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