1 /*
2 * Copyright (c) 2000, 2012, 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 #define MAX_UNROLL 16 // maximum number of unrolls for main loop
628
629 //------------------------------policy_unroll----------------------------------
630 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
631 // the loop is a CountedLoop and the body is small enough.
632 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
633
634 CountedLoopNode *cl = _head->as_CountedLoop();
635 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
636
637 if (!cl->is_valid_counted_loop())
638 return false; // Malformed counted loop
639
640 // Protect against over-unrolling.
641 // After split at least one iteration will be executed in pre-loop.
642 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
643
644 int future_unroll_ct = cl->unrolled_count() * 2;
645 if (future_unroll_ct > MAX_UNROLL) return false;
646
647 // Check for initial stride being a small enough constant
648 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
649
650 // Don't unroll if the next round of unrolling would push us
651 // over the expected trip count of the loop. One is subtracted
652 // from the expected trip count because the pre-loop normally
653 // executes 1 iteration.
654 if (UnrollLimitForProfileCheck > 0 &&
655 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
656 future_unroll_ct > UnrollLimitForProfileCheck &&
657 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
658 return false;
659 }
660
661 // When unroll count is greater than LoopUnrollMin, don't unroll if:
662 // the residual iterations are more than 10% of the trip count
663 // and rounds of "unroll,optimize" are not making significant progress
664 // Progress defined as current size less than 20% larger than previous size.
665 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
666 future_unroll_ct > LoopUnrollMin &&
667 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
668 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
669 return false;
670 }
671
672 Node *init_n = cl->init_trip();
673 Node *limit_n = cl->limit();
674 int stride_con = cl->stride_con();
675 // Non-constant bounds.
676 // Protect against over-unrolling when init or/and limit are not constant
677 // (so that trip_count's init value is maxint) but iv range is known.
678 if (init_n == NULL || !init_n->is_Con() ||
679 limit_n == NULL || !limit_n->is_Con()) {
680 Node* phi = cl->phi();
681 if (phi != NULL) {
682 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
683 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
684 int next_stride = stride_con * 2; // stride after this unroll
685 if (next_stride > 0) {
686 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
687 iv_type->_lo + next_stride > iv_type->_hi) {
688 return false; // over-unrolling
689 }
690 } else if (next_stride < 0) {
691 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
692 iv_type->_hi + next_stride < iv_type->_lo) {
693 return false; // over-unrolling
694 }
695 }
696 }
697 }
698
699 // After unroll limit will be adjusted: new_limit = limit-stride.
700 // Bailout if adjustment overflow.
701 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
702 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
703 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
704 return false; // overflow
705
706 // Adjust body_size to determine if we unroll or not
707 uint body_size = _body.size();
708 // Key test to unroll loop in CRC32 java code
709 int xors_in_loop = 0;
710 // Also count ModL, DivL and MulL which expand mightly
711 for (uint k = 0; k < _body.size(); k++) {
712 Node* n = _body.at(k);
713 switch (n->Opcode()) {
714 case Op_XorI: xors_in_loop++; break; // CRC32 java code
715 case Op_ModL: body_size += 30; break;
716 case Op_DivL: body_size += 30; break;
717 case Op_MulL: body_size += 10; break;
718 case Op_StrComp:
719 case Op_StrEquals:
720 case Op_StrIndexOf:
721 case Op_EncodeISOArray:
722 case Op_AryEq: {
723 // Do not unroll a loop with String intrinsics code.
724 // String intrinsics are large and have loops.
725 return false;
726 }
727 } // switch
728 }
729
730 // Check for being too big
731 if (body_size > (uint)LoopUnrollLimit) {
732 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
733 // Normal case: loop too big
734 return false;
735 }
736
737 // Unroll once! (Each trip will soon do double iterations)
738 return true;
739 }
740
741 //------------------------------policy_align-----------------------------------
742 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
743 // expression that does the alignment. Note that only one array base can be
744 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
745 // if we vectorize short memory ops into longer memory ops, we may want to
746 // increase alignment.
747 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
748 return false;
749 }
750
751 //------------------------------policy_range_check-----------------------------
752 // Return TRUE or FALSE if the loop should be range-check-eliminated.
753 // Actually we do iteration-splitting, a more powerful form of RCE.
754 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
755 if (!RangeCheckElimination) return false;
756
757 CountedLoopNode *cl = _head->as_CountedLoop();
758 // If we unrolled with no intention of doing RCE and we later
759 // changed our minds, we got no pre-loop. Either we need to
760 // make a new pre-loop, or we gotta disallow RCE.
761 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
762 Node *trip_counter = cl->phi();
763
764 // Check loop body for tests of trip-counter plus loop-invariant vs
765 // loop-invariant.
766 for (uint i = 0; i < _body.size(); i++) {
767 Node *iff = _body[i];
768 if (iff->Opcode() == Op_If) { // Test?
769
770 // Comparing trip+off vs limit
771 Node *bol = iff->in(1);
772 if (bol->req() != 2) continue; // dead constant test
773 if (!bol->is_Bool()) {
774 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
775 continue;
776 }
777 if (bol->as_Bool()->_test._test == BoolTest::ne)
778 continue; // not RC
779
780 Node *cmp = bol->in(1);
781
782 Node *rc_exp = cmp->in(1);
783 Node *limit = cmp->in(2);
784
785 Node *limit_c = phase->get_ctrl(limit);
786 if( limit_c == phase->C->top() )
787 return false; // Found dead test on live IF? No RCE!
788 if( is_member(phase->get_loop(limit_c) ) ) {
789 // Compare might have operands swapped; commute them
790 rc_exp = cmp->in(2);
791 limit = cmp->in(1);
792 limit_c = phase->get_ctrl(limit);
793 if( is_member(phase->get_loop(limit_c) ) )
794 continue; // Both inputs are loop varying; cannot RCE
795 }
796
797 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
798 continue;
799 }
800 // Yeah! Found a test like 'trip+off vs limit'
801 // Test is an IfNode, has 2 projections. If BOTH are in the loop
802 // we need loop unswitching instead of iteration splitting.
803 if( is_loop_exit(iff) )
804 return true; // Found reason to split iterations
805 } // End of is IF
806 }
807
808 return false;
809 }
810
811 //------------------------------policy_peel_only-------------------------------
812 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
813 // for unrolling loops with NO array accesses.
814 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
815
816 for( uint i = 0; i < _body.size(); i++ )
817 if( _body[i]->is_Mem() )
818 return false;
819
820 // No memory accesses at all!
821 return true;
822 }
823
824 //------------------------------clone_up_backedge_goo--------------------------
825 // If Node n lives in the back_ctrl block and cannot float, we clone a private
826 // version of n in preheader_ctrl block and return that, otherwise return n.
827 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
828 if( get_ctrl(n) != back_ctrl ) return n;
829
830 // Only visit once
831 if (visited.test_set(n->_idx)) {
832 Node *x = clones.find(n->_idx);
833 if (x != NULL)
834 return x;
835 return n;
836 }
837
838 Node *x = NULL; // If required, a clone of 'n'
839 // Check for 'n' being pinned in the backedge.
840 if( n->in(0) && n->in(0) == back_ctrl ) {
841 assert(clones.find(n->_idx) == NULL, "dead loop");
842 x = n->clone(); // Clone a copy of 'n' to preheader
843 clones.push(x, n->_idx);
844 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
845 }
846
847 // Recursive fixup any other input edges into x.
848 // If there are no changes we can just return 'n', otherwise
849 // we need to clone a private copy and change it.
850 for( uint i = 1; i < n->req(); i++ ) {
851 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
852 if( g != n->in(i) ) {
853 if( !x ) {
854 assert(clones.find(n->_idx) == NULL, "dead loop");
855 x = n->clone();
856 clones.push(x, n->_idx);
857 }
858 x->set_req(i, g);
859 }
860 }
861 if( x ) { // x can legally float to pre-header location
862 register_new_node( x, preheader_ctrl );
863 return x;
864 } else { // raise n to cover LCA of uses
865 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
866 }
867 return n;
868 }
869
870 //------------------------------insert_pre_post_loops--------------------------
871 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
872 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
873 // alignment. Useful to unroll loops that do no array accesses.
874 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
875
876 #ifndef PRODUCT
877 if (TraceLoopOpts) {
878 if (peel_only)
879 tty->print("PeelMainPost ");
880 else
881 tty->print("PreMainPost ");
882 loop->dump_head();
883 }
884 #endif
885 C->set_major_progress();
886
887 // Find common pieces of the loop being guarded with pre & post loops
888 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
889 assert( main_head->is_normal_loop(), "" );
890 CountedLoopEndNode *main_end = main_head->loopexit();
891 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
892 uint dd_main_head = dom_depth(main_head);
893 uint max = main_head->outcnt();
894
895 Node *pre_header= main_head->in(LoopNode::EntryControl);
896 Node *init = main_head->init_trip();
897 Node *incr = main_end ->incr();
898 Node *limit = main_end ->limit();
899 Node *stride = main_end ->stride();
900 Node *cmp = main_end ->cmp_node();
901 BoolTest::mask b_test = main_end->test_trip();
902
903 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
904 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
905 if( bol->outcnt() != 1 ) {
906 bol = bol->clone();
907 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
908 _igvn.hash_delete(main_end);
909 main_end->set_req(CountedLoopEndNode::TestValue, bol);
910 }
911 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
912 if( cmp->outcnt() != 1 ) {
913 cmp = cmp->clone();
914 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
915 _igvn.hash_delete(bol);
916 bol->set_req(1, cmp);
917 }
918
919 //------------------------------
920 // Step A: Create Post-Loop.
921 Node* main_exit = main_end->proj_out(false);
922 assert( main_exit->Opcode() == Op_IfFalse, "" );
923 int dd_main_exit = dom_depth(main_exit);
924
925 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
926 // loop pre-header illegally has 2 control users (old & new loops).
927 clone_loop( loop, old_new, dd_main_exit );
928 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
929 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
930 post_head->set_post_loop(main_head);
931
932 // Reduce the post-loop trip count.
933 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
934 post_end->_prob = PROB_FAIR;
935
936 // Build the main-loop normal exit.
937 IfFalseNode *new_main_exit = new (C) IfFalseNode(main_end);
938 _igvn.register_new_node_with_optimizer( new_main_exit );
939 set_idom(new_main_exit, main_end, dd_main_exit );
940 set_loop(new_main_exit, loop->_parent);
941
942 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
943 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
944 // (the main-loop trip-counter exit value) because we will be changing
945 // the exit value (via unrolling) so we cannot constant-fold away the zero
946 // trip guard until all unrolling is done.
947 Node *zer_opaq = new (C) Opaque1Node(C, incr);
948 Node *zer_cmp = new (C) CmpINode( zer_opaq, limit );
949 Node *zer_bol = new (C) BoolNode( zer_cmp, b_test );
950 register_new_node( zer_opaq, new_main_exit );
951 register_new_node( zer_cmp , new_main_exit );
952 register_new_node( zer_bol , new_main_exit );
953
954 // Build the IfNode
955 IfNode *zer_iff = new (C) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
956 _igvn.register_new_node_with_optimizer( zer_iff );
957 set_idom(zer_iff, new_main_exit, dd_main_exit);
958 set_loop(zer_iff, loop->_parent);
959
960 // Plug in the false-path, taken if we need to skip post-loop
961 _igvn.replace_input_of(main_exit, 0, zer_iff);
962 set_idom(main_exit, zer_iff, dd_main_exit);
963 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
964 // Make the true-path, must enter the post loop
965 Node *zer_taken = new (C) IfTrueNode( zer_iff );
966 _igvn.register_new_node_with_optimizer( zer_taken );
967 set_idom(zer_taken, zer_iff, dd_main_exit);
968 set_loop(zer_taken, loop->_parent);
969 // Plug in the true path
970 _igvn.hash_delete( post_head );
971 post_head->set_req(LoopNode::EntryControl, zer_taken);
972 set_idom(post_head, zer_taken, dd_main_exit);
973
974 Arena *a = Thread::current()->resource_area();
975 VectorSet visited(a);
976 Node_Stack clones(a, main_head->back_control()->outcnt());
977 // Step A3: Make the fall-in values to the post-loop come from the
978 // fall-out values of the main-loop.
979 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
980 Node* main_phi = main_head->fast_out(i);
981 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
982 Node *post_phi = old_new[main_phi->_idx];
983 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
984 post_head->init_control(),
985 main_phi->in(LoopNode::LoopBackControl),
986 visited, clones);
987 _igvn.hash_delete(post_phi);
988 post_phi->set_req( LoopNode::EntryControl, fallmain );
989 }
990 }
991
992 // Update local caches for next stanza
993 main_exit = new_main_exit;
994
995
996 //------------------------------
997 // Step B: Create Pre-Loop.
998
999 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1000 // loop pre-header illegally has 2 control users (old & new loops).
1001 clone_loop( loop, old_new, dd_main_head );
1002 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1003 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1004 pre_head->set_pre_loop(main_head);
1005 Node *pre_incr = old_new[incr->_idx];
1006
1007 // Reduce the pre-loop trip count.
1008 pre_end->_prob = PROB_FAIR;
1009
1010 // Find the pre-loop normal exit.
1011 Node* pre_exit = pre_end->proj_out(false);
1012 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1013 IfFalseNode *new_pre_exit = new (C) IfFalseNode(pre_end);
1014 _igvn.register_new_node_with_optimizer( new_pre_exit );
1015 set_idom(new_pre_exit, pre_end, dd_main_head);
1016 set_loop(new_pre_exit, loop->_parent);
1017
1018 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1019 // pre-loop, the main-loop may not execute at all. Later in life this
1020 // zero-trip guard will become the minimum-trip guard when we unroll
1021 // the main-loop.
1022 Node *min_opaq = new (C) Opaque1Node(C, limit);
1023 Node *min_cmp = new (C) CmpINode( pre_incr, min_opaq );
1024 Node *min_bol = new (C) BoolNode( min_cmp, b_test );
1025 register_new_node( min_opaq, new_pre_exit );
1026 register_new_node( min_cmp , new_pre_exit );
1027 register_new_node( min_bol , new_pre_exit );
1028
1029 // Build the IfNode (assume the main-loop is executed always).
1030 IfNode *min_iff = new (C) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1031 _igvn.register_new_node_with_optimizer( min_iff );
1032 set_idom(min_iff, new_pre_exit, dd_main_head);
1033 set_loop(min_iff, loop->_parent);
1034
1035 // Plug in the false-path, taken if we need to skip main-loop
1036 _igvn.hash_delete( pre_exit );
1037 pre_exit->set_req(0, min_iff);
1038 set_idom(pre_exit, min_iff, dd_main_head);
1039 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1040 // Make the true-path, must enter the main loop
1041 Node *min_taken = new (C) IfTrueNode( min_iff );
1042 _igvn.register_new_node_with_optimizer( min_taken );
1043 set_idom(min_taken, min_iff, dd_main_head);
1044 set_loop(min_taken, loop->_parent);
1045 // Plug in the true path
1046 _igvn.hash_delete( main_head );
1047 main_head->set_req(LoopNode::EntryControl, min_taken);
1048 set_idom(main_head, min_taken, dd_main_head);
1049
1050 visited.Clear();
1051 clones.clear();
1052 // Step B3: Make the fall-in values to the main-loop come from the
1053 // fall-out values of the pre-loop.
1054 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1055 Node* main_phi = main_head->fast_out(i2);
1056 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1057 Node *pre_phi = old_new[main_phi->_idx];
1058 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1059 main_head->init_control(),
1060 pre_phi->in(LoopNode::LoopBackControl),
1061 visited, clones);
1062 _igvn.hash_delete(main_phi);
1063 main_phi->set_req( LoopNode::EntryControl, fallpre );
1064 }
1065 }
1066
1067 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1068 // RCE and alignment may change this later.
1069 Node *cmp_end = pre_end->cmp_node();
1070 assert( cmp_end->in(2) == limit, "" );
1071 Node *pre_limit = new (C) AddINode( init, stride );
1072
1073 // Save the original loop limit in this Opaque1 node for
1074 // use by range check elimination.
1075 Node *pre_opaq = new (C) Opaque1Node(C, pre_limit, limit);
1076
1077 register_new_node( pre_limit, pre_head->in(0) );
1078 register_new_node( pre_opaq , pre_head->in(0) );
1079
1080 // Since no other users of pre-loop compare, I can hack limit directly
1081 assert( cmp_end->outcnt() == 1, "no other users" );
1082 _igvn.hash_delete(cmp_end);
1083 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1084
1085 // Special case for not-equal loop bounds:
1086 // Change pre loop test, main loop test, and the
1087 // main loop guard test to use lt or gt depending on stride
1088 // direction:
1089 // positive stride use <
1090 // negative stride use >
1091 //
1092 // not-equal test is kept for post loop to handle case
1093 // when init > limit when stride > 0 (and reverse).
1094
1095 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1096
1097 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1098 // Modify pre loop end condition
1099 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1100 BoolNode* new_bol0 = new (C) BoolNode(pre_bol->in(1), new_test);
1101 register_new_node( new_bol0, pre_head->in(0) );
1102 _igvn.hash_delete(pre_end);
1103 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
1104 // Modify main loop guard condition
1105 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1106 BoolNode* new_bol1 = new (C) BoolNode(min_bol->in(1), new_test);
1107 register_new_node( new_bol1, new_pre_exit );
1108 _igvn.hash_delete(min_iff);
1109 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1110 // Modify main loop end condition
1111 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1112 BoolNode* new_bol2 = new (C) BoolNode(main_bol->in(1), new_test);
1113 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1114 _igvn.hash_delete(main_end);
1115 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
1116 }
1117
1118 // Flag main loop
1119 main_head->set_main_loop();
1120 if( peel_only ) main_head->set_main_no_pre_loop();
1121
1122 // Subtract a trip count for the pre-loop.
1123 main_head->set_trip_count(main_head->trip_count() - 1);
1124
1125 // It's difficult to be precise about the trip-counts
1126 // for the pre/post loops. They are usually very short,
1127 // so guess that 4 trips is a reasonable value.
1128 post_head->set_profile_trip_cnt(4.0);
1129 pre_head->set_profile_trip_cnt(4.0);
1130
1131 // Now force out all loop-invariant dominating tests. The optimizer
1132 // finds some, but we _know_ they are all useless.
1133 peeled_dom_test_elim(loop,old_new);
1134 }
1135
1136 //------------------------------is_invariant-----------------------------
1137 // Return true if n is invariant
1138 bool IdealLoopTree::is_invariant(Node* n) const {
1139 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1140 if (n_c->is_top()) return false;
1141 return !is_member(_phase->get_loop(n_c));
1142 }
1143
1144
1145 //------------------------------do_unroll--------------------------------------
1146 // Unroll the loop body one step - make each trip do 2 iterations.
1147 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1148 assert(LoopUnrollLimit, "");
1149 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1150 CountedLoopEndNode *loop_end = loop_head->loopexit();
1151 assert(loop_end, "");
1152 #ifndef PRODUCT
1153 if (PrintOpto && VerifyLoopOptimizations) {
1154 tty->print("Unrolling ");
1155 loop->dump_head();
1156 } else if (TraceLoopOpts) {
1157 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1158 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1159 } else {
1160 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1161 }
1162 loop->dump_head();
1163 }
1164 #endif
1165
1166 // Remember loop node count before unrolling to detect
1167 // if rounds of unroll,optimize are making progress
1168 loop_head->set_node_count_before_unroll(loop->_body.size());
1169
1170 Node *ctrl = loop_head->in(LoopNode::EntryControl);
1171 Node *limit = loop_head->limit();
1172 Node *init = loop_head->init_trip();
1173 Node *stride = loop_head->stride();
1174
1175 Node *opaq = NULL;
1176 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1177 // Search for zero-trip guard.
1178 assert( loop_head->is_main_loop(), "" );
1179 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1180 Node *iff = ctrl->in(0);
1181 assert( iff->Opcode() == Op_If, "" );
1182 Node *bol = iff->in(1);
1183 assert( bol->Opcode() == Op_Bool, "" );
1184 Node *cmp = bol->in(1);
1185 assert( cmp->Opcode() == Op_CmpI, "" );
1186 opaq = cmp->in(2);
1187 // Occasionally it's possible for a zero-trip guard Opaque1 node to be
1188 // optimized away and then another round of loop opts attempted.
1189 // We can not optimize this particular loop in that case.
1190 if (opaq->Opcode() != Op_Opaque1)
1191 return; // Cannot find zero-trip guard! Bail out!
1192 // Zero-trip test uses an 'opaque' node which is not shared.
1193 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1194 }
1195
1196 C->set_major_progress();
1197
1198 Node* new_limit = NULL;
1199 if (UnrollLimitCheck) {
1200 int stride_con = stride->get_int();
1201 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1202 uint old_trip_count = loop_head->trip_count();
1203 // Verify that unroll policy result is still valid.
1204 assert(old_trip_count > 1 &&
1205 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1206
1207 // Adjust loop limit to keep valid iterations number after unroll.
1208 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1209 // which may overflow.
1210 if (!adjust_min_trip) {
1211 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1212 "odd trip count for maximally unroll");
1213 // Don't need to adjust limit for maximally unroll since trip count is even.
1214 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1215 // Loop's limit is constant. Loop's init could be constant when pre-loop
1216 // become peeled iteration.
1217 jlong init_con = init->get_int();
1218 // We can keep old loop limit if iterations count stays the same:
1219 // old_trip_count == new_trip_count * 2
1220 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1221 // so we also don't need to adjust zero trip test.
1222 jlong limit_con = limit->get_int();
1223 // (stride_con*2) not overflow since stride_con <= 8.
1224 int new_stride_con = stride_con * 2;
1225 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1226 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1227 // New trip count should satisfy next conditions.
1228 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1229 uint new_trip_count = (uint)trip_count;
1230 adjust_min_trip = (old_trip_count != new_trip_count*2);
1231 }
1232
1233 if (adjust_min_trip) {
1234 // Step 2: Adjust the trip limit if it is called for.
1235 // The adjustment amount is -stride. Need to make sure if the
1236 // adjustment underflows or overflows, then the main loop is skipped.
1237 Node* cmp = loop_end->cmp_node();
1238 assert(cmp->in(2) == limit, "sanity");
1239 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1240
1241 // Verify that policy_unroll result is still valid.
1242 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1243 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1244 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1245
1246 if (limit->is_Con()) {
1247 // The check in policy_unroll and the assert above guarantee
1248 // no underflow if limit is constant.
1249 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1250 set_ctrl(new_limit, C->root());
1251 } else {
1252 // Limit is not constant.
1253 if (loop_head->unrolled_count() == 1) { // only for first unroll
1254 // Separate limit by Opaque node in case it is an incremented
1255 // variable from previous loop to avoid using pre-incremented
1256 // value which could increase register pressure.
1257 // Otherwise reorg_offsets() optimization will create a separate
1258 // Opaque node for each use of trip-counter and as result
1259 // zero trip guard limit will be different from loop limit.
1260 assert(has_ctrl(opaq), "should have it");
1261 Node* opaq_ctrl = get_ctrl(opaq);
1262 limit = new (C) Opaque2Node( C, limit );
1263 register_new_node( limit, opaq_ctrl );
1264 }
1265 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) ||
1266 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) {
1267 // No underflow.
1268 new_limit = new (C) SubINode(limit, stride);
1269 } else {
1270 // (limit - stride) may underflow.
1271 // Clamp the adjustment value with MININT or MAXINT:
1272 //
1273 // new_limit = limit-stride
1274 // if (stride > 0)
1275 // new_limit = (limit < new_limit) ? MININT : new_limit;
1276 // else
1277 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1278 //
1279 BoolTest::mask bt = loop_end->test_trip();
1280 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1281 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1282 set_ctrl(adj_max, C->root());
1283 Node* old_limit = NULL;
1284 Node* adj_limit = NULL;
1285 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1286 if (loop_head->unrolled_count() > 1 &&
1287 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1288 limit->in(CMoveNode::IfTrue) == adj_max &&
1289 bol->as_Bool()->_test._test == bt &&
1290 bol->in(1)->Opcode() == Op_CmpI &&
1291 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1292 // Loop was unrolled before.
1293 // Optimize the limit to avoid nested CMove:
1294 // use original limit as old limit.
1295 old_limit = bol->in(1)->in(1);
1296 // Adjust previous adjusted limit.
1297 adj_limit = limit->in(CMoveNode::IfFalse);
1298 adj_limit = new (C) SubINode(adj_limit, stride);
1299 } else {
1300 old_limit = limit;
1301 adj_limit = new (C) SubINode(limit, stride);
1302 }
1303 assert(old_limit != NULL && adj_limit != NULL, "");
1304 register_new_node( adj_limit, ctrl ); // adjust amount
1305 Node* adj_cmp = new (C) CmpINode(old_limit, adj_limit);
1306 register_new_node( adj_cmp, ctrl );
1307 Node* adj_bool = new (C) BoolNode(adj_cmp, bt);
1308 register_new_node( adj_bool, ctrl );
1309 new_limit = new (C) CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1310 }
1311 register_new_node(new_limit, ctrl);
1312 }
1313 assert(new_limit != NULL, "");
1314 // Replace in loop test.
1315 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1316 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1317 // Don't need to create new test since only one user.
1318 _igvn.hash_delete(cmp);
1319 cmp->set_req(2, new_limit);
1320 } else {
1321 // Create new test since it is shared.
1322 Node* ctrl2 = loop_end->in(0);
1323 Node* cmp2 = cmp->clone();
1324 cmp2->set_req(2, new_limit);
1325 register_new_node(cmp2, ctrl2);
1326 Node* bol2 = loop_end->in(1)->clone();
1327 bol2->set_req(1, cmp2);
1328 register_new_node(bol2, ctrl2);
1329 _igvn.hash_delete(loop_end);
1330 loop_end->set_req(1, bol2);
1331 }
1332 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1333 // Make it a 1-trip test (means at least 2 trips).
1334
1335 // Guard test uses an 'opaque' node which is not shared. Hence I
1336 // can edit it's inputs directly. Hammer in the new limit for the
1337 // minimum-trip guard.
1338 assert(opaq->outcnt() == 1, "");
1339 _igvn.hash_delete(opaq);
1340 opaq->set_req(1, new_limit);
1341 }
1342
1343 // Adjust max trip count. The trip count is intentionally rounded
1344 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1345 // the main, unrolled, part of the loop will never execute as it is protected
1346 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1347 // and later determined that part of the unrolled loop was dead.
1348 loop_head->set_trip_count(old_trip_count / 2);
1349
1350 // Double the count of original iterations in the unrolled loop body.
1351 loop_head->double_unrolled_count();
1352
1353 } else { // LoopLimitCheck
1354
1355 // Adjust max trip count. The trip count is intentionally rounded
1356 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1357 // the main, unrolled, part of the loop will never execute as it is protected
1358 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1359 // and later determined that part of the unrolled loop was dead.
1360 loop_head->set_trip_count(loop_head->trip_count() / 2);
1361
1362 // Double the count of original iterations in the unrolled loop body.
1363 loop_head->double_unrolled_count();
1364
1365 // -----------
1366 // Step 2: Cut back the trip counter for an unroll amount of 2.
1367 // Loop will normally trip (limit - init)/stride_con. Since it's a
1368 // CountedLoop this is exact (stride divides limit-init exactly).
1369 // We are going to double the loop body, so we want to knock off any
1370 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
1371 Node *span = new (C) SubINode( limit, init );
1372 register_new_node( span, ctrl );
1373 Node *trip = new (C) DivINode( 0, span, stride );
1374 register_new_node( trip, ctrl );
1375 Node *mtwo = _igvn.intcon(-2);
1376 set_ctrl(mtwo, C->root());
1377 Node *rond = new (C) AndINode( trip, mtwo );
1378 register_new_node( rond, ctrl );
1379 Node *spn2 = new (C) MulINode( rond, stride );
1380 register_new_node( spn2, ctrl );
1381 new_limit = new (C) AddINode( spn2, init );
1382 register_new_node( new_limit, ctrl );
1383
1384 // Hammer in the new limit
1385 Node *ctrl2 = loop_end->in(0);
1386 Node *cmp2 = new (C) CmpINode( loop_head->incr(), new_limit );
1387 register_new_node( cmp2, ctrl2 );
1388 Node *bol2 = new (C) BoolNode( cmp2, loop_end->test_trip() );
1389 register_new_node( bol2, ctrl2 );
1390 _igvn.hash_delete(loop_end);
1391 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
1392
1393 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1394 // Make it a 1-trip test (means at least 2 trips).
1395 if( adjust_min_trip ) {
1396 assert( new_limit != NULL, "" );
1397 // Guard test uses an 'opaque' node which is not shared. Hence I
1398 // can edit it's inputs directly. Hammer in the new limit for the
1399 // minimum-trip guard.
1400 assert( opaq->outcnt() == 1, "" );
1401 _igvn.hash_delete(opaq);
1402 opaq->set_req(1, new_limit);
1403 }
1404 } // LoopLimitCheck
1405
1406 // ---------
1407 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1408 // represents the odd iterations; since the loop trips an even number of
1409 // times its backedge is never taken. Kill the backedge.
1410 uint dd = dom_depth(loop_head);
1411 clone_loop( loop, old_new, dd );
1412
1413 // Make backedges of the clone equal to backedges of the original.
1414 // Make the fall-in from the original come from the fall-out of the clone.
1415 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1416 Node* phi = loop_head->fast_out(j);
1417 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1418 Node *newphi = old_new[phi->_idx];
1419 _igvn.hash_delete( phi );
1420 _igvn.hash_delete( newphi );
1421
1422 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1423 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1424 phi ->set_req(LoopNode::LoopBackControl, C->top());
1425 }
1426 }
1427 Node *clone_head = old_new[loop_head->_idx];
1428 _igvn.hash_delete( clone_head );
1429 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1430 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1431 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1432 loop->_head = clone_head; // New loop header
1433
1434 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1435 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1436
1437 // Kill the clone's backedge
1438 Node *newcle = old_new[loop_end->_idx];
1439 _igvn.hash_delete( newcle );
1440 Node *one = _igvn.intcon(1);
1441 set_ctrl(one, C->root());
1442 newcle->set_req(1, one);
1443 // Force clone into same loop body
1444 uint max = loop->_body.size();
1445 for( uint k = 0; k < max; k++ ) {
1446 Node *old = loop->_body.at(k);
1447 Node *nnn = old_new[old->_idx];
1448 loop->_body.push(nnn);
1449 if (!has_ctrl(old))
1450 set_loop(nnn, loop);
1451 }
1452
1453 loop->record_for_igvn();
1454 }
1455
1456 //------------------------------do_maximally_unroll----------------------------
1457
1458 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1459 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1460 assert(cl->has_exact_trip_count(), "trip count is not exact");
1461 assert(cl->trip_count() > 0, "");
1462 #ifndef PRODUCT
1463 if (TraceLoopOpts) {
1464 tty->print("MaxUnroll %d ", cl->trip_count());
1465 loop->dump_head();
1466 }
1467 #endif
1468
1469 // If loop is tripping an odd number of times, peel odd iteration
1470 if ((cl->trip_count() & 1) == 1) {
1471 do_peeling(loop, old_new);
1472 }
1473
1474 // Now its tripping an even number of times remaining. Double loop body.
1475 // Do not adjust pre-guards; they are not needed and do not exist.
1476 if (cl->trip_count() > 0) {
1477 assert((cl->trip_count() & 1) == 0, "missed peeling");
1478 do_unroll(loop, old_new, false);
1479 }
1480 }
1481
1482 //------------------------------dominates_backedge---------------------------------
1483 // Returns true if ctrl is executed on every complete iteration
1484 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1485 assert(ctrl->is_CFG(), "must be control");
1486 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1487 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1488 }
1489
1490 //------------------------------adjust_limit-----------------------------------
1491 // Helper function for add_constraint().
1492 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1493 // Compute "I :: (limit-offset)/scale"
1494 Node *con = new (C) SubINode(rc_limit, offset);
1495 register_new_node(con, pre_ctrl);
1496 Node *X = new (C) DivINode(0, con, scale);
1497 register_new_node(X, pre_ctrl);
1498
1499 // Adjust loop limit
1500 loop_limit = (stride_con > 0)
1501 ? (Node*)(new (C) MinINode(loop_limit, X))
1502 : (Node*)(new (C) MaxINode(loop_limit, X));
1503 register_new_node(loop_limit, pre_ctrl);
1504 return loop_limit;
1505 }
1506
1507 //------------------------------add_constraint---------------------------------
1508 // Constrain the main loop iterations so the conditions:
1509 // low_limit <= scale_con * I + offset < upper_limit
1510 // always holds true. That is, either increase the number of iterations in
1511 // the pre-loop or the post-loop until the condition holds true in the main
1512 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1513 // stride and scale are constants (offset and limit often are).
1514 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 ) {
1515 // For positive stride, the pre-loop limit always uses a MAX function
1516 // and the main loop a MIN function. For negative stride these are
1517 // reversed.
1518
1519 // Also for positive stride*scale the affine function is increasing, so the
1520 // pre-loop must check for underflow and the post-loop for overflow.
1521 // Negative stride*scale reverses this; pre-loop checks for overflow and
1522 // post-loop for underflow.
1523
1524 Node *scale = _igvn.intcon(scale_con);
1525 set_ctrl(scale, C->root());
1526
1527 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1528 // The overflow limit: scale*I+offset < upper_limit
1529 // For main-loop compute
1530 // ( if (scale > 0) /* and stride > 0 */
1531 // I < (upper_limit-offset)/scale
1532 // else /* scale < 0 and stride < 0 */
1533 // I > (upper_limit-offset)/scale
1534 // )
1535 //
1536 // (upper_limit-offset) may overflow or underflow.
1537 // But it is fine since main loop will either have
1538 // less iterations or will be skipped in such case.
1539 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1540
1541 // The underflow limit: low_limit <= scale*I+offset.
1542 // For pre-loop compute
1543 // NOT(scale*I+offset >= low_limit)
1544 // scale*I+offset < low_limit
1545 // ( if (scale > 0) /* and stride > 0 */
1546 // I < (low_limit-offset)/scale
1547 // else /* scale < 0 and stride < 0 */
1548 // I > (low_limit-offset)/scale
1549 // )
1550
1551 if (low_limit->get_int() == -max_jint) {
1552 if (!RangeLimitCheck) return;
1553 // We need this guard when scale*pre_limit+offset >= limit
1554 // due to underflow. So we need execute pre-loop until
1555 // scale*I+offset >= min_int. But (min_int-offset) will
1556 // underflow when offset > 0 and X will be > original_limit
1557 // when stride > 0. To avoid it we replace positive offset with 0.
1558 //
1559 // Also (min_int+1 == -max_int) is used instead of min_int here
1560 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1561 Node* shift = _igvn.intcon(31);
1562 set_ctrl(shift, C->root());
1563 Node* sign = new (C) RShiftINode(offset, shift);
1564 register_new_node(sign, pre_ctrl);
1565 offset = new (C) AndINode(offset, sign);
1566 register_new_node(offset, pre_ctrl);
1567 } else {
1568 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1569 // The only problem we have here when offset == min_int
1570 // since (0-min_int) == min_int. It may be fine for stride > 0
1571 // but for stride < 0 X will be < original_limit. To avoid it
1572 // max(pre_limit, original_limit) is used in do_range_check().
1573 }
1574 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1575 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1576
1577 } else { // stride_con*scale_con < 0
1578 // For negative stride*scale pre-loop checks for overflow and
1579 // post-loop for underflow.
1580 //
1581 // The overflow limit: scale*I+offset < upper_limit
1582 // For pre-loop compute
1583 // NOT(scale*I+offset < upper_limit)
1584 // scale*I+offset >= upper_limit
1585 // scale*I+offset+1 > upper_limit
1586 // ( if (scale < 0) /* and stride > 0 */
1587 // I < (upper_limit-(offset+1))/scale
1588 // else /* scale > 0 and stride < 0 */
1589 // I > (upper_limit-(offset+1))/scale
1590 // )
1591 //
1592 // (upper_limit-offset-1) may underflow or overflow.
1593 // To avoid it min(pre_limit, original_limit) is used
1594 // in do_range_check() for stride > 0 and max() for < 0.
1595 Node *one = _igvn.intcon(1);
1596 set_ctrl(one, C->root());
1597
1598 Node *plus_one = new (C) AddINode(offset, one);
1599 register_new_node( plus_one, pre_ctrl );
1600 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1601 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1602
1603 if (low_limit->get_int() == -max_jint) {
1604 if (!RangeLimitCheck) return;
1605 // We need this guard when scale*main_limit+offset >= limit
1606 // due to underflow. So we need execute main-loop while
1607 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1608 // underflow when (offset+1) > 0 and X will be < main_limit
1609 // when scale < 0 (and stride > 0). To avoid it we replace
1610 // positive (offset+1) with 0.
1611 //
1612 // Also (min_int+1 == -max_int) is used instead of min_int here
1613 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1614 Node* shift = _igvn.intcon(31);
1615 set_ctrl(shift, C->root());
1616 Node* sign = new (C) RShiftINode(plus_one, shift);
1617 register_new_node(sign, pre_ctrl);
1618 plus_one = new (C) AndINode(plus_one, sign);
1619 register_new_node(plus_one, pre_ctrl);
1620 } else {
1621 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1622 // The only problem we have here when offset == max_int
1623 // since (max_int+1) == min_int and (0-min_int) == min_int.
1624 // But it is fine since main loop will either have
1625 // less iterations or will be skipped in such case.
1626 }
1627 // The underflow limit: low_limit <= scale*I+offset.
1628 // For main-loop compute
1629 // scale*I+offset+1 > low_limit
1630 // ( if (scale < 0) /* and stride > 0 */
1631 // I < (low_limit-(offset+1))/scale
1632 // else /* scale > 0 and stride < 0 */
1633 // I > (low_limit-(offset+1))/scale
1634 // )
1635
1636 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1637 }
1638 }
1639
1640
1641 //------------------------------is_scaled_iv---------------------------------
1642 // Return true if exp is a constant times an induction var
1643 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1644 if (exp == iv) {
1645 if (p_scale != NULL) {
1646 *p_scale = 1;
1647 }
1648 return true;
1649 }
1650 int opc = exp->Opcode();
1651 if (opc == Op_MulI) {
1652 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1653 if (p_scale != NULL) {
1654 *p_scale = exp->in(2)->get_int();
1655 }
1656 return true;
1657 }
1658 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1659 if (p_scale != NULL) {
1660 *p_scale = exp->in(1)->get_int();
1661 }
1662 return true;
1663 }
1664 } else if (opc == Op_LShiftI) {
1665 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1666 if (p_scale != NULL) {
1667 *p_scale = 1 << exp->in(2)->get_int();
1668 }
1669 return true;
1670 }
1671 }
1672 return false;
1673 }
1674
1675 //-----------------------------is_scaled_iv_plus_offset------------------------------
1676 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1677 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1678 if (is_scaled_iv(exp, iv, p_scale)) {
1679 if (p_offset != NULL) {
1680 Node *zero = _igvn.intcon(0);
1681 set_ctrl(zero, C->root());
1682 *p_offset = zero;
1683 }
1684 return true;
1685 }
1686 int opc = exp->Opcode();
1687 if (opc == Op_AddI) {
1688 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1689 if (p_offset != NULL) {
1690 *p_offset = exp->in(2);
1691 }
1692 return true;
1693 }
1694 if (exp->in(2)->is_Con()) {
1695 Node* offset2 = NULL;
1696 if (depth < 2 &&
1697 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1698 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1699 if (p_offset != NULL) {
1700 Node *ctrl_off2 = get_ctrl(offset2);
1701 Node* offset = new (C) AddINode(offset2, exp->in(2));
1702 register_new_node(offset, ctrl_off2);
1703 *p_offset = offset;
1704 }
1705 return true;
1706 }
1707 }
1708 } else if (opc == Op_SubI) {
1709 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1710 if (p_offset != NULL) {
1711 Node *zero = _igvn.intcon(0);
1712 set_ctrl(zero, C->root());
1713 Node *ctrl_off = get_ctrl(exp->in(2));
1714 Node* offset = new (C) SubINode(zero, exp->in(2));
1715 register_new_node(offset, ctrl_off);
1716 *p_offset = offset;
1717 }
1718 return true;
1719 }
1720 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1721 if (p_offset != NULL) {
1722 *p_scale *= -1;
1723 *p_offset = exp->in(1);
1724 }
1725 return true;
1726 }
1727 }
1728 return false;
1729 }
1730
1731 //------------------------------do_range_check---------------------------------
1732 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1733 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1734 #ifndef PRODUCT
1735 if (PrintOpto && VerifyLoopOptimizations) {
1736 tty->print("Range Check Elimination ");
1737 loop->dump_head();
1738 } else if (TraceLoopOpts) {
1739 tty->print("RangeCheck ");
1740 loop->dump_head();
1741 }
1742 #endif
1743 assert(RangeCheckElimination, "");
1744 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1745 assert(cl->is_main_loop(), "");
1746
1747 // protect against stride not being a constant
1748 if (!cl->stride_is_con())
1749 return;
1750
1751 // Find the trip counter; we are iteration splitting based on it
1752 Node *trip_counter = cl->phi();
1753 // Find the main loop limit; we will trim it's iterations
1754 // to not ever trip end tests
1755 Node *main_limit = cl->limit();
1756
1757 // Need to find the main-loop zero-trip guard
1758 Node *ctrl = cl->in(LoopNode::EntryControl);
1759 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1760 Node *iffm = ctrl->in(0);
1761 assert(iffm->Opcode() == Op_If, "");
1762 Node *bolzm = iffm->in(1);
1763 assert(bolzm->Opcode() == Op_Bool, "");
1764 Node *cmpzm = bolzm->in(1);
1765 assert(cmpzm->is_Cmp(), "");
1766 Node *opqzm = cmpzm->in(2);
1767 // Can not optimize a loop if zero-trip Opaque1 node is optimized
1768 // away and then another round of loop opts attempted.
1769 if (opqzm->Opcode() != Op_Opaque1)
1770 return;
1771 assert(opqzm->in(1) == main_limit, "do not understand situation");
1772
1773 // Find the pre-loop limit; we will expand it's iterations to
1774 // not ever trip low tests.
1775 Node *p_f = iffm->in(0);
1776 assert(p_f->Opcode() == Op_IfFalse, "");
1777 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1778 assert(pre_end->loopnode()->is_pre_loop(), "");
1779 Node *pre_opaq1 = pre_end->limit();
1780 // Occasionally it's possible for a pre-loop Opaque1 node to be
1781 // optimized away and then another round of loop opts attempted.
1782 // We can not optimize this particular loop in that case.
1783 if (pre_opaq1->Opcode() != Op_Opaque1)
1784 return;
1785 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1786 Node *pre_limit = pre_opaq->in(1);
1787
1788 // Where do we put new limit calculations
1789 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1790
1791 // Ensure the original loop limit is available from the
1792 // pre-loop Opaque1 node.
1793 Node *orig_limit = pre_opaq->original_loop_limit();
1794 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1795 return;
1796
1797 // Must know if its a count-up or count-down loop
1798
1799 int stride_con = cl->stride_con();
1800 Node *zero = _igvn.intcon(0);
1801 Node *one = _igvn.intcon(1);
1802 // Use symmetrical int range [-max_jint,max_jint]
1803 Node *mini = _igvn.intcon(-max_jint);
1804 set_ctrl(zero, C->root());
1805 set_ctrl(one, C->root());
1806 set_ctrl(mini, C->root());
1807
1808 // Range checks that do not dominate the loop backedge (ie.
1809 // conditionally executed) can lengthen the pre loop limit beyond
1810 // the original loop limit. To prevent this, the pre limit is
1811 // (for stride > 0) MINed with the original loop limit (MAXed
1812 // stride < 0) when some range_check (rc) is conditionally
1813 // executed.
1814 bool conditional_rc = false;
1815
1816 // Check loop body for tests of trip-counter plus loop-invariant vs
1817 // loop-invariant.
1818 for( uint i = 0; i < loop->_body.size(); i++ ) {
1819 Node *iff = loop->_body[i];
1820 if( iff->Opcode() == Op_If ) { // Test?
1821
1822 // Test is an IfNode, has 2 projections. If BOTH are in the loop
1823 // we need loop unswitching instead of iteration splitting.
1824 Node *exit = loop->is_loop_exit(iff);
1825 if( !exit ) continue;
1826 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1827
1828 // Get boolean condition to test
1829 Node *i1 = iff->in(1);
1830 if( !i1->is_Bool() ) continue;
1831 BoolNode *bol = i1->as_Bool();
1832 BoolTest b_test = bol->_test;
1833 // Flip sense of test if exit condition is flipped
1834 if( flip )
1835 b_test = b_test.negate();
1836
1837 // Get compare
1838 Node *cmp = bol->in(1);
1839
1840 // Look for trip_counter + offset vs limit
1841 Node *rc_exp = cmp->in(1);
1842 Node *limit = cmp->in(2);
1843 jint scale_con= 1; // Assume trip counter not scaled
1844
1845 Node *limit_c = get_ctrl(limit);
1846 if( loop->is_member(get_loop(limit_c) ) ) {
1847 // Compare might have operands swapped; commute them
1848 b_test = b_test.commute();
1849 rc_exp = cmp->in(2);
1850 limit = cmp->in(1);
1851 limit_c = get_ctrl(limit);
1852 if( loop->is_member(get_loop(limit_c) ) )
1853 continue; // Both inputs are loop varying; cannot RCE
1854 }
1855 // Here we know 'limit' is loop invariant
1856
1857 // 'limit' maybe pinned below the zero trip test (probably from a
1858 // previous round of rce), in which case, it can't be used in the
1859 // zero trip test expression which must occur before the zero test's if.
1860 if( limit_c == ctrl ) {
1861 continue; // Don't rce this check but continue looking for other candidates.
1862 }
1863
1864 // Check for scaled induction variable plus an offset
1865 Node *offset = NULL;
1866
1867 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1868 continue;
1869 }
1870
1871 Node *offset_c = get_ctrl(offset);
1872 if( loop->is_member( get_loop(offset_c) ) )
1873 continue; // Offset is not really loop invariant
1874 // Here we know 'offset' is loop invariant.
1875
1876 // As above for the 'limit', the 'offset' maybe pinned below the
1877 // zero trip test.
1878 if( offset_c == ctrl ) {
1879 continue; // Don't rce this check but continue looking for other candidates.
1880 }
1881 #ifdef ASSERT
1882 if (TraceRangeLimitCheck) {
1883 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
1884 bol->dump(2);
1885 }
1886 #endif
1887 // At this point we have the expression as:
1888 // scale_con * trip_counter + offset :: limit
1889 // where scale_con, offset and limit are loop invariant. Trip_counter
1890 // monotonically increases by stride_con, a constant. Both (or either)
1891 // stride_con and scale_con can be negative which will flip about the
1892 // sense of the test.
1893
1894 // Adjust pre and main loop limits to guard the correct iteration set
1895 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1896 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1897 // The underflow and overflow limits: 0 <= scale*I+offset < limit
1898 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
1899 if (!conditional_rc) {
1900 // (0-offset)/scale could be outside of loop iterations range.
1901 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1902 }
1903 } else {
1904 #ifndef PRODUCT
1905 if( PrintOpto )
1906 tty->print_cr("missed RCE opportunity");
1907 #endif
1908 continue; // In release mode, ignore it
1909 }
1910 } else { // Otherwise work on normal compares
1911 switch( b_test._test ) {
1912 case BoolTest::gt:
1913 // Fall into GE case
1914 case BoolTest::ge:
1915 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
1916 scale_con = -scale_con;
1917 offset = new (C) SubINode( zero, offset );
1918 register_new_node( offset, pre_ctrl );
1919 limit = new (C) SubINode( zero, limit );
1920 register_new_node( limit, pre_ctrl );
1921 // Fall into LE case
1922 case BoolTest::le:
1923 if (b_test._test != BoolTest::gt) {
1924 // Convert X <= Y to X < Y+1
1925 limit = new (C) AddINode( limit, one );
1926 register_new_node( limit, pre_ctrl );
1927 }
1928 // Fall into LT case
1929 case BoolTest::lt:
1930 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
1931 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
1932 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
1933 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
1934 if (!conditional_rc) {
1935 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
1936 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
1937 // still be outside of loop range.
1938 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1939 }
1940 break;
1941 default:
1942 #ifndef PRODUCT
1943 if( PrintOpto )
1944 tty->print_cr("missed RCE opportunity");
1945 #endif
1946 continue; // Unhandled case
1947 }
1948 }
1949
1950 // Kill the eliminated test
1951 C->set_major_progress();
1952 Node *kill_con = _igvn.intcon( 1-flip );
1953 set_ctrl(kill_con, C->root());
1954 _igvn.replace_input_of(iff, 1, kill_con);
1955 // Find surviving projection
1956 assert(iff->is_If(), "");
1957 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
1958 // Find loads off the surviving projection; remove their control edge
1959 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1960 Node* cd = dp->fast_out(i); // Control-dependent node
1961 if( cd->is_Load() ) { // Loads can now float around in the loop
1962 // Allow the load to float around in the loop, or before it
1963 // but NOT before the pre-loop.
1964 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
1965 --i;
1966 --imax;
1967 }
1968 }
1969
1970 } // End of is IF
1971
1972 }
1973
1974 // Update loop limits
1975 if (conditional_rc) {
1976 pre_limit = (stride_con > 0) ? (Node*)new (C) MinINode(pre_limit, orig_limit)
1977 : (Node*)new (C) MaxINode(pre_limit, orig_limit);
1978 register_new_node(pre_limit, pre_ctrl);
1979 }
1980 _igvn.hash_delete(pre_opaq);
1981 pre_opaq->set_req(1, pre_limit);
1982
1983 // Note:: we are making the main loop limit no longer precise;
1984 // need to round up based on stride.
1985 cl->set_nonexact_trip_count();
1986 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
1987 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
1988 // Hopefully, compiler will optimize for powers of 2.
1989 Node *ctrl = get_ctrl(main_limit);
1990 Node *stride = cl->stride();
1991 Node *init = cl->init_trip();
1992 Node *span = new (C) SubINode(main_limit,init);
1993 register_new_node(span,ctrl);
1994 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
1995 Node *add = new (C) AddINode(span,rndup);
1996 register_new_node(add,ctrl);
1997 Node *div = new (C) DivINode(0,add,stride);
1998 register_new_node(div,ctrl);
1999 Node *mul = new (C) MulINode(div,stride);
2000 register_new_node(mul,ctrl);
2001 Node *newlim = new (C) AddINode(mul,init);
2002 register_new_node(newlim,ctrl);
2003 main_limit = newlim;
2004 }
2005
2006 Node *main_cle = cl->loopexit();
2007 Node *main_bol = main_cle->in(1);
2008 // Hacking loop bounds; need private copies of exit test
2009 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2010 _igvn.hash_delete(main_cle);
2011 main_bol = main_bol->clone();// Clone a private BoolNode
2012 register_new_node( main_bol, main_cle->in(0) );
2013 main_cle->set_req(1,main_bol);
2014 }
2015 Node *main_cmp = main_bol->in(1);
2016 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2017 _igvn.hash_delete(main_bol);
2018 main_cmp = main_cmp->clone();// Clone a private CmpNode
2019 register_new_node( main_cmp, main_cle->in(0) );
2020 main_bol->set_req(1,main_cmp);
2021 }
2022 // Hack the now-private loop bounds
2023 _igvn.replace_input_of(main_cmp, 2, main_limit);
2024 // The OpaqueNode is unshared by design
2025 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2026 _igvn.replace_input_of(opqzm, 1, main_limit);
2027 }
2028
2029 //------------------------------DCE_loop_body----------------------------------
2030 // Remove simplistic dead code from loop body
2031 void IdealLoopTree::DCE_loop_body() {
2032 for( uint i = 0; i < _body.size(); i++ )
2033 if( _body.at(i)->outcnt() == 0 )
2034 _body.map( i--, _body.pop() );
2035 }
2036
2037
2038 //------------------------------adjust_loop_exit_prob--------------------------
2039 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2040 // Replace with a 1-in-10 exit guess.
2041 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2042 Node *test = tail();
2043 while( test != _head ) {
2044 uint top = test->Opcode();
2045 if( top == Op_IfTrue || top == Op_IfFalse ) {
2046 int test_con = ((ProjNode*)test)->_con;
2047 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2048 IfNode *iff = test->in(0)->as_If();
2049 if( iff->outcnt() == 2 ) { // Ignore dead tests
2050 Node *bol = iff->in(1);
2051 if( bol && bol->req() > 1 && bol->in(1) &&
2052 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2053 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2054 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2055 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2056 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2057 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2058 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2059 return; // Allocation loops RARELY take backedge
2060 // Find the OTHER exit path from the IF
2061 Node* ex = iff->proj_out(1-test_con);
2062 float p = iff->_prob;
2063 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2064 if( top == Op_IfTrue ) {
2065 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2066 iff->_prob = PROB_STATIC_FREQUENT;
2067 }
2068 } else {
2069 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2070 iff->_prob = PROB_STATIC_INFREQUENT;
2071 }
2072 }
2073 }
2074 }
2075 }
2076 test = phase->idom(test);
2077 }
2078 }
2079
2080
2081 //------------------------------policy_do_remove_empty_loop--------------------
2082 // Micro-benchmark spamming. Policy is to always remove empty loops.
2083 // The 'DO' part is to replace the trip counter with the value it will
2084 // have on the last iteration. This will break the loop.
2085 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2086 // Minimum size must be empty loop
2087 if (_body.size() > EMPTY_LOOP_SIZE)
2088 return false;
2089
2090 if (!_head->is_CountedLoop())
2091 return false; // Dead loop
2092 CountedLoopNode *cl = _head->as_CountedLoop();
2093 if (!cl->is_valid_counted_loop())
2094 return false; // Malformed loop
2095 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2096 return false; // Infinite loop
2097
2098 #ifdef ASSERT
2099 // Ensure only one phi which is the iv.
2100 Node* iv = NULL;
2101 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2102 Node* n = cl->fast_out(i);
2103 if (n->Opcode() == Op_Phi) {
2104 assert(iv == NULL, "Too many phis" );
2105 iv = n;
2106 }
2107 }
2108 assert(iv == cl->phi(), "Wrong phi" );
2109 #endif
2110
2111 // main and post loops have explicitly created zero trip guard
2112 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2113 if (needs_guard) {
2114 // Skip guard if values not overlap.
2115 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2116 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2117 int stride_con = cl->stride_con();
2118 if (stride_con > 0) {
2119 needs_guard = (init_t->_hi >= limit_t->_lo);
2120 } else {
2121 needs_guard = (init_t->_lo <= limit_t->_hi);
2122 }
2123 }
2124 if (needs_guard) {
2125 // Check for an obvious zero trip guard.
2126 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2127 if (inctrl->Opcode() == Op_IfTrue) {
2128 // The test should look like just the backedge of a CountedLoop
2129 Node* iff = inctrl->in(0);
2130 if (iff->is_If()) {
2131 Node* bol = iff->in(1);
2132 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
2133 Node* cmp = bol->in(1);
2134 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
2135 needs_guard = false;
2136 }
2137 }
2138 }
2139 }
2140 }
2141
2142 #ifndef PRODUCT
2143 if (PrintOpto) {
2144 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2145 this->dump_head();
2146 } else if (TraceLoopOpts) {
2147 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2148 this->dump_head();
2149 }
2150 #endif
2151
2152 if (needs_guard) {
2153 // Peel the loop to ensure there's a zero trip guard
2154 Node_List old_new;
2155 phase->do_peeling(this, old_new);
2156 }
2157
2158 // Replace the phi at loop head with the final value of the last
2159 // iteration. Then the CountedLoopEnd will collapse (backedge never
2160 // taken) and all loop-invariant uses of the exit values will be correct.
2161 Node *phi = cl->phi();
2162 Node *exact_limit = phase->exact_limit(this);
2163 if (exact_limit != cl->limit()) {
2164 // We also need to replace the original limit to collapse loop exit.
2165 Node* cmp = cl->loopexit()->cmp_node();
2166 assert(cl->limit() == cmp->in(2), "sanity");
2167 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2168 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2169 }
2170 // Note: the final value after increment should not overflow since
2171 // counted loop has limit check predicate.
2172 Node *final = new (phase->C) SubINode( exact_limit, cl->stride() );
2173 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2174 phase->_igvn.replace_node(phi,final);
2175 phase->C->set_major_progress();
2176 return true;
2177 }
2178
2179 //------------------------------policy_do_one_iteration_loop-------------------
2180 // Convert one iteration loop into normal code.
2181 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2182 if (!_head->as_Loop()->is_valid_counted_loop())
2183 return false; // Only for counted loop
2184
2185 CountedLoopNode *cl = _head->as_CountedLoop();
2186 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2187 return false;
2188 }
2189
2190 #ifndef PRODUCT
2191 if(TraceLoopOpts) {
2192 tty->print("OneIteration ");
2193 this->dump_head();
2194 }
2195 #endif
2196
2197 Node *init_n = cl->init_trip();
2198 #ifdef ASSERT
2199 // Loop boundaries should be constant since trip count is exact.
2200 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2201 #endif
2202 // Replace the phi at loop head with the value of the init_trip.
2203 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2204 // and all loop-invariant uses of the exit values will be correct.
2205 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2206 phase->C->set_major_progress();
2207 return true;
2208 }
2209
2210 //=============================================================================
2211 //------------------------------iteration_split_impl---------------------------
2212 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2213 // Compute exact loop trip count if possible.
2214 compute_exact_trip_count(phase);
2215
2216 // Convert one iteration loop into normal code.
2217 if (policy_do_one_iteration_loop(phase))
2218 return true;
2219
2220 // Check and remove empty loops (spam micro-benchmarks)
2221 if (policy_do_remove_empty_loop(phase))
2222 return true; // Here we removed an empty loop
2223
2224 bool should_peel = policy_peeling(phase); // Should we peel?
2225
2226 bool should_unswitch = policy_unswitching(phase);
2227
2228 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2229 // This removes loop-invariant tests (usually null checks).
2230 if (!_head->is_CountedLoop()) { // Non-counted loop
2231 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2232 // Partial peel succeeded so terminate this round of loop opts
2233 return false;
2234 }
2235 if (should_peel) { // Should we peel?
2236 #ifndef PRODUCT
2237 if (PrintOpto) tty->print_cr("should_peel");
2238 #endif
2239 phase->do_peeling(this,old_new);
2240 } else if (should_unswitch) {
2241 phase->do_unswitching(this, old_new);
2242 }
2243 return true;
2244 }
2245 CountedLoopNode *cl = _head->as_CountedLoop();
2246
2247 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2248
2249 // Do nothing special to pre- and post- loops
2250 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2251
2252 // Compute loop trip count from profile data
2253 compute_profile_trip_cnt(phase);
2254
2255 // Before attempting fancy unrolling, RCE or alignment, see if we want
2256 // to completely unroll this loop or do loop unswitching.
2257 if (cl->is_normal_loop()) {
2258 if (should_unswitch) {
2259 phase->do_unswitching(this, old_new);
2260 return true;
2261 }
2262 bool should_maximally_unroll = policy_maximally_unroll(phase);
2263 if (should_maximally_unroll) {
2264 // Here we did some unrolling and peeling. Eventually we will
2265 // completely unroll this loop and it will no longer be a loop.
2266 phase->do_maximally_unroll(this,old_new);
2267 return true;
2268 }
2269 }
2270
2271 // Skip next optimizations if running low on nodes. Note that
2272 // policy_unswitching and policy_maximally_unroll have this check.
2273 uint nodes_left = MaxNodeLimit - (uint) phase->C->live_nodes();
2274 if ((2 * _body.size()) > nodes_left) {
2275 return true;
2276 }
2277
2278 // Counted loops may be peeled, may need some iterations run up
2279 // front for RCE, and may want to align loop refs to a cache
2280 // line. Thus we clone a full loop up front whose trip count is
2281 // at least 1 (if peeling), but may be several more.
2282
2283 // The main loop will start cache-line aligned with at least 1
2284 // iteration of the unrolled body (zero-trip test required) and
2285 // will have some range checks removed.
2286
2287 // A post-loop will finish any odd iterations (leftover after
2288 // unrolling), plus any needed for RCE purposes.
2289
2290 bool should_unroll = policy_unroll(phase);
2291
2292 bool should_rce = policy_range_check(phase);
2293
2294 bool should_align = policy_align(phase);
2295
2296 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2297 // need a pre-loop. We may still need to peel an initial iteration but
2298 // we will not be needing an unknown number of pre-iterations.
2299 //
2300 // Basically, if may_rce_align reports FALSE first time through,
2301 // we will not be able to later do RCE or Aligning on this loop.
2302 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2303
2304 // If we have any of these conditions (RCE, alignment, unrolling) met, then
2305 // we switch to the pre-/main-/post-loop model. This model also covers
2306 // peeling.
2307 if (should_rce || should_align || should_unroll) {
2308 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
2309 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2310
2311 // Adjust the pre- and main-loop limits to let the pre and post loops run
2312 // with full checks, but the main-loop with no checks. Remove said
2313 // checks from the main body.
2314 if (should_rce)
2315 phase->do_range_check(this,old_new);
2316
2317 // Double loop body for unrolling. Adjust the minimum-trip test (will do
2318 // twice as many iterations as before) and the main body limit (only do
2319 // an even number of trips). If we are peeling, we might enable some RCE
2320 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2321 // peeling.
2322 if (should_unroll && !should_peel)
2323 phase->do_unroll(this,old_new, true);
2324
2325 // Adjust the pre-loop limits to align the main body
2326 // iterations.
2327 if (should_align)
2328 Unimplemented();
2329
2330 } else { // Else we have an unchanged counted loop
2331 if (should_peel) // Might want to peel but do nothing else
2332 phase->do_peeling(this,old_new);
2333 }
2334 return true;
2335 }
2336
2337
2338 //=============================================================================
2339 //------------------------------iteration_split--------------------------------
2340 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2341 // Recursively iteration split nested loops
2342 if (_child && !_child->iteration_split(phase, old_new))
2343 return false;
2344
2345 // Clean out prior deadwood
2346 DCE_loop_body();
2347
2348
2349 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2350 // Replace with a 1-in-10 exit guess.
2351 if (_parent /*not the root loop*/ &&
2352 !_irreducible &&
2353 // Also ignore the occasional dead backedge
2354 !tail()->is_top()) {
2355 adjust_loop_exit_prob(phase);
2356 }
2357
2358 // Gate unrolling, RCE and peeling efforts.
2359 if (!_child && // If not an inner loop, do not split
2360 !_irreducible &&
2361 _allow_optimizations &&
2362 !tail()->is_top()) { // Also ignore the occasional dead backedge
2363 if (!_has_call) {
2364 if (!iteration_split_impl(phase, old_new)) {
2365 return false;
2366 }
2367 } else if (policy_unswitching(phase)) {
2368 phase->do_unswitching(this, old_new);
2369 }
2370 }
2371
2372 // Minor offset re-organization to remove loop-fallout uses of
2373 // trip counter when there was no major reshaping.
2374 phase->reorg_offsets(this);
2375
2376 if (_next && !_next->iteration_split(phase, old_new))
2377 return false;
2378 return true;
2379 }
2380
2381
2382 //=============================================================================
2383 // Process all the loops in the loop tree and replace any fill
2384 // patterns with an intrisc version.
2385 bool PhaseIdealLoop::do_intrinsify_fill() {
2386 bool changed = false;
2387 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2388 IdealLoopTree* lpt = iter.current();
2389 changed |= intrinsify_fill(lpt);
2390 }
2391 return changed;
2392 }
2393
2394
2395 // Examine an inner loop looking for a a single store of an invariant
2396 // value in a unit stride loop,
2397 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2398 Node*& shift, Node*& con) {
2399 const char* msg = NULL;
2400 Node* msg_node = NULL;
2401
2402 store_value = NULL;
2403 con = NULL;
2404 shift = NULL;
2405
2406 // Process the loop looking for stores. If there are multiple
2407 // stores or extra control flow give at this point.
2408 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2409 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2410 Node* n = lpt->_body.at(i);
2411 if (n->outcnt() == 0) continue; // Ignore dead
2412 if (n->is_Store()) {
2413 if (store != NULL) {
2414 msg = "multiple stores";
2415 break;
2416 }
2417 int opc = n->Opcode();
2418 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
2419 msg = "oop fills not handled";
2420 break;
2421 }
2422 Node* value = n->in(MemNode::ValueIn);
2423 if (!lpt->is_invariant(value)) {
2424 msg = "variant store value";
2425 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2426 msg = "not array address";
2427 }
2428 store = n;
2429 store_value = value;
2430 } else if (n->is_If() && n != head->loopexit()) {
2431 msg = "extra control flow";
2432 msg_node = n;
2433 }
2434 }
2435
2436 if (store == NULL) {
2437 // No store in loop
2438 return false;
2439 }
2440
2441 if (msg == NULL && head->stride_con() != 1) {
2442 // could handle negative strides too
2443 if (head->stride_con() < 0) {
2444 msg = "negative stride";
2445 } else {
2446 msg = "non-unit stride";
2447 }
2448 }
2449
2450 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2451 msg = "can't handle store address";
2452 msg_node = store->in(MemNode::Address);
2453 }
2454
2455 if (msg == NULL &&
2456 (!store->in(MemNode::Memory)->is_Phi() ||
2457 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2458 msg = "store memory isn't proper phi";
2459 msg_node = store->in(MemNode::Memory);
2460 }
2461
2462 // Make sure there is an appropriate fill routine
2463 BasicType t = store->as_Mem()->memory_type();
2464 const char* fill_name;
2465 if (msg == NULL &&
2466 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2467 msg = "unsupported store";
2468 msg_node = store;
2469 }
2470
2471 if (msg != NULL) {
2472 #ifndef PRODUCT
2473 if (TraceOptimizeFill) {
2474 tty->print_cr("not fill intrinsic candidate: %s", msg);
2475 if (msg_node != NULL) msg_node->dump();
2476 }
2477 #endif
2478 return false;
2479 }
2480
2481 // Make sure the address expression can be handled. It should be
2482 // head->phi * elsize + con. head->phi might have a ConvI2L.
2483 Node* elements[4];
2484 Node* conv = NULL;
2485 bool found_index = false;
2486 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2487 for (int e = 0; e < count; e++) {
2488 Node* n = elements[e];
2489 if (n->is_Con() && con == NULL) {
2490 con = n;
2491 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2492 Node* value = n->in(1);
2493 #ifdef _LP64
2494 if (value->Opcode() == Op_ConvI2L) {
2495 conv = value;
2496 value = value->in(1);
2497 }
2498 #endif
2499 if (value != head->phi()) {
2500 msg = "unhandled shift in address";
2501 } else {
2502 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2503 msg = "scale doesn't match";
2504 } else {
2505 found_index = true;
2506 shift = n;
2507 }
2508 }
2509 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2510 if (n->in(1) == head->phi()) {
2511 found_index = true;
2512 conv = n;
2513 } else {
2514 msg = "unhandled input to ConvI2L";
2515 }
2516 } else if (n == head->phi()) {
2517 // no shift, check below for allowed cases
2518 found_index = true;
2519 } else {
2520 msg = "unhandled node in address";
2521 msg_node = n;
2522 }
2523 }
2524
2525 if (count == -1) {
2526 msg = "malformed address expression";
2527 msg_node = store;
2528 }
2529
2530 if (!found_index) {
2531 msg = "missing use of index";
2532 }
2533
2534 // byte sized items won't have a shift
2535 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2536 msg = "can't find shift";
2537 msg_node = store;
2538 }
2539
2540 if (msg != NULL) {
2541 #ifndef PRODUCT
2542 if (TraceOptimizeFill) {
2543 tty->print_cr("not fill intrinsic: %s", msg);
2544 if (msg_node != NULL) msg_node->dump();
2545 }
2546 #endif
2547 return false;
2548 }
2549
2550 // No make sure all the other nodes in the loop can be handled
2551 VectorSet ok(Thread::current()->resource_area());
2552
2553 // store related values are ok
2554 ok.set(store->_idx);
2555 ok.set(store->in(MemNode::Memory)->_idx);
2556
2557 // Loop structure is ok
2558 ok.set(head->_idx);
2559 ok.set(head->loopexit()->_idx);
2560 ok.set(head->phi()->_idx);
2561 ok.set(head->incr()->_idx);
2562 ok.set(head->loopexit()->cmp_node()->_idx);
2563 ok.set(head->loopexit()->in(1)->_idx);
2564
2565 // Address elements are ok
2566 if (con) ok.set(con->_idx);
2567 if (shift) ok.set(shift->_idx);
2568 if (conv) ok.set(conv->_idx);
2569
2570 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2571 Node* n = lpt->_body.at(i);
2572 if (n->outcnt() == 0) continue; // Ignore dead
2573 if (ok.test(n->_idx)) continue;
2574 // Backedge projection is ok
2575 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue;
2576 if (!n->is_AddP()) {
2577 msg = "unhandled node";
2578 msg_node = n;
2579 break;
2580 }
2581 }
2582
2583 // Make sure no unexpected values are used outside the loop
2584 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2585 Node* n = lpt->_body.at(i);
2586 // These values can be replaced with other nodes if they are used
2587 // outside the loop.
2588 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2589 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2590 Node* use = iter.get();
2591 if (!lpt->_body.contains(use)) {
2592 msg = "node is used outside loop";
2593 // lpt->_body.dump();
2594 msg_node = n;
2595 break;
2596 }
2597 }
2598 }
2599
2600 #ifdef ASSERT
2601 if (TraceOptimizeFill) {
2602 if (msg != NULL) {
2603 tty->print_cr("no fill intrinsic: %s", msg);
2604 if (msg_node != NULL) msg_node->dump();
2605 } else {
2606 tty->print_cr("fill intrinsic for:");
2607 }
2608 store->dump();
2609 if (Verbose) {
2610 lpt->_body.dump();
2611 }
2612 }
2613 #endif
2614
2615 return msg == NULL;
2616 }
2617
2618
2619
2620 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2621 // Only for counted inner loops
2622 if (!lpt->is_counted() || !lpt->is_inner()) {
2623 return false;
2624 }
2625
2626 // Must have constant stride
2627 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2628 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
2629 return false;
2630 }
2631
2632 // Check that the body only contains a store of a loop invariant
2633 // value that is indexed by the loop phi.
2634 Node* store = NULL;
2635 Node* store_value = NULL;
2636 Node* shift = NULL;
2637 Node* offset = NULL;
2638 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2639 return false;
2640 }
2641
2642 #ifndef PRODUCT
2643 if (TraceLoopOpts) {
2644 tty->print("ArrayFill ");
2645 lpt->dump_head();
2646 }
2647 #endif
2648
2649 // Now replace the whole loop body by a call to a fill routine that
2650 // covers the same region as the loop.
2651 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2652
2653 // Build an expression for the beginning of the copy region
2654 Node* index = head->init_trip();
2655 #ifdef _LP64
2656 index = new (C) ConvI2LNode(index);
2657 _igvn.register_new_node_with_optimizer(index);
2658 #endif
2659 if (shift != NULL) {
2660 // byte arrays don't require a shift but others do.
2661 index = new (C) LShiftXNode(index, shift->in(2));
2662 _igvn.register_new_node_with_optimizer(index);
2663 }
2664 index = new (C) AddPNode(base, base, index);
2665 _igvn.register_new_node_with_optimizer(index);
2666 Node* from = new (C) AddPNode(base, index, offset);
2667 _igvn.register_new_node_with_optimizer(from);
2668 // Compute the number of elements to copy
2669 Node* len = new (C) SubINode(head->limit(), head->init_trip());
2670 _igvn.register_new_node_with_optimizer(len);
2671
2672 BasicType t = store->as_Mem()->memory_type();
2673 bool aligned = false;
2674 if (offset != NULL && head->init_trip()->is_Con()) {
2675 int element_size = type2aelembytes(t);
2676 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2677 }
2678
2679 // Build a call to the fill routine
2680 const char* fill_name;
2681 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2682 assert(fill != NULL, "what?");
2683
2684 // Convert float/double to int/long for fill routines
2685 if (t == T_FLOAT) {
2686 store_value = new (C) MoveF2INode(store_value);
2687 _igvn.register_new_node_with_optimizer(store_value);
2688 } else if (t == T_DOUBLE) {
2689 store_value = new (C) MoveD2LNode(store_value);
2690 _igvn.register_new_node_with_optimizer(store_value);
2691 }
2692
2693 Node* mem_phi = store->in(MemNode::Memory);
2694 Node* result_ctrl;
2695 Node* result_mem;
2696 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2697 CallLeafNode *call = new (C) CallLeafNoFPNode(call_type, fill,
2698 fill_name, TypeAryPtr::get_array_body_type(t));
2699 call->init_req(TypeFunc::Parms+0, from);
2700 call->init_req(TypeFunc::Parms+1, store_value);
2701 #ifdef _LP64
2702 len = new (C) ConvI2LNode(len);
2703 _igvn.register_new_node_with_optimizer(len);
2704 #endif
2705 call->init_req(TypeFunc::Parms+2, len);
2706 #ifdef _LP64
2707 call->init_req(TypeFunc::Parms+3, C->top());
2708 #endif
2709 call->init_req( TypeFunc::Control, head->init_control());
2710 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o
2711 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) );
2712 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) );
2713 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) );
2714 _igvn.register_new_node_with_optimizer(call);
2715 result_ctrl = new (C) ProjNode(call,TypeFunc::Control);
2716 _igvn.register_new_node_with_optimizer(result_ctrl);
2717 result_mem = new (C) ProjNode(call,TypeFunc::Memory);
2718 _igvn.register_new_node_with_optimizer(result_mem);
2719
2720 /* Disable following optimization until proper fix (add missing checks).
2721
2722 // If this fill is tightly coupled to an allocation and overwrites
2723 // the whole body, allow it to take over the zeroing.
2724 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2725 if (alloc != NULL && alloc->is_AllocateArray()) {
2726 Node* length = alloc->as_AllocateArray()->Ideal_length();
2727 if (head->limit() == length &&
2728 head->init_trip() == _igvn.intcon(0)) {
2729 if (TraceOptimizeFill) {
2730 tty->print_cr("Eliminated zeroing in allocation");
2731 }
2732 alloc->maybe_set_complete(&_igvn);
2733 } else {
2734 #ifdef ASSERT
2735 if (TraceOptimizeFill) {
2736 tty->print_cr("filling array but bounds don't match");
2737 alloc->dump();
2738 head->init_trip()->dump();
2739 head->limit()->dump();
2740 length->dump();
2741 }
2742 #endif
2743 }
2744 }
2745 */
2746
2747 // Redirect the old control and memory edges that are outside the loop.
2748 Node* exit = head->loopexit()->proj_out(0);
2749 // Sometimes the memory phi of the head is used as the outgoing
2750 // state of the loop. It's safe in this case to replace it with the
2751 // result_mem.
2752 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
2753 _igvn.replace_node(exit, result_ctrl);
2754 _igvn.replace_node(store, result_mem);
2755 // Any uses the increment outside of the loop become the loop limit.
2756 _igvn.replace_node(head->incr(), head->limit());
2757
2758 // Disconnect the head from the loop.
2759 for (uint i = 0; i < lpt->_body.size(); i++) {
2760 Node* n = lpt->_body.at(i);
2761 _igvn.replace_node(n, C->top());
2762 }
2763
2764 return true;
2765 }