421 assert(new_limit != NULL, "");
1422 // Replace in loop test.
1423 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1424 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1425 // Don't need to create new test since only one user.
1426 _igvn.hash_delete(cmp);
1427 cmp->set_req(2, new_limit);
1428 } else {
1429 // Create new test since it is shared.
1430 Node* ctrl2 = loop_end->in(0);
1431 Node* cmp2 = cmp->clone();
1432 cmp2->set_req(2, new_limit);
1433 register_new_node(cmp2, ctrl2);
1434 Node* bol2 = loop_end->in(1)->clone();
1435 bol2->set_req(1, cmp2);
1436 register_new_node(bol2, ctrl2);
1437 _igvn.replace_input_of(loop_end, 1, bol2);
1438 }
1439 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1440 // Make it a 1-trip test (means at least 2 trips).
1441
1442 // Guard test uses an 'opaque' node which is not shared. Hence I
1443 // can edit it's inputs directly. Hammer in the new limit for the
1444 // minimum-trip guard.
1445 assert(opaq->outcnt() == 1, "");
1446 _igvn.replace_input_of(opaq, 1, new_limit);
1447 }
1448
1449 // Adjust max trip count. The trip count is intentionally rounded
1450 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1451 // the main, unrolled, part of the loop will never execute as it is protected
1452 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1453 // and later determined that part of the unrolled loop was dead.
1454 loop_head->set_trip_count(old_trip_count / 2);
1455
1456 // Double the count of original iterations in the unrolled loop body.
1457 loop_head->double_unrolled_count();
1458
1459 } else { // LoopLimitCheck
1460
1461 // Adjust max trip count. The trip count is intentionally rounded
1462 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1463 // the main, unrolled, part of the loop will never execute as it is protected
1464 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1465 // and later determined that part of the unrolled loop was dead.
1466 loop_head->set_trip_count(loop_head->trip_count() / 2);
1467
1468 // Double the count of original iterations in the unrolled loop body.
1469 loop_head->double_unrolled_count();
1470
1471 // -----------
1472 // Step 2: Cut back the trip counter for an unroll amount of 2.
1473 // Loop will normally trip (limit - init)/stride_con. Since it's a
1474 // CountedLoop this is exact (stride divides limit-init exactly).
1475 // We are going to double the loop body, so we want to knock off any
1476 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
1477 Node *span = new SubINode( limit, init );
1478 register_new_node( span, ctrl );
1479 Node *trip = new DivINode( 0, span, stride );
1480 register_new_node( trip, ctrl );
1481 Node *mtwo = _igvn.intcon(-2);
1482 set_ctrl(mtwo, C->root());
1483 Node *rond = new AndINode( trip, mtwo );
1484 register_new_node( rond, ctrl );
1485 Node *spn2 = new MulINode( rond, stride );
1486 register_new_node( spn2, ctrl );
1487 new_limit = new AddINode( spn2, init );
1488 register_new_node( new_limit, ctrl );
1489
1490 // Hammer in the new limit
1491 Node *ctrl2 = loop_end->in(0);
1492 Node *cmp2 = new CmpINode( loop_head->incr(), new_limit );
1493 register_new_node( cmp2, ctrl2 );
1494 Node *bol2 = new BoolNode( cmp2, loop_end->test_trip() );
1495 register_new_node( bol2, ctrl2 );
1496 _igvn.replace_input_of(loop_end, CountedLoopEndNode::TestValue, bol2);
1497
1498 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1499 // Make it a 1-trip test (means at least 2 trips).
1500 if( adjust_min_trip ) {
1501 assert( new_limit != NULL, "" );
1502 // Guard test uses an 'opaque' node which is not shared. Hence I
1503 // can edit it's inputs directly. Hammer in the new limit for the
1504 // minimum-trip guard.
1505 assert( opaq->outcnt() == 1, "" );
1506 _igvn.hash_delete(opaq);
1507 opaq->set_req(1, new_limit);
1508 }
1509 } // LoopLimitCheck
1510
1511 // ---------
1512 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1513 // represents the odd iterations; since the loop trips an even number of
1514 // times its backedge is never taken. Kill the backedge.
1515 uint dd = dom_depth(loop_head);
1516 clone_loop( loop, old_new, dd );
1517
1518 // Make backedges of the clone equal to backedges of the original.
1519 // Make the fall-in from the original come from the fall-out of the clone.
1520 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1521 Node* phi = loop_head->fast_out(j);
1522 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1523 Node *newphi = old_new[phi->_idx];
1524 _igvn.hash_delete( phi );
1525 _igvn.hash_delete( newphi );
1526
1527 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1528 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1529 phi ->set_req(LoopNode::LoopBackControl, C->top());
1530 }
1531 }
1532 Node *clone_head = old_new[loop_head->_idx];
1533 _igvn.hash_delete( clone_head );
1534 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1535 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1536 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1537 loop->_head = clone_head; // New loop header
1538
1539 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1540 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1541
1542 // Kill the clone's backedge
1543 Node *newcle = old_new[loop_end->_idx];
1544 _igvn.hash_delete( newcle );
1545 Node *one = _igvn.intcon(1);
1546 set_ctrl(one, C->root());
1547 newcle->set_req(1, one);
1548 // Force clone into same loop body
1549 uint max = loop->_body.size();
1550 for( uint k = 0; k < max; k++ ) {
1551 Node *old = loop->_body.at(k);
1552 Node *nnn = old_new[old->_idx];
1553 loop->_body.push(nnn);
1554 if (!has_ctrl(old))
1555 set_loop(nnn, loop);
1556 }
1557
1558 loop->record_for_igvn();
1559
1560 #ifndef PRODUCT
1561 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) {
1562 tty->print("\nnew loop after unroll\n"); loop->dump_head();
1563 for (uint i = 0; i < loop->_body.size(); i++) {
1564 loop->_body.at(i)->dump();
1565 }
1566 if(C->clone_map().is_debug()) {
1567 tty->print("\nCloneMap\n");
1568 Dict* dict = C->clone_map().dict();
1569 DictI i(dict);
1570 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size());
1571 for (int ii = 0; i.test(); ++i, ++ii) {
1572 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key));
1573 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen());
1574 if (ii % 10 == 9) {
1575 tty->print_cr(" ");
1576 }
1577 }
1578 tty->print_cr(" ");
1579 }
1580 }
1581 #endif
1582
1583 }
1584
1585 //------------------------------do_maximally_unroll----------------------------
1586
1587 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1588 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1589 assert(cl->has_exact_trip_count(), "trip count is not exact");
1590 assert(cl->trip_count() > 0, "");
1591 #ifndef PRODUCT
1592 if (TraceLoopOpts) {
1593 tty->print("MaxUnroll %d ", cl->trip_count());
1594 loop->dump_head();
1595 }
1596 #endif
1597
1598 // If loop is tripping an odd number of times, peel odd iteration
1599 if ((cl->trip_count() & 1) == 1) {
1600 do_peeling(loop, old_new);
1601 }
1602
1603 // Now its tripping an even number of times remaining. Double loop body.
1604 // Do not adjust pre-guards; they are not needed and do not exist.
1605 if (cl->trip_count() > 0) {
1606 assert((cl->trip_count() & 1) == 0, "missed peeling");
1607 do_unroll(loop, old_new, false);
1608 }
1609 }
1610
1611 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) {
1612 if (SuperWordReductions == false) return;
1613
1614 CountedLoopNode* loop_head = loop->_head->as_CountedLoop();
1615 if (loop_head->unrolled_count() > 1) {
1616 return;
1617 }
1618
1619 Node* trip_phi = loop_head->phi();
1620 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) {
1621 Node* phi = loop_head->fast_out(i);
1622 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) {
1623 // For definitions which are loop inclusive and not tripcounts.
1624 Node* def_node = phi->in(LoopNode::LoopBackControl);
1625
1626 if (def_node != NULL) {
1627 Node* n_ctrl = get_ctrl(def_node);
1628 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) {
1629 // Now test it to see if it fits the standard pattern for a reduction operator.
1630 int opc = def_node->Opcode();
1631 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) {
1632 if (!def_node->is_reduction()) { // Not marked yet
1633 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it
1634 bool ok = false;
1635 for (unsigned j = 1; j < def_node->req(); j++) {
1636 Node* in = def_node->in(j);
1637 if (in == phi) {
1638 ok = true;
1639 break;
1640 }
1641 }
1642
1643 // do nothing if we did not match the initial criteria
1644 if (ok == false) {
1645 continue;
1646 }
1647
1648 // The result of the reduction must not be used in the loop
1649 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) {
1650 Node* u = def_node->fast_out(i);
1651 if (has_ctrl(u) && !loop->is_member(get_loop(get_ctrl(u)))) {
1652 continue;
1653 }
1654 if (u == phi) {
1655 continue;
1656 }
1657 ok = false;
1658 }
1659
1660 // iff the uses conform
1661 if (ok) {
1662 def_node->add_flag(Node::Flag_is_reduction);
1663 loop_head->mark_has_reductions();
1664 }
1665 }
1666 }
1667 }
1668 }
1669 }
1670 }
1671 }
1672
1673 //------------------------------dominates_backedge---------------------------------
1674 // Returns true if ctrl is executed on every complete iteration
1675 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1676 assert(ctrl->is_CFG(), "must be control");
1677 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1678 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1679 }
1680
1681 //------------------------------adjust_limit-----------------------------------
1682 // Helper function for add_constraint().
1683 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1684 // Compute "I :: (limit-offset)/scale"
1685 Node *con = new SubINode(rc_limit, offset);
1686 register_new_node(con, pre_ctrl);
1687 Node *X = new DivINode(0, con, scale);
1688 register_new_node(X, pre_ctrl);
1689
1690 // Adjust loop limit
1691 loop_limit = (stride_con > 0)
1692 ? (Node*)(new MinINode(loop_limit, X))
1693 : (Node*)(new MaxINode(loop_limit, X));
1694 register_new_node(loop_limit, pre_ctrl);
1695 return loop_limit;
1696 }
1697
1698 //------------------------------add_constraint---------------------------------
1699 // Constrain the main loop iterations so the conditions:
1700 // low_limit <= scale_con * I + offset < upper_limit
1701 // always holds true. That is, either increase the number of iterations in
1702 // the pre-loop or the post-loop until the condition holds true in the main
1703 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1704 // stride and scale are constants (offset and limit often are).
1705 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 ) {
1706 // For positive stride, the pre-loop limit always uses a MAX function
1707 // and the main loop a MIN function. For negative stride these are
1708 // reversed.
1709
1710 // Also for positive stride*scale the affine function is increasing, so the
1711 // pre-loop must check for underflow and the post-loop for overflow.
1712 // Negative stride*scale reverses this; pre-loop checks for overflow and
1713 // post-loop for underflow.
1714
1715 Node *scale = _igvn.intcon(scale_con);
1716 set_ctrl(scale, C->root());
1717
1718 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1719 // The overflow limit: scale*I+offset < upper_limit
1720 // For main-loop compute
1721 // ( if (scale > 0) /* and stride > 0 */
1722 // I < (upper_limit-offset)/scale
1723 // else /* scale < 0 and stride < 0 */
1724 // I > (upper_limit-offset)/scale
1725 // )
1726 //
1727 // (upper_limit-offset) may overflow or underflow.
1728 // But it is fine since main loop will either have
1729 // less iterations or will be skipped in such case.
1730 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1731
1732 // The underflow limit: low_limit <= scale*I+offset.
1733 // For pre-loop compute
1734 // NOT(scale*I+offset >= low_limit)
1735 // scale*I+offset < low_limit
1736 // ( if (scale > 0) /* and stride > 0 */
1737 // I < (low_limit-offset)/scale
1738 // else /* scale < 0 and stride < 0 */
1739 // I > (low_limit-offset)/scale
1740 // )
1741
1742 if (low_limit->get_int() == -max_jint) {
1743 if (!RangeLimitCheck) return;
1744 // We need this guard when scale*pre_limit+offset >= limit
1745 // due to underflow. So we need execute pre-loop until
1746 // scale*I+offset >= min_int. But (min_int-offset) will
1747 // underflow when offset > 0 and X will be > original_limit
1748 // when stride > 0. To avoid it we replace positive offset with 0.
1749 //
1750 // Also (min_int+1 == -max_int) is used instead of min_int here
1751 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1752 Node* shift = _igvn.intcon(31);
1753 set_ctrl(shift, C->root());
1754 Node* sign = new RShiftINode(offset, shift);
1755 register_new_node(sign, pre_ctrl);
1756 offset = new AndINode(offset, sign);
1757 register_new_node(offset, pre_ctrl);
1758 } else {
1759 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1760 // The only problem we have here when offset == min_int
1761 // since (0-min_int) == min_int. It may be fine for stride > 0
1762 // but for stride < 0 X will be < original_limit. To avoid it
1763 // max(pre_limit, original_limit) is used in do_range_check().
1764 }
1765 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1766 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1767
1768 } else { // stride_con*scale_con < 0
1769 // For negative stride*scale pre-loop checks for overflow and
1770 // post-loop for underflow.
1771 //
1772 // The overflow limit: scale*I+offset < upper_limit
1773 // For pre-loop compute
1774 // NOT(scale*I+offset < upper_limit)
1775 // scale*I+offset >= upper_limit
1776 // scale*I+offset+1 > upper_limit
1777 // ( if (scale < 0) /* and stride > 0 */
1778 // I < (upper_limit-(offset+1))/scale
1779 // else /* scale > 0 and stride < 0 */
1780 // I > (upper_limit-(offset+1))/scale
1781 // )
1782 //
1783 // (upper_limit-offset-1) may underflow or overflow.
1784 // To avoid it min(pre_limit, original_limit) is used
1785 // in do_range_check() for stride > 0 and max() for < 0.
1786 Node *one = _igvn.intcon(1);
1787 set_ctrl(one, C->root());
1788
1789 Node *plus_one = new AddINode(offset, one);
1790 register_new_node( plus_one, pre_ctrl );
1791 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1792 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1793
1794 if (low_limit->get_int() == -max_jint) {
1795 if (!RangeLimitCheck) return;
1796 // We need this guard when scale*main_limit+offset >= limit
1797 // due to underflow. So we need execute main-loop while
1798 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1799 // underflow when (offset+1) > 0 and X will be < main_limit
1800 // when scale < 0 (and stride > 0). To avoid it we replace
1801 // positive (offset+1) with 0.
1802 //
1803 // Also (min_int+1 == -max_int) is used instead of min_int here
1804 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1805 Node* shift = _igvn.intcon(31);
1806 set_ctrl(shift, C->root());
1807 Node* sign = new RShiftINode(plus_one, shift);
1808 register_new_node(sign, pre_ctrl);
1809 plus_one = new AndINode(plus_one, sign);
1810 register_new_node(plus_one, pre_ctrl);
1811 } else {
1812 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1813 // The only problem we have here when offset == max_int
1814 // since (max_int+1) == min_int and (0-min_int) == min_int.
1815 // But it is fine since main loop will either have
1816 // less iterations or will be skipped in such case.
1817 }
1818 // The underflow limit: low_limit <= scale*I+offset.
1819 // For main-loop compute
1820 // scale*I+offset+1 > low_limit
1821 // ( if (scale < 0) /* and stride > 0 */
1822 // I < (low_limit-(offset+1))/scale
1823 // else /* scale > 0 and stride < 0 */
1824 // I > (low_limit-(offset+1))/scale
1825 // )
1826
1827 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1828 }
1829 }
1830
1831
1832 //------------------------------is_scaled_iv---------------------------------
1833 // Return true if exp is a constant times an induction var
1834 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1835 if (exp == iv) {
1836 if (p_scale != NULL) {
1837 *p_scale = 1;
1838 }
1839 return true;
1840 }
1841 int opc = exp->Opcode();
1842 if (opc == Op_MulI) {
1843 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1844 if (p_scale != NULL) {
1845 *p_scale = exp->in(2)->get_int();
1846 }
1847 return true;
1848 }
1849 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1850 if (p_scale != NULL) {
1851 *p_scale = exp->in(1)->get_int();
1852 }
1853 return true;
1854 }
1855 } else if (opc == Op_LShiftI) {
1856 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1857 if (p_scale != NULL) {
1858 *p_scale = 1 << exp->in(2)->get_int();
1859 }
1860 return true;
1861 }
1862 }
1863 return false;
1864 }
1865
1866 //-----------------------------is_scaled_iv_plus_offset------------------------------
1867 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1868 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1869 if (is_scaled_iv(exp, iv, p_scale)) {
1870 if (p_offset != NULL) {
1871 Node *zero = _igvn.intcon(0);
1872 set_ctrl(zero, C->root());
1873 *p_offset = zero;
1874 }
1875 return true;
1876 }
1877 int opc = exp->Opcode();
1878 if (opc == Op_AddI) {
1879 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1880 if (p_offset != NULL) {
1881 *p_offset = exp->in(2);
1882 }
1883 return true;
1884 }
1885 if (exp->in(2)->is_Con()) {
1886 Node* offset2 = NULL;
1887 if (depth < 2 &&
1888 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1889 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1890 if (p_offset != NULL) {
1891 Node *ctrl_off2 = get_ctrl(offset2);
1892 Node* offset = new AddINode(offset2, exp->in(2));
1893 register_new_node(offset, ctrl_off2);
1894 *p_offset = offset;
1895 }
1896 return true;
1897 }
1898 }
1899 } else if (opc == Op_SubI) {
1900 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1901 if (p_offset != NULL) {
1902 Node *zero = _igvn.intcon(0);
1903 set_ctrl(zero, C->root());
1904 Node *ctrl_off = get_ctrl(exp->in(2));
1905 Node* offset = new SubINode(zero, exp->in(2));
1906 register_new_node(offset, ctrl_off);
1907 *p_offset = offset;
1908 }
1909 return true;
1910 }
1911 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1912 if (p_offset != NULL) {
1913 *p_scale *= -1;
1914 *p_offset = exp->in(1);
1915 }
1916 return true;
1917 }
1918 }
1919 return false;
1920 }
1921
1922 //------------------------------do_range_check---------------------------------
1923 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1924 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1925 #ifndef PRODUCT
1926 if (PrintOpto && VerifyLoopOptimizations) {
1927 tty->print("Range Check Elimination ");
1928 loop->dump_head();
1929 } else if (TraceLoopOpts) {
1930 tty->print("RangeCheck ");
1931 loop->dump_head();
1932 }
1933 #endif
1934 assert(RangeCheckElimination, "");
1935 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1936 assert(cl->is_main_loop(), "");
1937
1938 // protect against stride not being a constant
1939 if (!cl->stride_is_con())
1940 return;
1941
1942 // Find the trip counter; we are iteration splitting based on it
1943 Node *trip_counter = cl->phi();
1944 // Find the main loop limit; we will trim it's iterations
1945 // to not ever trip end tests
1946 Node *main_limit = cl->limit();
1947
1948 // Need to find the main-loop zero-trip guard
1949 Node *ctrl = cl->in(LoopNode::EntryControl);
1950 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1951 Node *iffm = ctrl->in(0);
1952 assert(iffm->Opcode() == Op_If, "");
1953 Node *bolzm = iffm->in(1);
1954 assert(bolzm->Opcode() == Op_Bool, "");
1955 Node *cmpzm = bolzm->in(1);
1956 assert(cmpzm->is_Cmp(), "");
1957 Node *opqzm = cmpzm->in(2);
1958 // Can not optimize a loop if zero-trip Opaque1 node is optimized
1959 // away and then another round of loop opts attempted.
1960 if (opqzm->Opcode() != Op_Opaque1)
1961 return;
1962 assert(opqzm->in(1) == main_limit, "do not understand situation");
1963
1964 // Find the pre-loop limit; we will expand its iterations to
1965 // not ever trip low tests.
1966 Node *p_f = iffm->in(0);
1967 // pre loop may have been optimized out
1968 if (p_f->Opcode() != Op_IfFalse) {
1969 return;
1970 }
1971 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1972 assert(pre_end->loopnode()->is_pre_loop(), "");
1973 Node *pre_opaq1 = pre_end->limit();
1974 // Occasionally it's possible for a pre-loop Opaque1 node to be
1975 // optimized away and then another round of loop opts attempted.
1976 // We can not optimize this particular loop in that case.
1977 if (pre_opaq1->Opcode() != Op_Opaque1)
1978 return;
1979 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1980 Node *pre_limit = pre_opaq->in(1);
1981
1982 // Where do we put new limit calculations
1983 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1984
1985 // Ensure the original loop limit is available from the
1986 // pre-loop Opaque1 node.
1987 Node *orig_limit = pre_opaq->original_loop_limit();
1988 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1989 return;
1990
1991 // Must know if its a count-up or count-down loop
1992
1993 int stride_con = cl->stride_con();
1994 Node *zero = _igvn.intcon(0);
1995 Node *one = _igvn.intcon(1);
1996 // Use symmetrical int range [-max_jint,max_jint]
1997 Node *mini = _igvn.intcon(-max_jint);
1998 set_ctrl(zero, C->root());
1999 set_ctrl(one, C->root());
2000 set_ctrl(mini, C->root());
2001
2002 // Range checks that do not dominate the loop backedge (ie.
2003 // conditionally executed) can lengthen the pre loop limit beyond
2004 // the original loop limit. To prevent this, the pre limit is
2005 // (for stride > 0) MINed with the original loop limit (MAXed
2006 // stride < 0) when some range_check (rc) is conditionally
2007 // executed.
2008 bool conditional_rc = false;
2009
2010 // Check loop body for tests of trip-counter plus loop-invariant vs
2011 // loop-invariant.
2012 for( uint i = 0; i < loop->_body.size(); i++ ) {
2013 Node *iff = loop->_body[i];
2014 if( iff->Opcode() == Op_If ) { // Test?
2015
2016 // Test is an IfNode, has 2 projections. If BOTH are in the loop
2017 // we need loop unswitching instead of iteration splitting.
2018 Node *exit = loop->is_loop_exit(iff);
2019 if( !exit ) continue;
2020 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
2021
2022 // Get boolean condition to test
2023 Node *i1 = iff->in(1);
2024 if( !i1->is_Bool() ) continue;
2025 BoolNode *bol = i1->as_Bool();
2026 BoolTest b_test = bol->_test;
2027 // Flip sense of test if exit condition is flipped
2028 if( flip )
2029 b_test = b_test.negate();
2030
2031 // Get compare
2032 Node *cmp = bol->in(1);
2033
2034 // Look for trip_counter + offset vs limit
2035 Node *rc_exp = cmp->in(1);
2036 Node *limit = cmp->in(2);
2037 jint scale_con= 1; // Assume trip counter not scaled
2038
2039 Node *limit_c = get_ctrl(limit);
2040 if( loop->is_member(get_loop(limit_c) ) ) {
2041 // Compare might have operands swapped; commute them
2042 b_test = b_test.commute();
2043 rc_exp = cmp->in(2);
2044 limit = cmp->in(1);
2045 limit_c = get_ctrl(limit);
2046 if( loop->is_member(get_loop(limit_c) ) )
2047 continue; // Both inputs are loop varying; cannot RCE
2048 }
2049 // Here we know 'limit' is loop invariant
2050
2051 // 'limit' maybe pinned below the zero trip test (probably from a
2052 // previous round of rce), in which case, it can't be used in the
2053 // zero trip test expression which must occur before the zero test's if.
2054 if( limit_c == ctrl ) {
2055 continue; // Don't rce this check but continue looking for other candidates.
2056 }
2057
2058 // Check for scaled induction variable plus an offset
2059 Node *offset = NULL;
2060
2061 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
2062 continue;
2063 }
2064
2065 Node *offset_c = get_ctrl(offset);
2066 if( loop->is_member( get_loop(offset_c) ) )
2067 continue; // Offset is not really loop invariant
2068 // Here we know 'offset' is loop invariant.
2069
2070 // As above for the 'limit', the 'offset' maybe pinned below the
2071 // zero trip test.
2072 if( offset_c == ctrl ) {
2073 continue; // Don't rce this check but continue looking for other candidates.
2074 }
2075 #ifdef ASSERT
2076 if (TraceRangeLimitCheck) {
2077 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
2078 bol->dump(2);
2079 }
2080 #endif
2081 // At this point we have the expression as:
2082 // scale_con * trip_counter + offset :: limit
2083 // where scale_con, offset and limit are loop invariant. Trip_counter
2084 // monotonically increases by stride_con, a constant. Both (or either)
2085 // stride_con and scale_con can be negative which will flip about the
2086 // sense of the test.
2087
2088 // Adjust pre and main loop limits to guard the correct iteration set
2089 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
2090 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
2091 // The underflow and overflow limits: 0 <= scale*I+offset < limit
2092 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
2093 if (!conditional_rc) {
2094 // (0-offset)/scale could be outside of loop iterations range.
2095 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
2096 }
2097 } else {
2098 #ifndef PRODUCT
2099 if( PrintOpto )
2100 tty->print_cr("missed RCE opportunity");
2101 #endif
2102 continue; // In release mode, ignore it
2103 }
2104 } else { // Otherwise work on normal compares
2105 switch( b_test._test ) {
2106 case BoolTest::gt:
2107 // Fall into GE case
2108 case BoolTest::ge:
2109 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
2110 scale_con = -scale_con;
2111 offset = new SubINode( zero, offset );
2112 register_new_node( offset, pre_ctrl );
2113 limit = new SubINode( zero, limit );
2114 register_new_node( limit, pre_ctrl );
2115 // Fall into LE case
2116 case BoolTest::le:
2117 if (b_test._test != BoolTest::gt) {
2118 // Convert X <= Y to X < Y+1
2119 limit = new AddINode( limit, one );
2120 register_new_node( limit, pre_ctrl );
2121 }
2122 // Fall into LT case
2123 case BoolTest::lt:
2124 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
2125 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
2126 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
2127 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
2128 if (!conditional_rc) {
2129 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
2130 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
2131 // still be outside of loop range.
2132 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
2133 }
2134 break;
2135 default:
2136 #ifndef PRODUCT
2137 if( PrintOpto )
2138 tty->print_cr("missed RCE opportunity");
2139 #endif
2140 continue; // Unhandled case
2141 }
2142 }
2143
2144 // Kill the eliminated test
2145 C->set_major_progress();
2146 Node *kill_con = _igvn.intcon( 1-flip );
2147 set_ctrl(kill_con, C->root());
2148 _igvn.replace_input_of(iff, 1, kill_con);
2149 // Find surviving projection
2150 assert(iff->is_If(), "");
2151 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2152 // Find loads off the surviving projection; remove their control edge
2153 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2154 Node* cd = dp->fast_out(i); // Control-dependent node
2155 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop
2156 // Allow the load to float around in the loop, or before it
2157 // but NOT before the pre-loop.
2158 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
2159 --i;
2160 --imax;
2161 }
2162 }
2163
2164 } // End of is IF
2165
2166 }
2167
2168 // Update loop limits
2169 if (conditional_rc) {
2170 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit)
2171 : (Node*)new MaxINode(pre_limit, orig_limit);
2172 register_new_node(pre_limit, pre_ctrl);
2173 }
2174 _igvn.replace_input_of(pre_opaq, 1, pre_limit);
2175
2176 // Note:: we are making the main loop limit no longer precise;
2177 // need to round up based on stride.
2178 cl->set_nonexact_trip_count();
2179 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
2180 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
2181 // Hopefully, compiler will optimize for powers of 2.
2182 Node *ctrl = get_ctrl(main_limit);
2183 Node *stride = cl->stride();
2184 Node *init = cl->init_trip()->uncast();
2185 Node *span = new SubINode(main_limit,init);
2186 register_new_node(span,ctrl);
2187 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
2188 Node *add = new AddINode(span,rndup);
2189 register_new_node(add,ctrl);
2190 Node *div = new DivINode(0,add,stride);
2191 register_new_node(div,ctrl);
2192 Node *mul = new MulINode(div,stride);
2193 register_new_node(mul,ctrl);
2194 Node *newlim = new AddINode(mul,init);
2195 register_new_node(newlim,ctrl);
2196 main_limit = newlim;
2197 }
2198
2199 Node *main_cle = cl->loopexit();
2200 Node *main_bol = main_cle->in(1);
2201 // Hacking loop bounds; need private copies of exit test
2202 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2203 main_bol = main_bol->clone();// Clone a private BoolNode
2204 register_new_node( main_bol, main_cle->in(0) );
2205 _igvn.replace_input_of(main_cle, 1, main_bol);
2206 }
2207 Node *main_cmp = main_bol->in(1);
2208 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2209 main_cmp = main_cmp->clone();// Clone a private CmpNode
2210 register_new_node( main_cmp, main_cle->in(0) );
2211 _igvn.replace_input_of(main_bol, 1, main_cmp);
2212 }
2213 // Hack the now-private loop bounds
2214 _igvn.replace_input_of(main_cmp, 2, main_limit);
2215 // The OpaqueNode is unshared by design
2216 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2217 _igvn.replace_input_of(opqzm, 1, main_limit);
2218 }
2219
2220 //------------------------------DCE_loop_body----------------------------------
2221 // Remove simplistic dead code from loop body
2222 void IdealLoopTree::DCE_loop_body() {
2223 for( uint i = 0; i < _body.size(); i++ )
2224 if( _body.at(i)->outcnt() == 0 )
2225 _body.map( i--, _body.pop() );
2226 }
2227
2228
2229 //------------------------------adjust_loop_exit_prob--------------------------
2230 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2231 // Replace with a 1-in-10 exit guess.
2232 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2233 Node *test = tail();
2234 while( test != _head ) {
2235 uint top = test->Opcode();
2236 if( top == Op_IfTrue || top == Op_IfFalse ) {
2237 int test_con = ((ProjNode*)test)->_con;
2238 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2239 IfNode *iff = test->in(0)->as_If();
2240 if( iff->outcnt() == 2 ) { // Ignore dead tests
2241 Node *bol = iff->in(1);
2242 if( bol && bol->req() > 1 && bol->in(1) &&
2243 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2244 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2245 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2246 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2247 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2248 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2249 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2250 return; // Allocation loops RARELY take backedge
2251 // Find the OTHER exit path from the IF
2252 Node* ex = iff->proj_out(1-test_con);
2253 float p = iff->_prob;
2254 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2255 if( top == Op_IfTrue ) {
2256 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2257 iff->_prob = PROB_STATIC_FREQUENT;
2258 }
2259 } else {
2260 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2261 iff->_prob = PROB_STATIC_INFREQUENT;
2262 }
2263 }
2264 }
2265 }
2266 }
2267 test = phase->idom(test);
2268 }
2269 }
2270
2271 #ifdef ASSERT
2272 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) {
2273 Node *ctrl = cl->in(LoopNode::EntryControl);
2274 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
2275 Node *iffm = ctrl->in(0);
2276 assert(iffm->Opcode() == Op_If, "");
2277 Node *p_f = iffm->in(0);
2278 assert(p_f->Opcode() == Op_IfFalse, "");
2279 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2280 assert(pre_end->loopnode()->is_pre_loop(), "");
2281 return pre_end->loopnode();
2282 }
2283 #endif
2284
2285 // Remove the main and post loops and make the pre loop execute all
2286 // iterations. Useful when the pre loop is found empty.
2287 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) {
2288 CountedLoopEndNode* pre_end = cl->loopexit();
2289 Node* pre_cmp = pre_end->cmp_node();
2290 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) {
2291 // Only safe to remove the main loop if the compiler optimized it
2292 // out based on an unknown number of iterations
2293 return;
2294 }
2295
2296 // Can we find the main loop?
2297 if (_next == NULL) {
2298 return;
2299 }
2300
2301 Node* next_head = _next->_head;
2302 if (!next_head->is_CountedLoop()) {
2303 return;
2304 }
2305
2306 CountedLoopNode* main_head = next_head->as_CountedLoop();
2307 if (!main_head->is_main_loop()) {
2308 return;
2309 }
2310
2311 assert(locate_pre_from_main(main_head) == cl, "bad main loop");
2312 Node* main_iff = main_head->in(LoopNode::EntryControl)->in(0);
2313
2314 // Remove the Opaque1Node of the pre loop and make it execute all iterations
2315 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2));
2316 // Remove the Opaque1Node of the main loop so it can be optimized out
2317 Node* main_cmp = main_iff->in(1)->in(1);
2318 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?");
2319 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1));
2320 }
2321
2322 //------------------------------policy_do_remove_empty_loop--------------------
2323 // Micro-benchmark spamming. Policy is to always remove empty loops.
2324 // The 'DO' part is to replace the trip counter with the value it will
2325 // have on the last iteration. This will break the loop.
2326 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2327 // Minimum size must be empty loop
2328 if (_body.size() > EMPTY_LOOP_SIZE)
2329 return false;
2330
2331 if (!_head->is_CountedLoop())
2332 return false; // Dead loop
2333 CountedLoopNode *cl = _head->as_CountedLoop();
2334 if (!cl->is_valid_counted_loop())
2335 return false; // Malformed loop
2336 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2337 return false; // Infinite loop
2338
2339 if (cl->is_pre_loop()) {
2340 // If the loop we are removing is a pre-loop then the main and
2341 // post loop can be removed as well
2342 remove_main_post_loops(cl, phase);
2343 }
2344
2345 #ifdef ASSERT
2346 // Ensure only one phi which is the iv.
2347 Node* iv = NULL;
2348 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2349 Node* n = cl->fast_out(i);
2350 if (n->Opcode() == Op_Phi) {
2351 assert(iv == NULL, "Too many phis" );
2352 iv = n;
2353 }
2354 }
2355 assert(iv == cl->phi(), "Wrong phi" );
2356 #endif
2357
2358 // main and post loops have explicitly created zero trip guard
2359 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2360 if (needs_guard) {
2361 // Skip guard if values not overlap.
2362 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2363 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2364 int stride_con = cl->stride_con();
2365 if (stride_con > 0) {
2366 needs_guard = (init_t->_hi >= limit_t->_lo);
2367 } else {
2368 needs_guard = (init_t->_lo <= limit_t->_hi);
2369 }
2370 }
2371 if (needs_guard) {
2372 // Check for an obvious zero trip guard.
2373 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2374 if (inctrl->Opcode() == Op_IfTrue) {
2375 // The test should look like just the backedge of a CountedLoop
2376 Node* iff = inctrl->in(0);
2377 if (iff->is_If()) {
2378 Node* bol = iff->in(1);
2379 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
2380 Node* cmp = bol->in(1);
2381 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
2382 needs_guard = false;
2383 }
2384 }
2385 }
2386 }
2387 }
2388
2389 #ifndef PRODUCT
2390 if (PrintOpto) {
2391 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2392 this->dump_head();
2393 } else if (TraceLoopOpts) {
2394 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2395 this->dump_head();
2396 }
2397 #endif
2398
2399 if (needs_guard) {
2400 // Peel the loop to ensure there's a zero trip guard
2401 Node_List old_new;
2402 phase->do_peeling(this, old_new);
2403 }
2404
2405 // Replace the phi at loop head with the final value of the last
2406 // iteration. Then the CountedLoopEnd will collapse (backedge never
2407 // taken) and all loop-invariant uses of the exit values will be correct.
2408 Node *phi = cl->phi();
2409 Node *exact_limit = phase->exact_limit(this);
2410 if (exact_limit != cl->limit()) {
2411 // We also need to replace the original limit to collapse loop exit.
2412 Node* cmp = cl->loopexit()->cmp_node();
2413 assert(cl->limit() == cmp->in(2), "sanity");
2414 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2415 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2416 }
2417 // Note: the final value after increment should not overflow since
2418 // counted loop has limit check predicate.
2419 Node *final = new SubINode( exact_limit, cl->stride() );
2420 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2421 phase->_igvn.replace_node(phi,final);
2422 phase->C->set_major_progress();
2423 return true;
2424 }
2425
2426 //------------------------------policy_do_one_iteration_loop-------------------
2427 // Convert one iteration loop into normal code.
2428 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2429 if (!_head->as_Loop()->is_valid_counted_loop())
2430 return false; // Only for counted loop
2431
2432 CountedLoopNode *cl = _head->as_CountedLoop();
2433 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2434 return false;
2435 }
2436
2437 #ifndef PRODUCT
2438 if(TraceLoopOpts) {
2439 tty->print("OneIteration ");
2440 this->dump_head();
2441 }
2442 #endif
2443
2444 Node *init_n = cl->init_trip();
2445 #ifdef ASSERT
2446 // Loop boundaries should be constant since trip count is exact.
2447 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2448 #endif
2449 // Replace the phi at loop head with the value of the init_trip.
2450 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2451 // and all loop-invariant uses of the exit values will be correct.
2452 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2453 phase->C->set_major_progress();
2454 return true;
2455 }
2456
2457 //=============================================================================
2458 //------------------------------iteration_split_impl---------------------------
2459 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2460 // Compute exact loop trip count if possible.
2461 compute_exact_trip_count(phase);
2462
2463 // Convert one iteration loop into normal code.
2464 if (policy_do_one_iteration_loop(phase))
2465 return true;
2466
2467 // Check and remove empty loops (spam micro-benchmarks)
2468 if (policy_do_remove_empty_loop(phase))
2469 return true; // Here we removed an empty loop
2470
2471 bool should_peel = policy_peeling(phase); // Should we peel?
2472
2473 bool should_unswitch = policy_unswitching(phase);
2474
2475 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2476 // This removes loop-invariant tests (usually null checks).
2477 if (!_head->is_CountedLoop()) { // Non-counted loop
2478 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2479 // Partial peel succeeded so terminate this round of loop opts
2480 return false;
2481 }
2482 if (should_peel) { // Should we peel?
2483 #ifndef PRODUCT
2484 if (PrintOpto) tty->print_cr("should_peel");
2485 #endif
2486 phase->do_peeling(this,old_new);
2487 } else if (should_unswitch) {
2488 phase->do_unswitching(this, old_new);
2489 }
2490 return true;
2491 }
2492 CountedLoopNode *cl = _head->as_CountedLoop();
2493
2494 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2495
2496 // Do nothing special to pre- and post- loops
2497 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2498
2499 // Compute loop trip count from profile data
2500 compute_profile_trip_cnt(phase);
2501
2502 // Before attempting fancy unrolling, RCE or alignment, see if we want
2503 // to completely unroll this loop or do loop unswitching.
2504 if (cl->is_normal_loop()) {
2505 if (should_unswitch) {
2506 phase->do_unswitching(this, old_new);
2507 return true;
2508 }
2509 bool should_maximally_unroll = policy_maximally_unroll(phase);
2510 if (should_maximally_unroll) {
2511 // Here we did some unrolling and peeling. Eventually we will
2512 // completely unroll this loop and it will no longer be a loop.
2513 phase->do_maximally_unroll(this,old_new);
2514 return true;
2515 }
2516 }
2517
2518 // Skip next optimizations if running low on nodes. Note that
2519 // policy_unswitching and policy_maximally_unroll have this check.
2520 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes();
2521 if ((int)(2 * _body.size()) > nodes_left) {
2522 return true;
2523 }
2524
2525 // Counted loops may be peeled, may need some iterations run up
2526 // front for RCE, and may want to align loop refs to a cache
2527 // line. Thus we clone a full loop up front whose trip count is
2528 // at least 1 (if peeling), but may be several more.
2529
2530 // The main loop will start cache-line aligned with at least 1
2531 // iteration of the unrolled body (zero-trip test required) and
2532 // will have some range checks removed.
2533
2534 // A post-loop will finish any odd iterations (leftover after
2535 // unrolling), plus any needed for RCE purposes.
2536
2537 bool should_unroll = policy_unroll(phase);
2538
2539 bool should_rce = policy_range_check(phase);
2540
2541 bool should_align = policy_align(phase);
2542
2543 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2544 // need a pre-loop. We may still need to peel an initial iteration but
2545 // we will not be needing an unknown number of pre-iterations.
2546 //
2547 // Basically, if may_rce_align reports FALSE first time through,
2548 // we will not be able to later do RCE or Aligning on this loop.
2549 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2550
2551 // If we have any of these conditions (RCE, alignment, unrolling) met, then
2552 // we switch to the pre-/main-/post-loop model. This model also covers
2553 // peeling.
2554 if (should_rce || should_align || should_unroll) {
2555 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
2556 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2557
2558 // Adjust the pre- and main-loop limits to let the pre and post loops run
2559 // with full checks, but the main-loop with no checks. Remove said
2560 // checks from the main body.
2561 if (should_rce)
2562 phase->do_range_check(this,old_new);
2563
2564 // Double loop body for unrolling. Adjust the minimum-trip test (will do
2565 // twice as many iterations as before) and the main body limit (only do
2566 // an even number of trips). If we are peeling, we might enable some RCE
2567 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2568 // peeling.
2569 if (should_unroll && !should_peel) {
2570 phase->do_unroll(this, old_new, true);
2571 }
2572
2573 // Adjust the pre-loop limits to align the main body
2574 // iterations.
2575 if (should_align)
2576 Unimplemented();
2577
2578 } else { // Else we have an unchanged counted loop
2579 if (should_peel) // Might want to peel but do nothing else
2580 phase->do_peeling(this,old_new);
2581 }
2582 return true;
2583 }
2584
2585
2586 //=============================================================================
2587 //------------------------------iteration_split--------------------------------
2588 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2589 // Recursively iteration split nested loops
2590 if (_child && !_child->iteration_split(phase, old_new))
2591 return false;
2592
2593 // Clean out prior deadwood
2594 DCE_loop_body();
2595
2596
2597 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2598 // Replace with a 1-in-10 exit guess.
2599 if (_parent /*not the root loop*/ &&
2600 !_irreducible &&
2601 // Also ignore the occasional dead backedge
2602 !tail()->is_top()) {
2603 adjust_loop_exit_prob(phase);
2604 }
2605
2606 // Gate unrolling, RCE and peeling efforts.
2607 if (!_child && // If not an inner loop, do not split
2608 !_irreducible &&
2609 _allow_optimizations &&
2610 !tail()->is_top()) { // Also ignore the occasional dead backedge
2611 if (!_has_call) {
2612 if (!iteration_split_impl(phase, old_new)) {
2613 return false;
2614 }
2615 } else if (policy_unswitching(phase)) {
2616 phase->do_unswitching(this, old_new);
2617 }
2618 }
2619
2620 // Minor offset re-organization to remove loop-fallout uses of
2621 // trip counter when there was no major reshaping.
2622 phase->reorg_offsets(this);
2623
2624 if (_next && !_next->iteration_split(phase, old_new))
2625 return false;
2626 return true;
2627 }
2628
2629
2630 //=============================================================================
2631 // Process all the loops in the loop tree and replace any fill
2632 // patterns with an intrisc version.
2633 bool PhaseIdealLoop::do_intrinsify_fill() {
2634 bool changed = false;
2635 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2636 IdealLoopTree* lpt = iter.current();
2637 changed |= intrinsify_fill(lpt);
2638 }
2639 return changed;
2640 }
2641
2642
2643 // Examine an inner loop looking for a a single store of an invariant
2644 // value in a unit stride loop,
2645 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2646 Node*& shift, Node*& con) {
2647 const char* msg = NULL;
2648 Node* msg_node = NULL;
2649
2650 store_value = NULL;
2651 con = NULL;
2652 shift = NULL;
2653
2654 // Process the loop looking for stores. If there are multiple
2655 // stores or extra control flow give at this point.
2656 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2657 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2658 Node* n = lpt->_body.at(i);
2659 if (n->outcnt() == 0) continue; // Ignore dead
2660 if (n->is_Store()) {
2661 if (store != NULL) {
2662 msg = "multiple stores";
2663 break;
2664 }
2665 int opc = n->Opcode();
2666 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
2667 msg = "oop fills not handled";
2668 break;
2669 }
2670 Node* value = n->in(MemNode::ValueIn);
2671 if (!lpt->is_invariant(value)) {
2672 msg = "variant store value";
2673 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2674 msg = "not array address";
2675 }
2676 store = n;
2677 store_value = value;
2678 } else if (n->is_If() && n != head->loopexit()) {
2679 msg = "extra control flow";
2680 msg_node = n;
2681 }
2682 }
2683
2684 if (store == NULL) {
2685 // No store in loop
2686 return false;
2687 }
2688
2689 if (msg == NULL && head->stride_con() != 1) {
2690 // could handle negative strides too
2691 if (head->stride_con() < 0) {
2692 msg = "negative stride";
2693 } else {
2694 msg = "non-unit stride";
2695 }
2696 }
2697
2698 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2699 msg = "can't handle store address";
2700 msg_node = store->in(MemNode::Address);
2701 }
2702
2703 if (msg == NULL &&
2704 (!store->in(MemNode::Memory)->is_Phi() ||
2705 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2706 msg = "store memory isn't proper phi";
2707 msg_node = store->in(MemNode::Memory);
2708 }
2709
2710 // Make sure there is an appropriate fill routine
2711 BasicType t = store->as_Mem()->memory_type();
2712 const char* fill_name;
2713 if (msg == NULL &&
2714 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2715 msg = "unsupported store";
2716 msg_node = store;
2717 }
2718
2719 if (msg != NULL) {
2720 #ifndef PRODUCT
2721 if (TraceOptimizeFill) {
2722 tty->print_cr("not fill intrinsic candidate: %s", msg);
2723 if (msg_node != NULL) msg_node->dump();
2724 }
2725 #endif
2726 return false;
2727 }
2728
2729 // Make sure the address expression can be handled. It should be
2730 // head->phi * elsize + con. head->phi might have a ConvI2L.
2731 Node* elements[4];
2732 Node* conv = NULL;
2733 bool found_index = false;
2734 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2735 for (int e = 0; e < count; e++) {
2736 Node* n = elements[e];
2737 if (n->is_Con() && con == NULL) {
2738 con = n;
2739 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2740 Node* value = n->in(1);
2741 #ifdef _LP64
2742 if (value->Opcode() == Op_ConvI2L) {
2743 conv = value;
2744 value = value->in(1);
2745 }
2746 #endif
2747 if (value != head->phi()) {
2748 msg = "unhandled shift in address";
2749 } else {
2750 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2751 msg = "scale doesn't match";
2752 } else {
2753 found_index = true;
2754 shift = n;
2755 }
2756 }
2757 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2758 if (n->in(1) == head->phi()) {
2759 found_index = true;
2760 conv = n;
2761 } else {
2762 msg = "unhandled input to ConvI2L";
2763 }
2764 } else if (n == head->phi()) {
2765 // no shift, check below for allowed cases
2766 found_index = true;
2767 } else {
2768 msg = "unhandled node in address";
2769 msg_node = n;
2770 }
2771 }
2772
2773 if (count == -1) {
2774 msg = "malformed address expression";
2775 msg_node = store;
2776 }
2777
2778 if (!found_index) {
2779 msg = "missing use of index";
2780 }
2781
2782 // byte sized items won't have a shift
2783 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2784 msg = "can't find shift";
2785 msg_node = store;
2786 }
2787
2788 if (msg != NULL) {
2789 #ifndef PRODUCT
2790 if (TraceOptimizeFill) {
2791 tty->print_cr("not fill intrinsic: %s", msg);
2792 if (msg_node != NULL) msg_node->dump();
2793 }
2794 #endif
2795 return false;
2796 }
2797
2798 // No make sure all the other nodes in the loop can be handled
2799 VectorSet ok(Thread::current()->resource_area());
2800
2801 // store related values are ok
2802 ok.set(store->_idx);
2803 ok.set(store->in(MemNode::Memory)->_idx);
2804
2805 CountedLoopEndNode* loop_exit = head->loopexit();
2806 guarantee(loop_exit != NULL, "no loop exit node");
2807
2808 // Loop structure is ok
2809 ok.set(head->_idx);
2810 ok.set(loop_exit->_idx);
2811 ok.set(head->phi()->_idx);
2812 ok.set(head->incr()->_idx);
2813 ok.set(loop_exit->cmp_node()->_idx);
2814 ok.set(loop_exit->in(1)->_idx);
2815
2816 // Address elements are ok
2817 if (con) ok.set(con->_idx);
2818 if (shift) ok.set(shift->_idx);
2819 if (conv) ok.set(conv->_idx);
2820
2821 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2822 Node* n = lpt->_body.at(i);
2823 if (n->outcnt() == 0) continue; // Ignore dead
2824 if (ok.test(n->_idx)) continue;
2825 // Backedge projection is ok
2826 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
2827 if (!n->is_AddP()) {
2828 msg = "unhandled node";
2829 msg_node = n;
2830 break;
2831 }
2832 }
2833
2834 // Make sure no unexpected values are used outside the loop
2835 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2836 Node* n = lpt->_body.at(i);
2837 // These values can be replaced with other nodes if they are used
2838 // outside the loop.
2839 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2840 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2841 Node* use = iter.get();
2842 if (!lpt->_body.contains(use)) {
2843 msg = "node is used outside loop";
2844 // lpt->_body.dump();
2845 msg_node = n;
2846 break;
2847 }
2848 }
2849 }
2850
2851 #ifdef ASSERT
2852 if (TraceOptimizeFill) {
2853 if (msg != NULL) {
2854 tty->print_cr("no fill intrinsic: %s", msg);
2855 if (msg_node != NULL) msg_node->dump();
2856 } else {
2857 tty->print_cr("fill intrinsic for:");
2858 }
2859 store->dump();
2860 if (Verbose) {
2861 lpt->_body.dump();
2862 }
2863 }
2864 #endif
2865
2866 return msg == NULL;
2867 }
2868
2869
2870
2871 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2872 // Only for counted inner loops
2873 if (!lpt->is_counted() || !lpt->is_inner()) {
2874 return false;
2875 }
2876
2877 // Must have constant stride
2878 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2879 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
2880 return false;
2881 }
2882
2883 // Check that the body only contains a store of a loop invariant
2884 // value that is indexed by the loop phi.
2885 Node* store = NULL;
2886 Node* store_value = NULL;
2887 Node* shift = NULL;
2888 Node* offset = NULL;
2889 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2890 return false;
2891 }
2892
2893 #ifndef PRODUCT
2894 if (TraceLoopOpts) {
2895 tty->print("ArrayFill ");
2896 lpt->dump_head();
2897 }
2898 #endif
2899
2900 // Now replace the whole loop body by a call to a fill routine that
2901 // covers the same region as the loop.
2902 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2903
2904 // Build an expression for the beginning of the copy region
2905 Node* index = head->init_trip();
2906 #ifdef _LP64
2907 index = new ConvI2LNode(index);
2908 _igvn.register_new_node_with_optimizer(index);
2909 #endif
2910 if (shift != NULL) {
2911 // byte arrays don't require a shift but others do.
2912 index = new LShiftXNode(index, shift->in(2));
2913 _igvn.register_new_node_with_optimizer(index);
2914 }
2915 index = new AddPNode(base, base, index);
2916 _igvn.register_new_node_with_optimizer(index);
2917 Node* from = new AddPNode(base, index, offset);
2918 _igvn.register_new_node_with_optimizer(from);
2919 // Compute the number of elements to copy
2920 Node* len = new SubINode(head->limit(), head->init_trip());
2921 _igvn.register_new_node_with_optimizer(len);
2922
2923 BasicType t = store->as_Mem()->memory_type();
2924 bool aligned = false;
2925 if (offset != NULL && head->init_trip()->is_Con()) {
2926 int element_size = type2aelembytes(t);
2927 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2928 }
2929
2930 // Build a call to the fill routine
2931 const char* fill_name;
2932 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2933 assert(fill != NULL, "what?");
2934
2935 // Convert float/double to int/long for fill routines
2936 if (t == T_FLOAT) {
2937 store_value = new MoveF2INode(store_value);
2938 _igvn.register_new_node_with_optimizer(store_value);
2939 } else if (t == T_DOUBLE) {
2940 store_value = new MoveD2LNode(store_value);
2941 _igvn.register_new_node_with_optimizer(store_value);
2942 }
2943
2944 if (CCallingConventionRequiresIntsAsLongs &&
2945 // See StubRoutines::select_fill_function for types. FLOAT has been converted to INT.
2946 (t == T_FLOAT || t == T_INT || is_subword_type(t))) {
2947 store_value = new ConvI2LNode(store_value);
2948 _igvn.register_new_node_with_optimizer(store_value);
2949 }
2950
2951 Node* mem_phi = store->in(MemNode::Memory);
2952 Node* result_ctrl;
2953 Node* result_mem;
2954 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2955 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill,
2956 fill_name, TypeAryPtr::get_array_body_type(t));
2957 uint cnt = 0;
2958 call->init_req(TypeFunc::Parms + cnt++, from);
2959 call->init_req(TypeFunc::Parms + cnt++, store_value);
2960 if (CCallingConventionRequiresIntsAsLongs) {
2961 call->init_req(TypeFunc::Parms + cnt++, C->top());
2962 }
2963 #ifdef _LP64
2964 len = new ConvI2LNode(len);
2965 _igvn.register_new_node_with_optimizer(len);
2966 #endif
2967 call->init_req(TypeFunc::Parms + cnt++, len);
2968 #ifdef _LP64
2969 call->init_req(TypeFunc::Parms + cnt++, C->top());
2970 #endif
2971 call->init_req(TypeFunc::Control, head->init_control());
2972 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O.
2973 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl));
2974 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr));
2975 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
2976 _igvn.register_new_node_with_optimizer(call);
2977 result_ctrl = new ProjNode(call,TypeFunc::Control);
2978 _igvn.register_new_node_with_optimizer(result_ctrl);
2979 result_mem = new ProjNode(call,TypeFunc::Memory);
2980 _igvn.register_new_node_with_optimizer(result_mem);
2981
2982 /* Disable following optimization until proper fix (add missing checks).
2983
2984 // If this fill is tightly coupled to an allocation and overwrites
2985 // the whole body, allow it to take over the zeroing.
2986 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2987 if (alloc != NULL && alloc->is_AllocateArray()) {
2988 Node* length = alloc->as_AllocateArray()->Ideal_length();
2989 if (head->limit() == length &&
2990 head->init_trip() == _igvn.intcon(0)) {
2991 if (TraceOptimizeFill) {
2992 tty->print_cr("Eliminated zeroing in allocation");
2993 }
2994 alloc->maybe_set_complete(&_igvn);
2995 } else {
2996 #ifdef ASSERT
2997 if (TraceOptimizeFill) {
2998 tty->print_cr("filling array but bounds don't match");
2999 alloc->dump();
3000 head->init_trip()->dump();
3001 head->limit()->dump();
3002 length->dump();
3003 }
3004 #endif
3005 }
3006 }
3007 */
3008
3009 // Redirect the old control and memory edges that are outside the loop.
3010 Node* exit = head->loopexit()->proj_out(0);
3011 // Sometimes the memory phi of the head is used as the outgoing
3012 // state of the loop. It's safe in this case to replace it with the
3013 // result_mem.
3014 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
3015 _igvn.replace_node(exit, result_ctrl);
3016 _igvn.replace_node(store, result_mem);
3017 // Any uses the increment outside of the loop become the loop limit.
3018 _igvn.replace_node(head->incr(), head->limit());
3019
3020 // Disconnect the head from the loop.
3021 for (uint i = 0; i < lpt->_body.size(); i++) {
3022 Node* n = lpt->_body.at(i);
3023 _igvn.replace_node(n, C->top());
3024 }
3025
3026 return true;
3027 }