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