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