1 /*
2 * Copyright (c) 2000, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/convertnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/opaquenode.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/superword.hpp"
42 #include "opto/vectornode.hpp"
43
44 //------------------------------is_loop_exit-----------------------------------
45 // Given an IfNode, return the loop-exiting projection or NULL if both
46 // arms remain in the loop.
47 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
48 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
49 PhaseIdealLoop *phase = _phase;
50 // Test is an IfNode, has 2 projections. If BOTH are in the loop
51 // we need loop unswitching instead of peeling.
52 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
53 return iff->raw_out(0);
54 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
55 return iff->raw_out(1);
56 return NULL;
57 }
58
59
60 //=============================================================================
61
62
63 //------------------------------record_for_igvn----------------------------
64 // Put loop body on igvn work list
65 void IdealLoopTree::record_for_igvn() {
66 for( uint i = 0; i < _body.size(); i++ ) {
67 Node *n = _body.at(i);
68 _phase->_igvn._worklist.push(n);
69 }
70 // put body of outer strip mined loop on igvn work list as well
71 if (_head->is_CountedLoop() && _head->as_Loop()->is_strip_mined()) {
72 CountedLoopNode* l = _head->as_CountedLoop();
73 Node* outer_loop = l->outer_loop();
74 assert(outer_loop != NULL, "missing piece of strip mined loop");
75 _phase->_igvn._worklist.push(outer_loop);
76 Node* outer_loop_tail = l->outer_loop_tail();
77 assert(outer_loop_tail != NULL, "missing piece of strip mined loop");
78 _phase->_igvn._worklist.push(outer_loop_tail);
79 Node* outer_loop_end = l->outer_loop_end();
80 assert(outer_loop_end != NULL, "missing piece of strip mined loop");
81 _phase->_igvn._worklist.push(outer_loop_end);
82 Node* outer_safepoint = l->outer_safepoint();
83 assert(outer_safepoint != NULL, "missing piece of strip mined loop");
84 _phase->_igvn._worklist.push(outer_safepoint);
85 Node* cle_out = _head->as_CountedLoop()->loopexit()->proj_out(false);
86 assert(cle_out != NULL, "missing piece of strip mined loop");
87 _phase->_igvn._worklist.push(cle_out);
88 }
89 }
90
91 //------------------------------compute_exact_trip_count-----------------------
92 // Compute loop trip count if possible. Do not recalculate trip count for
93 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
94 void IdealLoopTree::compute_trip_count(PhaseIdealLoop* phase) {
95 if (!_head->as_Loop()->is_valid_counted_loop()) {
96 return;
97 }
98 CountedLoopNode* cl = _head->as_CountedLoop();
99 // Trip count may become nonexact for iteration split loops since
100 // RCE modifies limits. Note, _trip_count value is not reset since
101 // it is used to limit unrolling of main loop.
102 cl->set_nonexact_trip_count();
103
104 // Loop's test should be part of loop.
105 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
106 return; // Infinite loop
107
108 #ifdef ASSERT
109 BoolTest::mask bt = cl->loopexit()->test_trip();
110 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
111 bt == BoolTest::ne, "canonical test is expected");
112 #endif
113
114 Node* init_n = cl->init_trip();
115 Node* limit_n = cl->limit();
116 if (init_n != NULL && limit_n != NULL) {
117 // Use longs to avoid integer overflow.
118 int stride_con = cl->stride_con();
119 jlong init_con = phase->_igvn.type(init_n)->is_int()->_lo;
120 jlong limit_con = phase->_igvn.type(limit_n)->is_int()->_hi;
121 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
122 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
123 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
124 if (init_n->is_Con() && limit_n->is_Con()) {
125 // Set exact trip count.
126 cl->set_exact_trip_count((uint)trip_count);
127 } else if (cl->unrolled_count() == 1) {
128 // Set maximum trip count before unrolling.
129 cl->set_trip_count((uint)trip_count);
130 }
131 }
132 }
133 }
134
135 //------------------------------compute_profile_trip_cnt----------------------------
136 // Compute loop trip count from profile data as
137 // (backedge_count + loop_exit_count) / loop_exit_count
138 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
139 if (!_head->is_CountedLoop()) {
140 return;
141 }
142 CountedLoopNode* head = _head->as_CountedLoop();
143 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
144 return; // Already computed
145 }
146 float trip_cnt = (float)max_jint; // default is big
147
148 Node* back = head->in(LoopNode::LoopBackControl);
149 while (back != head) {
150 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
151 back->in(0) &&
152 back->in(0)->is_If() &&
153 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
154 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
155 break;
156 }
157 back = phase->idom(back);
158 }
159 if (back != head) {
160 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
161 back->in(0), "if-projection exists");
162 IfNode* back_if = back->in(0)->as_If();
163 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
164
165 // Now compute a loop exit count
166 float loop_exit_cnt = 0.0f;
167 for( uint i = 0; i < _body.size(); i++ ) {
168 Node *n = _body[i];
169 if( n->is_If() ) {
170 IfNode *iff = n->as_If();
171 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
172 Node *exit = is_loop_exit(iff);
173 if( exit ) {
174 float exit_prob = iff->_prob;
175 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
176 if (exit_prob > PROB_MIN) {
177 float exit_cnt = iff->_fcnt * exit_prob;
178 loop_exit_cnt += exit_cnt;
179 }
180 }
181 }
182 }
183 }
184 if (loop_exit_cnt > 0.0f) {
185 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
186 } else {
187 // No exit count so use
188 trip_cnt = loop_back_cnt;
189 }
190 }
191 #ifndef PRODUCT
192 if (TraceProfileTripCount) {
193 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
194 }
195 #endif
196 head->set_profile_trip_cnt(trip_cnt);
197 }
198
199 //---------------------is_invariant_addition-----------------------------
200 // Return nonzero index of invariant operand for an Add or Sub
201 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
202 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
203 int op = n->Opcode();
204 if (op == Op_AddI || op == Op_SubI) {
205 bool in1_invar = this->is_invariant(n->in(1));
206 bool in2_invar = this->is_invariant(n->in(2));
207 if (in1_invar && !in2_invar) return 1;
208 if (!in1_invar && in2_invar) return 2;
209 }
210 return 0;
211 }
212
213 //---------------------reassociate_add_sub-----------------------------
214 // Reassociate invariant add and subtract expressions:
215 //
216 // inv1 + (x + inv2) => ( inv1 + inv2) + x
217 // (x + inv2) + inv1 => ( inv1 + inv2) + x
218 // inv1 + (x - inv2) => ( inv1 - inv2) + x
219 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
220 // (x + inv2) - inv1 => (-inv1 + inv2) + x
221 // (x - inv2) + inv1 => ( inv1 - inv2) + x
222 // (x - inv2) - inv1 => (-inv1 - inv2) + x
223 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
224 // inv1 - (x - inv2) => ( inv1 + inv2) - x
225 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
226 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
227 // inv1 - (x + inv2) => ( inv1 - inv2) - x
228 //
229 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
230 if ((!n1->is_Add() && !n1->is_Sub()) || n1->outcnt() == 0) return NULL;
231 if (is_invariant(n1)) return NULL;
232 int inv1_idx = is_invariant_addition(n1, phase);
233 if (!inv1_idx) return NULL;
234 // Don't mess with add of constant (igvn moves them to expression tree root.)
235 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
236 Node* inv1 = n1->in(inv1_idx);
237 Node* n2 = n1->in(3 - inv1_idx);
238 int inv2_idx = is_invariant_addition(n2, phase);
239 if (!inv2_idx) return NULL;
240 Node* x = n2->in(3 - inv2_idx);
241 Node* inv2 = n2->in(inv2_idx);
242
243 bool neg_x = n2->is_Sub() && inv2_idx == 1;
244 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
245 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
246 if (n1->is_Sub() && inv1_idx == 1) {
247 neg_x = !neg_x;
248 neg_inv2 = !neg_inv2;
249 }
250 Node* inv1_c = phase->get_ctrl(inv1);
251 Node* inv2_c = phase->get_ctrl(inv2);
252 Node* n_inv1;
253 if (neg_inv1) {
254 Node *zero = phase->_igvn.intcon(0);
255 phase->set_ctrl(zero, phase->C->root());
256 n_inv1 = new SubINode(zero, inv1);
257 phase->register_new_node(n_inv1, inv1_c);
258 } else {
259 n_inv1 = inv1;
260 }
261 Node* inv;
262 if (neg_inv2) {
263 inv = new SubINode(n_inv1, inv2);
264 } else {
265 inv = new AddINode(n_inv1, inv2);
266 }
267 phase->register_new_node(inv, phase->get_early_ctrl(inv));
268
269 Node* addx;
270 if (neg_x) {
271 addx = new SubINode(inv, x);
272 } else {
273 addx = new AddINode(x, inv);
274 }
275 phase->register_new_node(addx, phase->get_ctrl(x));
276 phase->_igvn.replace_node(n1, addx);
277 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
278 _body.yank(n1);
279 return addx;
280 }
281
282 //---------------------reassociate_invariants-----------------------------
283 // Reassociate invariant expressions:
284 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
285 for (int i = _body.size() - 1; i >= 0; i--) {
286 Node *n = _body.at(i);
287 for (int j = 0; j < 5; j++) {
288 Node* nn = reassociate_add_sub(n, phase);
289 if (nn == NULL) break;
290 n = nn; // again
291 };
292 }
293 }
294
295 //------------------------------policy_peeling---------------------------------
296 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
297 // make some loop-invariant test (usually a null-check) happen before the loop.
298 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
299 Node *test = ((IdealLoopTree*)this)->tail();
300 int body_size = ((IdealLoopTree*)this)->_body.size();
301 // Peeling does loop cloning which can result in O(N^2) node construction
302 if( body_size > 255 /* Prevent overflow for large body_size */
303 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) {
304 return false; // too large to safely clone
305 }
306
307 // check for vectorized loops, any peeling done was already applied
308 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false;
309
310 while( test != _head ) { // Scan till run off top of loop
311 if( test->is_If() ) { // Test?
312 Node *ctrl = phase->get_ctrl(test->in(1));
313 if (ctrl->is_top())
314 return false; // Found dead test on live IF? No peeling!
315 // Standard IF only has one input value to check for loop invariance
316 assert(test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd || test->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
317 // Condition is not a member of this loop?
318 if( !is_member(phase->get_loop(ctrl)) &&
319 is_loop_exit(test) )
320 return true; // Found reason to peel!
321 }
322 // Walk up dominators to loop _head looking for test which is
323 // executed on every path thru loop.
324 test = phase->idom(test);
325 }
326 return false;
327 }
328
329 //------------------------------peeled_dom_test_elim---------------------------
330 // If we got the effect of peeling, either by actually peeling or by making
331 // a pre-loop which must execute at least once, we can remove all
332 // loop-invariant dominated tests in the main body.
333 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
334 bool progress = true;
335 while( progress ) {
336 progress = false; // Reset for next iteration
337 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
338 Node *test = prev->in(0);
339 while( test != loop->_head ) { // Scan till run off top of loop
340
341 int p_op = prev->Opcode();
342 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
343 test->is_If() && // Test?
344 !test->in(1)->is_Con() && // And not already obvious?
345 // Condition is not a member of this loop?
346 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
347 // Walk loop body looking for instances of this test
348 for( uint i = 0; i < loop->_body.size(); i++ ) {
349 Node *n = loop->_body.at(i);
350 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
351 // IfNode was dominated by version in peeled loop body
352 progress = true;
353 dominated_by( old_new[prev->_idx], n );
354 }
355 }
356 }
357 prev = test;
358 test = idom(test);
359 } // End of scan tests in loop
360
361 } // End of while( progress )
362 }
363
364 //------------------------------do_peeling-------------------------------------
365 // Peel the first iteration of the given loop.
366 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
367 // The pre-loop illegally has 2 control users (old & new loops).
368 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
369 // Do this by making the old-loop fall-in edges act as if they came
370 // around the loopback from the prior iteration (follow the old-loop
371 // backedges) and then map to the new peeled iteration. This leaves
372 // the pre-loop with only 1 user (the new peeled iteration), but the
373 // peeled-loop backedge has 2 users.
374 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
375 // extra backedge user.
376 //
377 // orig
378 //
379 // stmt1
380 // |
381 // v
382 // loop predicate
383 // |
384 // v
385 // loop<----+
386 // | |
387 // stmt2 |
388 // | |
389 // v |
390 // if ^
391 // / \ |
392 // / \ |
393 // v v |
394 // false true |
395 // / \ |
396 // / ----+
397 // |
398 // v
399 // exit
400 //
401 //
402 // after clone loop
403 //
404 // stmt1
405 // |
406 // v
407 // loop predicate
408 // / \
409 // clone / \ orig
410 // / \
411 // / \
412 // v v
413 // +---->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 // region
428 // |
429 // v
430 // exit
431 //
432 //
433 // after peel and predicate move
434 //
435 // stmt1
436 // /
437 // /
438 // clone / orig
439 // /
440 // / +----------+
441 // / | |
442 // / loop predicate |
443 // / | |
444 // v v |
445 // TOP-->loop clone loop<----+ |
446 // | | | |
447 // stmt2 clone stmt2 | |
448 // | | | ^
449 // v v | |
450 // if clone If ^ |
451 // / \ / \ | |
452 // / \ / \ | |
453 // v v v v | |
454 // true false false true | |
455 // | \ / \ | |
456 // | \ / ----+ ^
457 // | \ / |
458 // | 1v v2 |
459 // v region |
460 // | | |
461 // | v |
462 // | exit |
463 // | |
464 // +--------------->-----------------+
465 //
466 //
467 // final graph
468 //
469 // stmt1
470 // |
471 // v
472 // stmt2 clone
473 // |
474 // v
475 // if clone
476 // / |
477 // / |
478 // v v
479 // false true
480 // | |
481 // | v
482 // | loop predicate
483 // | |
484 // | v
485 // | loop<----+
486 // | | |
487 // | stmt2 |
488 // | | |
489 // | v |
490 // v if ^
491 // | / \ |
492 // | / \ |
493 // | v v |
494 // | false true |
495 // | | \ |
496 // v v --+
497 // region
498 // |
499 // v
500 // exit
501 //
502 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
503
504 C->set_major_progress();
505 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
506 // 'pre' loop from the main and the 'pre' can no longer have its
507 // iterations adjusted. Therefore, we need to declare this loop as
508 // no longer a 'main' loop; it will need new pre and post loops before
509 // we can do further RCE.
510 #ifndef PRODUCT
511 if (TraceLoopOpts) {
512 tty->print("Peel ");
513 loop->dump_head();
514 }
515 #endif
516 LoopNode* head = loop->_head->as_Loop();
517 bool counted_loop = head->is_CountedLoop();
518 if (counted_loop) {
519 CountedLoopNode *cl = head->as_CountedLoop();
520 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
521 cl->set_trip_count(cl->trip_count() - 1);
522 if (cl->is_main_loop()) {
523 cl->set_normal_loop();
524 #ifndef PRODUCT
525 if (PrintOpto && VerifyLoopOptimizations) {
526 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
527 loop->dump_head();
528 }
529 #endif
530 }
531 }
532 Node* entry = head->in(LoopNode::EntryControl);
533
534 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
535 // The pre-loop illegally has 2 control users (old & new loops).
536 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), ControlAroundStripMined);
537
538 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
539 // Do this by making the old-loop fall-in edges act as if they came
540 // around the loopback from the prior iteration (follow the old-loop
541 // backedges) and then map to the new peeled iteration. This leaves
542 // the pre-loop with only 1 user (the new peeled iteration), but the
543 // peeled-loop backedge has 2 users.
544 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
545 _igvn.hash_delete(head->skip_strip_mined());
546 head->skip_strip_mined()->set_req(LoopNode::EntryControl, new_entry);
547 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
548 Node* old = head->fast_out(j);
549 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
550 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
551 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
552 // Then loop body backedge value remains the same.
553 new_exit_value = old->in(LoopNode::LoopBackControl);
554 _igvn.hash_delete(old);
555 old->set_req(LoopNode::EntryControl, new_exit_value);
556 }
557 }
558
559
560 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
561 // extra backedge user.
562 Node* new_head = old_new[head->_idx];
563 _igvn.hash_delete(new_head);
564 new_head->set_req(LoopNode::LoopBackControl, C->top());
565 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
566 Node* use = new_head->fast_out(j2);
567 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
568 _igvn.hash_delete(use);
569 use->set_req(LoopNode::LoopBackControl, C->top());
570 }
571 }
572
573
574 // Step 4: Correct dom-depth info. Set to loop-head depth.
575 int dd = dom_depth(head);
576 set_idom(head, head->in(1), dd);
577 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
578 Node *old = loop->_body.at(j3);
579 Node *nnn = old_new[old->_idx];
580 if (!has_ctrl(nnn))
581 set_idom(nnn, idom(nnn), dd-1);
582 }
583
584 // Now force out all loop-invariant dominating tests. The optimizer
585 // finds some, but we _know_ they are all useless.
586 peeled_dom_test_elim(loop,old_new);
587
588 loop->record_for_igvn();
589 }
590
591 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
592
593 //------------------------------policy_maximally_unroll------------------------
594 // Calculate exact loop trip count and return true if loop can be maximally
595 // unrolled.
596 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
597 CountedLoopNode *cl = _head->as_CountedLoop();
598 assert(cl->is_normal_loop(), "");
599 if (!cl->is_valid_counted_loop())
600 return false; // Malformed counted loop
601
602 if (!cl->has_exact_trip_count()) {
603 // Trip count is not exact.
604 return false;
605 }
606
607 uint trip_count = cl->trip_count();
608 // Note, max_juint is used to indicate unknown trip count.
609 assert(trip_count > 1, "one iteration loop should be optimized out already");
610 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
611
612 // Real policy: if we maximally unroll, does it get too big?
613 // Allow the unrolled mess to get larger than standard loop
614 // size. After all, it will no longer be a loop.
615 uint body_size = _body.size();
616 uint unroll_limit = (uint)LoopUnrollLimit * 4;
617 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
618 if (trip_count > unroll_limit || body_size > unroll_limit) {
619 return false;
620 }
621
622 // Fully unroll a loop with few iterations regardless next
623 // conditions since following loop optimizations will split
624 // such loop anyway (pre-main-post).
625 if (trip_count <= 3)
626 return true;
627
628 // Take into account that after unroll conjoined heads and tails will fold,
629 // otherwise policy_unroll() may allow more unrolling than max unrolling.
630 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
631 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
632 if (body_size != tst_body_size) // Check for int overflow
633 return false;
634 if (new_body_size > unroll_limit ||
635 // Unrolling can result in a large amount of node construction
636 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) {
637 return false;
638 }
639
640 // Do not unroll a loop with String intrinsics code.
641 // String intrinsics are large and have loops.
642 for (uint k = 0; k < _body.size(); k++) {
643 Node* n = _body.at(k);
644 switch (n->Opcode()) {
645 case Op_StrComp:
646 case Op_StrEquals:
647 case Op_StrIndexOf:
648 case Op_StrIndexOfChar:
649 case Op_EncodeISOArray:
650 case Op_AryEq:
651 case Op_HasNegatives: {
652 return false;
653 }
654 #if INCLUDE_RTM_OPT
655 case Op_FastLock:
656 case Op_FastUnlock: {
657 // Don't unroll RTM locking code because it is large.
658 if (UseRTMLocking) {
659 return false;
660 }
661 }
662 #endif
663 } // switch
664 }
665
666 return true; // Do maximally unroll
667 }
668
669
670 //------------------------------policy_unroll----------------------------------
671 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
672 // the loop is a CountedLoop and the body is small enough.
673 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
674
675 CountedLoopNode *cl = _head->as_CountedLoop();
676 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
677
678 if (!cl->is_valid_counted_loop())
679 return false; // Malformed counted loop
680
681 // Protect against over-unrolling.
682 // After split at least one iteration will be executed in pre-loop.
683 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
684
685 _local_loop_unroll_limit = LoopUnrollLimit;
686 _local_loop_unroll_factor = 4;
687 int future_unroll_ct = cl->unrolled_count() * 2;
688 if (!cl->is_vectorized_loop()) {
689 if (future_unroll_ct > LoopMaxUnroll) return false;
690 } else {
691 // obey user constraints on vector mapped loops with additional unrolling applied
692 int unroll_constraint = (cl->slp_max_unroll()) ? cl->slp_max_unroll() : 1;
693 if ((future_unroll_ct / unroll_constraint) > LoopMaxUnroll) return false;
694 }
695
696 // Check for initial stride being a small enough constant
697 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
698
699 // Don't unroll if the next round of unrolling would push us
700 // over the expected trip count of the loop. One is subtracted
701 // from the expected trip count because the pre-loop normally
702 // executes 1 iteration.
703 if (UnrollLimitForProfileCheck > 0 &&
704 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
705 future_unroll_ct > UnrollLimitForProfileCheck &&
706 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
707 return false;
708 }
709
710 // When unroll count is greater than LoopUnrollMin, don't unroll if:
711 // the residual iterations are more than 10% of the trip count
712 // and rounds of "unroll,optimize" are not making significant progress
713 // Progress defined as current size less than 20% larger than previous size.
714 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
715 future_unroll_ct > LoopUnrollMin &&
716 (future_unroll_ct - 1) * (100 / LoopPercentProfileLimit) > cl->profile_trip_cnt() &&
717 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
718 return false;
719 }
720
721 Node *init_n = cl->init_trip();
722 Node *limit_n = cl->limit();
723 int stride_con = cl->stride_con();
724 // Non-constant bounds.
725 // Protect against over-unrolling when init or/and limit are not constant
726 // (so that trip_count's init value is maxint) but iv range is known.
727 if (init_n == NULL || !init_n->is_Con() ||
728 limit_n == NULL || !limit_n->is_Con()) {
729 Node* phi = cl->phi();
730 if (phi != NULL) {
731 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
732 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
733 int next_stride = stride_con * 2; // stride after this unroll
734 if (next_stride > 0) {
735 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
736 iv_type->_lo + next_stride > iv_type->_hi) {
737 return false; // over-unrolling
738 }
739 } else if (next_stride < 0) {
740 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
741 iv_type->_hi + next_stride < iv_type->_lo) {
742 return false; // over-unrolling
743 }
744 }
745 }
746 }
747
748 // After unroll limit will be adjusted: new_limit = limit-stride.
749 // Bailout if adjustment overflow.
750 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
751 if ((stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi)) ||
752 (stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)))
753 return false; // overflow
754
755 // Adjust body_size to determine if we unroll or not
756 uint body_size = _body.size();
757 // Key test to unroll loop in CRC32 java code
758 int xors_in_loop = 0;
759 // Also count ModL, DivL and MulL which expand mightly
760 for (uint k = 0; k < _body.size(); k++) {
761 Node* n = _body.at(k);
762 switch (n->Opcode()) {
763 case Op_XorI: xors_in_loop++; break; // CRC32 java code
764 case Op_ModL: body_size += 30; break;
765 case Op_DivL: body_size += 30; break;
766 case Op_MulL: body_size += 10; break;
767 case Op_StrComp:
768 case Op_StrEquals:
769 case Op_StrIndexOf:
770 case Op_StrIndexOfChar:
771 case Op_EncodeISOArray:
772 case Op_AryEq:
773 case Op_HasNegatives: {
774 // Do not unroll a loop with String intrinsics code.
775 // String intrinsics are large and have loops.
776 return false;
777 }
778 #if INCLUDE_RTM_OPT
779 case Op_FastLock:
780 case Op_FastUnlock: {
781 // Don't unroll RTM locking code because it is large.
782 if (UseRTMLocking) {
783 return false;
784 }
785 }
786 #endif
787 } // switch
788 }
789
790 if (UseSuperWord) {
791 if (!cl->is_reduction_loop()) {
792 phase->mark_reductions(this);
793 }
794
795 // Only attempt slp analysis when user controls do not prohibit it
796 if (LoopMaxUnroll > _local_loop_unroll_factor) {
797 // Once policy_slp_analysis succeeds, mark the loop with the
798 // maximal unroll factor so that we minimize analysis passes
799 if (future_unroll_ct >= _local_loop_unroll_factor) {
800 policy_unroll_slp_analysis(cl, phase, future_unroll_ct);
801 }
802 }
803 }
804
805 int slp_max_unroll_factor = cl->slp_max_unroll();
806 if ((LoopMaxUnroll < slp_max_unroll_factor) && FLAG_IS_DEFAULT(LoopMaxUnroll) && UseSubwordForMaxVector) {
807 LoopMaxUnroll = slp_max_unroll_factor;
808 }
809 if (cl->has_passed_slp()) {
810 if (slp_max_unroll_factor >= future_unroll_ct) return true;
811 // Normal case: loop too big
812 return false;
813 }
814
815 // Check for being too big
816 if (body_size > (uint)_local_loop_unroll_limit) {
817 if ((UseSubwordForMaxVector || xors_in_loop >= 4) && body_size < (uint)LoopUnrollLimit * 4) return true;
818 // Normal case: loop too big
819 return false;
820 }
821
822 if (cl->do_unroll_only()) {
823 if (TraceSuperWordLoopUnrollAnalysis) {
824 tty->print_cr("policy_unroll passed vector loop(vlen=%d,factor = %d)\n", slp_max_unroll_factor, future_unroll_ct);
825 }
826 }
827
828 // Unroll once! (Each trip will soon do double iterations)
829 return true;
830 }
831
832 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) {
833 // Enable this functionality target by target as needed
834 if (SuperWordLoopUnrollAnalysis) {
835 if (!cl->was_slp_analyzed()) {
836 SuperWord sw(phase);
837 sw.transform_loop(this, false);
838
839 // If the loop is slp canonical analyze it
840 if (sw.early_return() == false) {
841 sw.unrolling_analysis(_local_loop_unroll_factor);
842 }
843 }
844
845 if (cl->has_passed_slp()) {
846 int slp_max_unroll_factor = cl->slp_max_unroll();
847 if (slp_max_unroll_factor >= future_unroll_ct) {
848 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor;
849 if (new_limit > LoopUnrollLimit) {
850 if (TraceSuperWordLoopUnrollAnalysis) {
851 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit);
852 }
853 _local_loop_unroll_limit = new_limit;
854 }
855 }
856 }
857 }
858 }
859
860 //------------------------------policy_align-----------------------------------
861 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
862 // expression that does the alignment. Note that only one array base can be
863 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
864 // if we vectorize short memory ops into longer memory ops, we may want to
865 // increase alignment.
866 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
867 return false;
868 }
869
870 //------------------------------policy_range_check-----------------------------
871 // Return TRUE or FALSE if the loop should be range-check-eliminated.
872 // Actually we do iteration-splitting, a more powerful form of RCE.
873 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
874 if (!RangeCheckElimination) return false;
875
876 CountedLoopNode *cl = _head->as_CountedLoop();
877 // If we unrolled with no intention of doing RCE and we later
878 // changed our minds, we got no pre-loop. Either we need to
879 // make a new pre-loop, or we gotta disallow RCE.
880 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
881 Node *trip_counter = cl->phi();
882
883 // check for vectorized loops, some opts are no longer needed
884 if (cl->do_unroll_only()) return false;
885
886 // Check loop body for tests of trip-counter plus loop-invariant vs
887 // loop-invariant.
888 for (uint i = 0; i < _body.size(); i++) {
889 Node *iff = _body[i];
890 if (iff->Opcode() == Op_If ||
891 iff->Opcode() == Op_RangeCheck) { // Test?
892
893 // Comparing trip+off vs limit
894 Node *bol = iff->in(1);
895 if (bol->req() != 2) continue; // dead constant test
896 if (!bol->is_Bool()) {
897 assert(bol->Opcode() == Op_Conv2B, "predicate check only");
898 continue;
899 }
900 if (bol->as_Bool()->_test._test == BoolTest::ne)
901 continue; // not RC
902
903 Node *cmp = bol->in(1);
904 Node *rc_exp = cmp->in(1);
905 Node *limit = cmp->in(2);
906
907 Node *limit_c = phase->get_ctrl(limit);
908 if( limit_c == phase->C->top() )
909 return false; // Found dead test on live IF? No RCE!
910 if( is_member(phase->get_loop(limit_c) ) ) {
911 // Compare might have operands swapped; commute them
912 rc_exp = cmp->in(2);
913 limit = cmp->in(1);
914 limit_c = phase->get_ctrl(limit);
915 if( is_member(phase->get_loop(limit_c) ) )
916 continue; // Both inputs are loop varying; cannot RCE
917 }
918
919 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
920 continue;
921 }
922 // Yeah! Found a test like 'trip+off vs limit'
923 // Test is an IfNode, has 2 projections. If BOTH are in the loop
924 // we need loop unswitching instead of iteration splitting.
925 if( is_loop_exit(iff) )
926 return true; // Found reason to split iterations
927 } // End of is IF
928 }
929
930 return false;
931 }
932
933 //------------------------------policy_peel_only-------------------------------
934 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
935 // for unrolling loops with NO array accesses.
936 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
937 // check for vectorized loops, any peeling done was already applied
938 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false;
939
940 for( uint i = 0; i < _body.size(); i++ )
941 if( _body[i]->is_Mem() )
942 return false;
943
944 // No memory accesses at all!
945 return true;
946 }
947
948 //------------------------------clone_up_backedge_goo--------------------------
949 // If Node n lives in the back_ctrl block and cannot float, we clone a private
950 // version of n in preheader_ctrl block and return that, otherwise return n.
951 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
952 if( get_ctrl(n) != back_ctrl ) return n;
953
954 // Only visit once
955 if (visited.test_set(n->_idx)) {
956 Node *x = clones.find(n->_idx);
957 if (x != NULL)
958 return x;
959 return n;
960 }
961
962 Node *x = NULL; // If required, a clone of 'n'
963 // Check for 'n' being pinned in the backedge.
964 if( n->in(0) && n->in(0) == back_ctrl ) {
965 assert(clones.find(n->_idx) == NULL, "dead loop");
966 x = n->clone(); // Clone a copy of 'n' to preheader
967 clones.push(x, n->_idx);
968 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
969 }
970
971 // Recursive fixup any other input edges into x.
972 // If there are no changes we can just return 'n', otherwise
973 // we need to clone a private copy and change it.
974 for( uint i = 1; i < n->req(); i++ ) {
975 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
976 if( g != n->in(i) ) {
977 if( !x ) {
978 assert(clones.find(n->_idx) == NULL, "dead loop");
979 x = n->clone();
980 clones.push(x, n->_idx);
981 }
982 x->set_req(i, g);
983 }
984 }
985 if( x ) { // x can legally float to pre-header location
986 register_new_node( x, preheader_ctrl );
987 return x;
988 } else { // raise n to cover LCA of uses
989 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
990 }
991 return n;
992 }
993
994 Node* PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) {
995 Node* castii = new CastIINode(incr, TypeInt::INT, true);
996 castii->set_req(0, ctrl);
997 register_new_node(castii, ctrl);
998 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) {
999 Node* n = incr->fast_out(i);
1000 if (n->is_Phi() && n->in(0) == loop) {
1001 int nrep = n->replace_edge(incr, castii);
1002 return castii;
1003 }
1004 }
1005 return NULL;
1006 }
1007
1008 // Make a copy of the skeleton range check predicates before the main
1009 // loop and set the initial value of loop as input. After unrolling,
1010 // the range of values for the induction variable in the main loop can
1011 // fall outside the allowed range of values by the array access (main
1012 // loop is never executed). When that happens, range check
1013 // CastII/ConvI2L nodes cause some data paths to die. For consistency,
1014 // the control paths must die too but the range checks were removed by
1015 // predication. The range checks that we add here guarantee that they
1016 // do.
1017 void PhaseIdealLoop::duplicate_predicates(CountedLoopNode* pre_head, Node* min_taken, Node* castii,
1018 IdealLoopTree* outer_loop, LoopNode* outer_main_head,
1019 uint dd_main_head) {
1020 if (UseLoopPredicate) {
1021 Node* entry = pre_head->in(LoopNode::EntryControl);
1022 Node* predicate = NULL;
1023 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1024 if (predicate != NULL) {
1025 entry = entry->in(0)->in(0);
1026 }
1027 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1028 if (predicate != NULL) {
1029 IfNode* iff = entry->in(0)->as_If();
1030 ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
1031 Node* rgn = uncommon_proj->unique_ctrl_out();
1032 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
1033 assert(iff->in(1)->in(1)->Opcode() == Op_Opaque1, "unexpected predicate shape");
1034 entry = entry->in(0)->in(0);
1035 Node* prev_proj = min_taken;
1036 while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
1037 uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
1038 if (uncommon_proj->unique_ctrl_out() != rgn)
1039 break;
1040 iff = entry->in(0)->as_If();
1041 if (iff->in(1)->Opcode() == Op_Opaque4) {
1042 Node_Stack to_clone(2);
1043 to_clone.push(iff->in(1), 1);
1044 uint current = C->unique();
1045 Node* result = NULL;
1046 // Look for the opaque node to replace with the init value
1047 // and clone everything in between. We keep the Opaque4 node
1048 // so the duplicated predicates are eliminated once loop
1049 // opts are over: they are here only to keep the IR graph
1050 // consistent.
1051 do {
1052 Node* n = to_clone.node();
1053 uint i = to_clone.index();
1054 Node* m = n->in(i);
1055 int op = m->Opcode();
1056 if (m->is_Bool() ||
1057 m->is_Cmp() ||
1058 op == Op_AndL ||
1059 op == Op_OrL ||
1060 op == Op_RShiftL ||
1061 op == Op_LShiftL ||
1062 op == Op_AddL ||
1063 op == Op_AddI ||
1064 op == Op_MulL ||
1065 op == Op_MulI ||
1066 op == Op_SubL ||
1067 op == Op_SubI ||
1068 op == Op_ConvI2L) {
1069 to_clone.push(m, 1);
1070 continue;
1071 }
1072 if (op == Op_Opaque1) {
1073 if (n->_idx < current) {
1074 n = n->clone();
1075 }
1076 n->set_req(i, castii);
1077 register_new_node(n, min_taken);
1078 to_clone.set_node(n);
1079 }
1080 for (;;) {
1081 Node* cur = to_clone.node();
1082 uint j = to_clone.index();
1083 if (j+1 < cur->req()) {
1084 to_clone.set_index(j+1);
1085 break;
1086 }
1087 to_clone.pop();
1088 if (to_clone.size() == 0) {
1089 result = cur;
1090 break;
1091 }
1092 Node* next = to_clone.node();
1093 j = to_clone.index();
1094 if (cur->_idx >= current) {
1095 if (next->_idx < current) {
1096 next = next->clone();
1097 register_new_node(next, min_taken);
1098 to_clone.set_node(next);
1099 }
1100 assert(next->in(j) != cur, "input should have been cloned");
1101 next->set_req(j, cur);
1102 }
1103 }
1104 } while (result == NULL);
1105 assert(result->_idx >= current, "new node expected");
1106
1107 Node* proj = entry->clone();
1108 Node* other_proj = uncommon_proj->clone();
1109 Node* new_iff = iff->clone();
1110 new_iff->set_req(1, result);
1111 proj->set_req(0, new_iff);
1112 other_proj->set_req(0, new_iff);
1113 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
1114 register_new_node(frame, C->start());
1115 // It's impossible for the predicate to fail at runtime. Use
1116 // an Halt node.
1117 Node* halt = new HaltNode(other_proj, frame);
1118 C->root()->add_req(halt);
1119 new_iff->set_req(0, prev_proj);
1120
1121 register_control(new_iff, outer_loop->_parent, prev_proj);
1122 register_control(proj, outer_loop->_parent, new_iff);
1123 register_control(other_proj, _ltree_root, new_iff);
1124 register_control(halt, _ltree_root, other_proj);
1125
1126 prev_proj = proj;
1127 }
1128 entry = entry->in(0)->in(0);
1129 }
1130 _igvn.replace_input_of(outer_main_head, LoopNode::EntryControl, prev_proj);
1131 set_idom(outer_main_head, prev_proj, dd_main_head);
1132 }
1133 }
1134 }
1135
1136 //------------------------------insert_pre_post_loops--------------------------
1137 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
1138 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
1139 // alignment. Useful to unroll loops that do no array accesses.
1140 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
1141
1142 #ifndef PRODUCT
1143 if (TraceLoopOpts) {
1144 if (peel_only)
1145 tty->print("PeelMainPost ");
1146 else
1147 tty->print("PreMainPost ");
1148 loop->dump_head();
1149 }
1150 #endif
1151 C->set_major_progress();
1152
1153 // Find common pieces of the loop being guarded with pre & post loops
1154 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1155 assert( main_head->is_normal_loop(), "" );
1156 CountedLoopEndNode *main_end = main_head->loopexit();
1157 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
1158
1159 Node *pre_header= main_head->in(LoopNode::EntryControl);
1160 Node *init = main_head->init_trip();
1161 Node *incr = main_end ->incr();
1162 Node *limit = main_end ->limit();
1163 Node *stride = main_end ->stride();
1164 Node *cmp = main_end ->cmp_node();
1165 BoolTest::mask b_test = main_end->test_trip();
1166
1167 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
1168 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
1169 if( bol->outcnt() != 1 ) {
1170 bol = bol->clone();
1171 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
1172 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol);
1173 }
1174 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
1175 if( cmp->outcnt() != 1 ) {
1176 cmp = cmp->clone();
1177 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
1178 _igvn.replace_input_of(bol, 1, cmp);
1179 }
1180
1181 // Add the post loop
1182 CountedLoopNode *post_head = NULL;
1183 Node *main_exit = insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1184
1185 //------------------------------
1186 // Step B: Create Pre-Loop.
1187
1188 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1189 // loop pre-header illegally has 2 control users (old & new loops).
1190 LoopNode* outer_main_head = main_head;
1191 IdealLoopTree* outer_loop = loop;
1192 if (main_head->is_strip_mined()) {
1193 main_head->verify_strip_mined(1);
1194 outer_main_head = main_head->outer_loop();
1195 outer_loop = loop->_parent;
1196 assert(outer_loop->_head == outer_main_head, "broken loop tree");
1197 }
1198 uint dd_main_head = dom_depth(outer_main_head);
1199 clone_loop(loop, old_new, dd_main_head, ControlAroundStripMined);
1200 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1201 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1202 pre_head->set_pre_loop(main_head);
1203 Node *pre_incr = old_new[incr->_idx];
1204
1205 // Reduce the pre-loop trip count.
1206 pre_end->_prob = PROB_FAIR;
1207
1208 // Find the pre-loop normal exit.
1209 Node* pre_exit = pre_end->proj_out(false);
1210 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1211 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end);
1212 _igvn.register_new_node_with_optimizer( new_pre_exit );
1213 set_idom(new_pre_exit, pre_end, dd_main_head);
1214 set_loop(new_pre_exit, outer_loop->_parent);
1215
1216 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1217 // pre-loop, the main-loop may not execute at all. Later in life this
1218 // zero-trip guard will become the minimum-trip guard when we unroll
1219 // the main-loop.
1220 Node *min_opaq = new Opaque1Node(C, limit);
1221 Node *min_cmp = new CmpINode( pre_incr, min_opaq );
1222 Node *min_bol = new BoolNode( min_cmp, b_test );
1223 register_new_node( min_opaq, new_pre_exit );
1224 register_new_node( min_cmp , new_pre_exit );
1225 register_new_node( min_bol , new_pre_exit );
1226
1227 // Build the IfNode (assume the main-loop is executed always).
1228 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1229 _igvn.register_new_node_with_optimizer( min_iff );
1230 set_idom(min_iff, new_pre_exit, dd_main_head);
1231 set_loop(min_iff, outer_loop->_parent);
1232
1233 // Plug in the false-path, taken if we need to skip main-loop
1234 _igvn.hash_delete( pre_exit );
1235 pre_exit->set_req(0, min_iff);
1236 set_idom(pre_exit, min_iff, dd_main_head);
1237 set_idom(pre_exit->unique_ctrl_out(), min_iff, dd_main_head);
1238 // Make the true-path, must enter the main loop
1239 Node *min_taken = new IfTrueNode( min_iff );
1240 _igvn.register_new_node_with_optimizer( min_taken );
1241 set_idom(min_taken, min_iff, dd_main_head);
1242 set_loop(min_taken, outer_loop->_parent);
1243 // Plug in the true path
1244 _igvn.hash_delete(outer_main_head);
1245 outer_main_head->set_req(LoopNode::EntryControl, min_taken);
1246 set_idom(outer_main_head, min_taken, dd_main_head);
1247
1248 Arena *a = Thread::current()->resource_area();
1249 VectorSet visited(a);
1250 Node_Stack clones(a, main_head->back_control()->outcnt());
1251 // Step B3: Make the fall-in values to the main-loop come from the
1252 // fall-out values of the pre-loop.
1253 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1254 Node* main_phi = main_head->fast_out(i2);
1255 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1256 Node *pre_phi = old_new[main_phi->_idx];
1257 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1258 main_head->skip_strip_mined()->in(LoopNode::EntryControl),
1259 pre_phi->in(LoopNode::LoopBackControl),
1260 visited, clones);
1261 _igvn.hash_delete(main_phi);
1262 main_phi->set_req( LoopNode::EntryControl, fallpre );
1263 }
1264 }
1265
1266 // Nodes inside the loop may be control dependent on a predicate
1267 // that was moved before the preloop. If the back branch of the main
1268 // or post loops becomes dead, those nodes won't be dependent on the
1269 // test that guards that loop nest anymore which could lead to an
1270 // incorrect array access because it executes independently of the
1271 // test that was guarding the loop nest. We add a special CastII on
1272 // the if branch that enters the loop, between the input induction
1273 // variable value and the induction variable Phi to preserve correct
1274 // dependencies.
1275
1276 // CastII for the main loop:
1277 Node* castii = cast_incr_before_loop( pre_incr, min_taken, main_head );
1278 assert(castii != NULL, "no castII inserted");
1279 duplicate_predicates(pre_head, min_taken, castii, outer_loop, outer_main_head, dd_main_head);
1280
1281 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1282 // RCE and alignment may change this later.
1283 Node *cmp_end = pre_end->cmp_node();
1284 assert( cmp_end->in(2) == limit, "" );
1285 Node *pre_limit = new AddINode( init, stride );
1286
1287 // Save the original loop limit in this Opaque1 node for
1288 // use by range check elimination.
1289 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit);
1290
1291 register_new_node( pre_limit, pre_head->in(0) );
1292 register_new_node( pre_opaq , pre_head->in(0) );
1293
1294 // Since no other users of pre-loop compare, I can hack limit directly
1295 assert( cmp_end->outcnt() == 1, "no other users" );
1296 _igvn.hash_delete(cmp_end);
1297 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1298
1299 // Special case for not-equal loop bounds:
1300 // Change pre loop test, main loop test, and the
1301 // main loop guard test to use lt or gt depending on stride
1302 // direction:
1303 // positive stride use <
1304 // negative stride use >
1305 //
1306 // not-equal test is kept for post loop to handle case
1307 // when init > limit when stride > 0 (and reverse).
1308
1309 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1310
1311 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1312 // Modify pre loop end condition
1313 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1314 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test);
1315 register_new_node( new_bol0, pre_head->in(0) );
1316 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0);
1317 // Modify main loop guard condition
1318 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1319 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test);
1320 register_new_node( new_bol1, new_pre_exit );
1321 _igvn.hash_delete(min_iff);
1322 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1323 // Modify main loop end condition
1324 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1325 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test);
1326 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1327 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2);
1328 }
1329
1330 // Flag main loop
1331 main_head->set_main_loop();
1332 if( peel_only ) main_head->set_main_no_pre_loop();
1333
1334 // Subtract a trip count for the pre-loop.
1335 main_head->set_trip_count(main_head->trip_count() - 1);
1336
1337 // It's difficult to be precise about the trip-counts
1338 // for the pre/post loops. They are usually very short,
1339 // so guess that 4 trips is a reasonable value.
1340 post_head->set_profile_trip_cnt(4.0);
1341 pre_head->set_profile_trip_cnt(4.0);
1342
1343 // Now force out all loop-invariant dominating tests. The optimizer
1344 // finds some, but we _know_ they are all useless.
1345 peeled_dom_test_elim(loop,old_new);
1346 loop->record_for_igvn();
1347 }
1348
1349 //------------------------------insert_vector_post_loop------------------------
1350 // Insert a copy of the atomic unrolled vectorized main loop as a post loop,
1351 // unroll_policy has already informed us that more unrolling is about to happen to
1352 // the main loop. The resultant post loop will serve as a vectorized drain loop.
1353 void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1354 if (!loop->_head->is_CountedLoop()) return;
1355
1356 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1357
1358 // only process vectorized main loops
1359 if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return;
1360
1361 int slp_max_unroll_factor = cl->slp_max_unroll();
1362 int cur_unroll = cl->unrolled_count();
1363
1364 if (slp_max_unroll_factor == 0) return;
1365
1366 // only process atomic unroll vector loops (not super unrolled after vectorization)
1367 if (cur_unroll != slp_max_unroll_factor) return;
1368
1369 // we only ever process this one time
1370 if (cl->has_atomic_post_loop()) return;
1371
1372 #ifndef PRODUCT
1373 if (TraceLoopOpts) {
1374 tty->print("PostVector ");
1375 loop->dump_head();
1376 }
1377 #endif
1378 C->set_major_progress();
1379
1380 // Find common pieces of the loop being guarded with pre & post loops
1381 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1382 CountedLoopEndNode *main_end = main_head->loopexit();
1383 // diagnostic to show loop end is not properly formed
1384 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1385
1386 // mark this loop as processed
1387 main_head->mark_has_atomic_post_loop();
1388
1389 Node *incr = main_end->incr();
1390 Node *limit = main_end->limit();
1391
1392 // In this case we throw away the result as we are not using it to connect anything else.
1393 CountedLoopNode *post_head = NULL;
1394 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1395
1396 // It's difficult to be precise about the trip-counts
1397 // for post loops. They are usually very short,
1398 // so guess that unit vector trips is a reasonable value.
1399 post_head->set_profile_trip_cnt(cur_unroll);
1400
1401 // Now force out all loop-invariant dominating tests. The optimizer
1402 // finds some, but we _know_ they are all useless.
1403 peeled_dom_test_elim(loop, old_new);
1404 loop->record_for_igvn();
1405 }
1406
1407
1408 //-------------------------insert_scalar_rced_post_loop------------------------
1409 // Insert a copy of the rce'd main loop as a post loop,
1410 // We have not unrolled the main loop, so this is the right time to inject this.
1411 // Later we will examine the partner of this post loop pair which still has range checks
1412 // to see inject code which tests at runtime if the range checks are applicable.
1413 void PhaseIdealLoop::insert_scalar_rced_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1414 if (!loop->_head->is_CountedLoop()) return;
1415
1416 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1417
1418 // only process RCE'd main loops
1419 if (!cl->is_main_loop() || cl->range_checks_present()) return;
1420
1421 #ifndef PRODUCT
1422 if (TraceLoopOpts) {
1423 tty->print("PostScalarRce ");
1424 loop->dump_head();
1425 }
1426 #endif
1427 C->set_major_progress();
1428
1429 // Find common pieces of the loop being guarded with pre & post loops
1430 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1431 CountedLoopEndNode *main_end = main_head->loopexit();
1432 // diagnostic to show loop end is not properly formed
1433 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1434
1435 Node *incr = main_end->incr();
1436 Node *limit = main_end->limit();
1437
1438 // In this case we throw away the result as we are not using it to connect anything else.
1439 CountedLoopNode *post_head = NULL;
1440 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1441
1442 // It's difficult to be precise about the trip-counts
1443 // for post loops. They are usually very short,
1444 // so guess that unit vector trips is a reasonable value.
1445 post_head->set_profile_trip_cnt(4.0);
1446 post_head->set_is_rce_post_loop();
1447
1448 // Now force out all loop-invariant dominating tests. The optimizer
1449 // finds some, but we _know_ they are all useless.
1450 peeled_dom_test_elim(loop, old_new);
1451 loop->record_for_igvn();
1452 }
1453
1454
1455 //------------------------------insert_post_loop-------------------------------
1456 // Insert post loops. Add a post loop to the given loop passed.
1457 Node *PhaseIdealLoop::insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
1458 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
1459 Node *incr, Node *limit, CountedLoopNode *&post_head) {
1460 IfNode* outer_main_end = main_end;
1461 IdealLoopTree* outer_loop = loop;
1462 if (main_head->is_strip_mined()) {
1463 main_head->verify_strip_mined(1);
1464 outer_main_end = main_head->outer_loop_end();
1465 outer_loop = loop->_parent;
1466 assert(outer_loop->_head == main_head->in(LoopNode::EntryControl), "broken loop tree");
1467 }
1468
1469 //------------------------------
1470 // Step A: Create a new post-Loop.
1471 Node* main_exit = outer_main_end->proj_out(false);
1472 assert(main_exit->Opcode() == Op_IfFalse, "");
1473 int dd_main_exit = dom_depth(main_exit);
1474
1475 // Step A1: Clone the loop body of main. The clone becomes the post-loop.
1476 // The main loop pre-header illegally has 2 control users (old & new loops).
1477 clone_loop(loop, old_new, dd_main_exit, ControlAroundStripMined);
1478 assert(old_new[main_end->_idx]->Opcode() == Op_CountedLoopEnd, "");
1479 post_head = old_new[main_head->_idx]->as_CountedLoop();
1480 post_head->set_normal_loop();
1481 post_head->set_post_loop(main_head);
1482
1483 // Reduce the post-loop trip count.
1484 CountedLoopEndNode* post_end = old_new[main_end->_idx]->as_CountedLoopEnd();
1485 post_end->_prob = PROB_FAIR;
1486
1487 // Build the main-loop normal exit.
1488 IfFalseNode *new_main_exit = new IfFalseNode(outer_main_end);
1489 _igvn.register_new_node_with_optimizer(new_main_exit);
1490 set_idom(new_main_exit, outer_main_end, dd_main_exit);
1491 set_loop(new_main_exit, outer_loop->_parent);
1492
1493 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
1494 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
1495 // (the previous loop trip-counter exit value) because we will be changing
1496 // the exit value (via additional unrolling) so we cannot constant-fold away the zero
1497 // trip guard until all unrolling is done.
1498 Node *zer_opaq = new Opaque1Node(C, incr);
1499 Node *zer_cmp = new CmpINode(zer_opaq, limit);
1500 Node *zer_bol = new BoolNode(zer_cmp, main_end->test_trip());
1501 register_new_node(zer_opaq, new_main_exit);
1502 register_new_node(zer_cmp, new_main_exit);
1503 register_new_node(zer_bol, new_main_exit);
1504
1505 // Build the IfNode
1506 IfNode *zer_iff = new IfNode(new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN);
1507 _igvn.register_new_node_with_optimizer(zer_iff);
1508 set_idom(zer_iff, new_main_exit, dd_main_exit);
1509 set_loop(zer_iff, outer_loop->_parent);
1510
1511 // Plug in the false-path, taken if we need to skip this post-loop
1512 _igvn.replace_input_of(main_exit, 0, zer_iff);
1513 set_idom(main_exit, zer_iff, dd_main_exit);
1514 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
1515 // Make the true-path, must enter this post loop
1516 Node *zer_taken = new IfTrueNode(zer_iff);
1517 _igvn.register_new_node_with_optimizer(zer_taken);
1518 set_idom(zer_taken, zer_iff, dd_main_exit);
1519 set_loop(zer_taken, outer_loop->_parent);
1520 // Plug in the true path
1521 _igvn.hash_delete(post_head);
1522 post_head->set_req(LoopNode::EntryControl, zer_taken);
1523 set_idom(post_head, zer_taken, dd_main_exit);
1524
1525 Arena *a = Thread::current()->resource_area();
1526 VectorSet visited(a);
1527 Node_Stack clones(a, main_head->back_control()->outcnt());
1528 // Step A3: Make the fall-in values to the post-loop come from the
1529 // fall-out values of the main-loop.
1530 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
1531 Node* main_phi = main_head->fast_out(i);
1532 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0) {
1533 Node *cur_phi = old_new[main_phi->_idx];
1534 Node *fallnew = clone_up_backedge_goo(main_head->back_control(),
1535 post_head->init_control(),
1536 main_phi->in(LoopNode::LoopBackControl),
1537 visited, clones);
1538 _igvn.hash_delete(cur_phi);
1539 cur_phi->set_req(LoopNode::EntryControl, fallnew);
1540 }
1541 }
1542
1543 // CastII for the new post loop:
1544 Node* castii = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1545 assert(castii != NULL, "no castII inserted");
1546
1547 return new_main_exit;
1548 }
1549
1550 //------------------------------is_invariant-----------------------------
1551 // Return true if n is invariant
1552 bool IdealLoopTree::is_invariant(Node* n) const {
1553 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1554 if (n_c->is_top()) return false;
1555 return !is_member(_phase->get_loop(n_c));
1556 }
1557
1558
1559 //------------------------------do_unroll--------------------------------------
1560 // Unroll the loop body one step - make each trip do 2 iterations.
1561 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1562 assert(LoopUnrollLimit, "");
1563 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1564 CountedLoopEndNode *loop_end = loop_head->loopexit();
1565 #ifndef PRODUCT
1566 if (PrintOpto && VerifyLoopOptimizations) {
1567 tty->print("Unrolling ");
1568 loop->dump_head();
1569 } else if (TraceLoopOpts) {
1570 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1571 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1572 } else {
1573 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1574 }
1575 loop->dump_head();
1576 }
1577
1578 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
1579 Arena* arena = Thread::current()->resource_area();
1580 Node_Stack stack(arena, C->live_nodes() >> 2);
1581 Node_List rpo_list;
1582 VectorSet visited(arena);
1583 visited.set(loop_head->_idx);
1584 rpo( loop_head, stack, visited, rpo_list );
1585 dump(loop, rpo_list.size(), rpo_list );
1586 }
1587 #endif
1588
1589 // Remember loop node count before unrolling to detect
1590 // if rounds of unroll,optimize are making progress
1591 loop_head->set_node_count_before_unroll(loop->_body.size());
1592
1593 Node *ctrl = loop_head->skip_strip_mined()->in(LoopNode::EntryControl);
1594 Node *limit = loop_head->limit();
1595 Node *init = loop_head->init_trip();
1596 Node *stride = loop_head->stride();
1597
1598 Node *opaq = NULL;
1599 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1600 // Search for zero-trip guard.
1601
1602 // Check the shape of the graph at the loop entry. If an inappropriate
1603 // graph shape is encountered, the compiler bails out loop unrolling;
1604 // compilation of the method will still succeed.
1605 if (!is_canonical_loop_entry(loop_head)) {
1606 return;
1607 }
1608 opaq = loop_head->skip_predicates()->in(0)->in(1)->in(1)->in(2);
1609 // Zero-trip test uses an 'opaque' node which is not shared.
1610 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1611 }
1612
1613 C->set_major_progress();
1614
1615 Node* new_limit = NULL;
1616 int stride_con = stride->get_int();
1617 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1618 uint old_trip_count = loop_head->trip_count();
1619 // Verify that unroll policy result is still valid.
1620 assert(old_trip_count > 1 &&
1621 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1622
1623 // Adjust loop limit to keep valid iterations number after unroll.
1624 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1625 // which may overflow.
1626 if (!adjust_min_trip) {
1627 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1628 "odd trip count for maximally unroll");
1629 // Don't need to adjust limit for maximally unroll since trip count is even.
1630 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1631 // Loop's limit is constant. Loop's init could be constant when pre-loop
1632 // become peeled iteration.
1633 jlong init_con = init->get_int();
1634 // We can keep old loop limit if iterations count stays the same:
1635 // old_trip_count == new_trip_count * 2
1636 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1637 // so we also don't need to adjust zero trip test.
1638 jlong limit_con = limit->get_int();
1639 // (stride_con*2) not overflow since stride_con <= 8.
1640 int new_stride_con = stride_con * 2;
1641 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1642 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1643 // New trip count should satisfy next conditions.
1644 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1645 uint new_trip_count = (uint)trip_count;
1646 adjust_min_trip = (old_trip_count != new_trip_count*2);
1647 }
1648
1649 if (adjust_min_trip) {
1650 // Step 2: Adjust the trip limit if it is called for.
1651 // The adjustment amount is -stride. Need to make sure if the
1652 // adjustment underflows or overflows, then the main loop is skipped.
1653 Node* cmp = loop_end->cmp_node();
1654 assert(cmp->in(2) == limit, "sanity");
1655 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1656
1657 // Verify that policy_unroll result is still valid.
1658 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1659 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1660 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1661
1662 if (limit->is_Con()) {
1663 // The check in policy_unroll and the assert above guarantee
1664 // no underflow if limit is constant.
1665 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1666 set_ctrl(new_limit, C->root());
1667 } else {
1668 // Limit is not constant.
1669 if (loop_head->unrolled_count() == 1) { // only for first unroll
1670 // Separate limit by Opaque node in case it is an incremented
1671 // variable from previous loop to avoid using pre-incremented
1672 // value which could increase register pressure.
1673 // Otherwise reorg_offsets() optimization will create a separate
1674 // Opaque node for each use of trip-counter and as result
1675 // zero trip guard limit will be different from loop limit.
1676 assert(has_ctrl(opaq), "should have it");
1677 Node* opaq_ctrl = get_ctrl(opaq);
1678 limit = new Opaque2Node( C, limit );
1679 register_new_node( limit, opaq_ctrl );
1680 }
1681 if ((stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo)) ||
1682 (stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi))) {
1683 // No underflow.
1684 new_limit = new SubINode(limit, stride);
1685 } else {
1686 // (limit - stride) may underflow.
1687 // Clamp the adjustment value with MININT or MAXINT:
1688 //
1689 // new_limit = limit-stride
1690 // if (stride > 0)
1691 // new_limit = (limit < new_limit) ? MININT : new_limit;
1692 // else
1693 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1694 //
1695 BoolTest::mask bt = loop_end->test_trip();
1696 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1697 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1698 set_ctrl(adj_max, C->root());
1699 Node* old_limit = NULL;
1700 Node* adj_limit = NULL;
1701 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1702 if (loop_head->unrolled_count() > 1 &&
1703 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1704 limit->in(CMoveNode::IfTrue) == adj_max &&
1705 bol->as_Bool()->_test._test == bt &&
1706 bol->in(1)->Opcode() == Op_CmpI &&
1707 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1708 // Loop was unrolled before.
1709 // Optimize the limit to avoid nested CMove:
1710 // use original limit as old limit.
1711 old_limit = bol->in(1)->in(1);
1712 // Adjust previous adjusted limit.
1713 adj_limit = limit->in(CMoveNode::IfFalse);
1714 adj_limit = new SubINode(adj_limit, stride);
1715 } else {
1716 old_limit = limit;
1717 adj_limit = new SubINode(limit, stride);
1718 }
1719 assert(old_limit != NULL && adj_limit != NULL, "");
1720 register_new_node( adj_limit, ctrl ); // adjust amount
1721 Node* adj_cmp = new CmpINode(old_limit, adj_limit);
1722 register_new_node( adj_cmp, ctrl );
1723 Node* adj_bool = new BoolNode(adj_cmp, bt);
1724 register_new_node( adj_bool, ctrl );
1725 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1726 }
1727 register_new_node(new_limit, ctrl);
1728 }
1729 assert(new_limit != NULL, "");
1730 // Replace in loop test.
1731 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1732 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1733 // Don't need to create new test since only one user.
1734 _igvn.hash_delete(cmp);
1735 cmp->set_req(2, new_limit);
1736 } else {
1737 // Create new test since it is shared.
1738 Node* ctrl2 = loop_end->in(0);
1739 Node* cmp2 = cmp->clone();
1740 cmp2->set_req(2, new_limit);
1741 register_new_node(cmp2, ctrl2);
1742 Node* bol2 = loop_end->in(1)->clone();
1743 bol2->set_req(1, cmp2);
1744 register_new_node(bol2, ctrl2);
1745 _igvn.replace_input_of(loop_end, 1, bol2);
1746 }
1747 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1748 // Make it a 1-trip test (means at least 2 trips).
1749
1750 // Guard test uses an 'opaque' node which is not shared. Hence I
1751 // can edit it's inputs directly. Hammer in the new limit for the
1752 // minimum-trip guard.
1753 assert(opaq->outcnt() == 1, "");
1754 _igvn.replace_input_of(opaq, 1, new_limit);
1755 }
1756
1757 // Adjust max trip count. The trip count is intentionally rounded
1758 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1759 // the main, unrolled, part of the loop will never execute as it is protected
1760 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1761 // and later determined that part of the unrolled loop was dead.
1762 loop_head->set_trip_count(old_trip_count / 2);
1763
1764 // Double the count of original iterations in the unrolled loop body.
1765 loop_head->double_unrolled_count();
1766
1767 // ---------
1768 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1769 // represents the odd iterations; since the loop trips an even number of
1770 // times its backedge is never taken. Kill the backedge.
1771 uint dd = dom_depth(loop_head);
1772 clone_loop(loop, old_new, dd, IgnoreStripMined);
1773
1774 // Make backedges of the clone equal to backedges of the original.
1775 // Make the fall-in from the original come from the fall-out of the clone.
1776 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1777 Node* phi = loop_head->fast_out(j);
1778 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1779 Node *newphi = old_new[phi->_idx];
1780 _igvn.hash_delete( phi );
1781 _igvn.hash_delete( newphi );
1782
1783 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1784 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1785 phi ->set_req(LoopNode::LoopBackControl, C->top());
1786 }
1787 }
1788 Node *clone_head = old_new[loop_head->_idx];
1789 _igvn.hash_delete( clone_head );
1790 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1791 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1792 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1793 loop->_head = clone_head; // New loop header
1794
1795 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1796 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1797
1798 // Kill the clone's backedge
1799 Node *newcle = old_new[loop_end->_idx];
1800 _igvn.hash_delete( newcle );
1801 Node *one = _igvn.intcon(1);
1802 set_ctrl(one, C->root());
1803 newcle->set_req(1, one);
1804 // Force clone into same loop body
1805 uint max = loop->_body.size();
1806 for( uint k = 0; k < max; k++ ) {
1807 Node *old = loop->_body.at(k);
1808 Node *nnn = old_new[old->_idx];
1809 loop->_body.push(nnn);
1810 if (!has_ctrl(old))
1811 set_loop(nnn, loop);
1812 }
1813
1814 loop->record_for_igvn();
1815 loop_head->clear_strip_mined();
1816
1817 #ifndef PRODUCT
1818 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
1819 tty->print("\nnew loop after unroll\n"); loop->dump_head();
1820 for (uint i = 0; i < loop->_body.size(); i++) {
1821 loop->_body.at(i)->dump();
1822 }
1823 if(C->clone_map().is_debug()) {
1824 tty->print("\nCloneMap\n");
1825 Dict* dict = C->clone_map().dict();
1826 DictI i(dict);
1827 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size());
1828 for (int ii = 0; i.test(); ++i, ++ii) {
1829 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key));
1830 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen());
1831 if (ii % 10 == 9) {
1832 tty->print_cr(" ");
1833 }
1834 }
1835 tty->print_cr(" ");
1836 }
1837 }
1838 #endif
1839
1840 }
1841
1842 //------------------------------do_maximally_unroll----------------------------
1843
1844 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1845 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1846 assert(cl->has_exact_trip_count(), "trip count is not exact");
1847 assert(cl->trip_count() > 0, "");
1848 #ifndef PRODUCT
1849 if (TraceLoopOpts) {
1850 tty->print("MaxUnroll %d ", cl->trip_count());
1851 loop->dump_head();
1852 }
1853 #endif
1854
1855 // If loop is tripping an odd number of times, peel odd iteration
1856 if ((cl->trip_count() & 1) == 1) {
1857 do_peeling(loop, old_new);
1858 }
1859
1860 // Now its tripping an even number of times remaining. Double loop body.
1861 // Do not adjust pre-guards; they are not needed and do not exist.
1862 if (cl->trip_count() > 0) {
1863 assert((cl->trip_count() & 1) == 0, "missed peeling");
1864 do_unroll(loop, old_new, false);
1865 }
1866 }
1867
1868 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) {
1869 if (SuperWordReductions == false) return;
1870
1871 CountedLoopNode* loop_head = loop->_head->as_CountedLoop();
1872 if (loop_head->unrolled_count() > 1) {
1873 return;
1874 }
1875
1876 Node* trip_phi = loop_head->phi();
1877 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) {
1878 Node* phi = loop_head->fast_out(i);
1879 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) {
1880 // For definitions which are loop inclusive and not tripcounts.
1881 Node* def_node = phi->in(LoopNode::LoopBackControl);
1882
1883 if (def_node != NULL) {
1884 Node* n_ctrl = get_ctrl(def_node);
1885 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) {
1886 // Now test it to see if it fits the standard pattern for a reduction operator.
1887 int opc = def_node->Opcode();
1888 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) {
1889 if (!def_node->is_reduction()) { // Not marked yet
1890 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it
1891 bool ok = false;
1892 for (unsigned j = 1; j < def_node->req(); j++) {
1893 Node* in = def_node->in(j);
1894 if (in == phi) {
1895 ok = true;
1896 break;
1897 }
1898 }
1899
1900 // do nothing if we did not match the initial criteria
1901 if (ok == false) {
1902 continue;
1903 }
1904
1905 // The result of the reduction must not be used in the loop
1906 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) {
1907 Node* u = def_node->fast_out(i);
1908 if (!loop->is_member(get_loop(ctrl_or_self(u)))) {
1909 continue;
1910 }
1911 if (u == phi) {
1912 continue;
1913 }
1914 ok = false;
1915 }
1916
1917 // iff the uses conform
1918 if (ok) {
1919 def_node->add_flag(Node::Flag_is_reduction);
1920 loop_head->mark_has_reductions();
1921 }
1922 }
1923 }
1924 }
1925 }
1926 }
1927 }
1928 }
1929
1930 //------------------------------adjust_limit-----------------------------------
1931 // Helper function for add_constraint().
1932 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1933 // Compute "I :: (limit-offset)/scale"
1934 Node *con = new SubINode(rc_limit, offset);
1935 register_new_node(con, pre_ctrl);
1936 Node *X = new DivINode(0, con, scale);
1937 register_new_node(X, pre_ctrl);
1938
1939 // Adjust loop limit
1940 loop_limit = (stride_con > 0)
1941 ? (Node*)(new MinINode(loop_limit, X))
1942 : (Node*)(new MaxINode(loop_limit, X));
1943 register_new_node(loop_limit, pre_ctrl);
1944 return loop_limit;
1945 }
1946
1947 //------------------------------add_constraint---------------------------------
1948 // Constrain the main loop iterations so the conditions:
1949 // low_limit <= scale_con * I + offset < upper_limit
1950 // always holds true. That is, either increase the number of iterations in
1951 // the pre-loop or the post-loop until the condition holds true in the main
1952 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1953 // stride and scale are constants (offset and limit often are).
1954 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 ) {
1955 // For positive stride, the pre-loop limit always uses a MAX function
1956 // and the main loop a MIN function. For negative stride these are
1957 // reversed.
1958
1959 // Also for positive stride*scale the affine function is increasing, so the
1960 // pre-loop must check for underflow and the post-loop for overflow.
1961 // Negative stride*scale reverses this; pre-loop checks for overflow and
1962 // post-loop for underflow.
1963
1964 Node *scale = _igvn.intcon(scale_con);
1965 set_ctrl(scale, C->root());
1966
1967 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1968 // The overflow limit: scale*I+offset < upper_limit
1969 // For main-loop compute
1970 // ( if (scale > 0) /* and stride > 0 */
1971 // I < (upper_limit-offset)/scale
1972 // else /* scale < 0 and stride < 0 */
1973 // I > (upper_limit-offset)/scale
1974 // )
1975 //
1976 // (upper_limit-offset) may overflow or underflow.
1977 // But it is fine since main loop will either have
1978 // less iterations or will be skipped in such case.
1979 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1980
1981 // The underflow limit: low_limit <= scale*I+offset.
1982 // For pre-loop compute
1983 // NOT(scale*I+offset >= low_limit)
1984 // scale*I+offset < low_limit
1985 // ( if (scale > 0) /* and stride > 0 */
1986 // I < (low_limit-offset)/scale
1987 // else /* scale < 0 and stride < 0 */
1988 // I > (low_limit-offset)/scale
1989 // )
1990
1991 if (low_limit->get_int() == -max_jint) {
1992 // We need this guard when scale*pre_limit+offset >= limit
1993 // due to underflow. So we need execute pre-loop until
1994 // scale*I+offset >= min_int. But (min_int-offset) will
1995 // underflow when offset > 0 and X will be > original_limit
1996 // when stride > 0. To avoid it we replace positive offset with 0.
1997 //
1998 // Also (min_int+1 == -max_int) is used instead of min_int here
1999 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
2000 Node* shift = _igvn.intcon(31);
2001 set_ctrl(shift, C->root());
2002 Node* sign = new RShiftINode(offset, shift);
2003 register_new_node(sign, pre_ctrl);
2004 offset = new AndINode(offset, sign);
2005 register_new_node(offset, pre_ctrl);
2006 } else {
2007 assert(low_limit->get_int() == 0, "wrong low limit for range check");
2008 // The only problem we have here when offset == min_int
2009 // since (0-min_int) == min_int. It may be fine for stride > 0
2010 // but for stride < 0 X will be < original_limit. To avoid it
2011 // max(pre_limit, original_limit) is used in do_range_check().
2012 }
2013 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
2014 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
2015
2016 } else { // stride_con*scale_con < 0
2017 // For negative stride*scale pre-loop checks for overflow and
2018 // post-loop for underflow.
2019 //
2020 // The overflow limit: scale*I+offset < upper_limit
2021 // For pre-loop compute
2022 // NOT(scale*I+offset < upper_limit)
2023 // scale*I+offset >= upper_limit
2024 // scale*I+offset+1 > upper_limit
2025 // ( if (scale < 0) /* and stride > 0 */
2026 // I < (upper_limit-(offset+1))/scale
2027 // else /* scale > 0 and stride < 0 */
2028 // I > (upper_limit-(offset+1))/scale
2029 // )
2030 //
2031 // (upper_limit-offset-1) may underflow or overflow.
2032 // To avoid it min(pre_limit, original_limit) is used
2033 // in do_range_check() for stride > 0 and max() for < 0.
2034 Node *one = _igvn.intcon(1);
2035 set_ctrl(one, C->root());
2036
2037 Node *plus_one = new AddINode(offset, one);
2038 register_new_node( plus_one, pre_ctrl );
2039 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
2040 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
2041
2042 if (low_limit->get_int() == -max_jint) {
2043 // We need this guard when scale*main_limit+offset >= limit
2044 // due to underflow. So we need execute main-loop while
2045 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
2046 // underflow when (offset+1) > 0 and X will be < main_limit
2047 // when scale < 0 (and stride > 0). To avoid it we replace
2048 // positive (offset+1) with 0.
2049 //
2050 // Also (min_int+1 == -max_int) is used instead of min_int here
2051 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
2052 Node* shift = _igvn.intcon(31);
2053 set_ctrl(shift, C->root());
2054 Node* sign = new RShiftINode(plus_one, shift);
2055 register_new_node(sign, pre_ctrl);
2056 plus_one = new AndINode(plus_one, sign);
2057 register_new_node(plus_one, pre_ctrl);
2058 } else {
2059 assert(low_limit->get_int() == 0, "wrong low limit for range check");
2060 // The only problem we have here when offset == max_int
2061 // since (max_int+1) == min_int and (0-min_int) == min_int.
2062 // But it is fine since main loop will either have
2063 // less iterations or will be skipped in such case.
2064 }
2065 // The underflow limit: low_limit <= scale*I+offset.
2066 // For main-loop compute
2067 // scale*I+offset+1 > low_limit
2068 // ( if (scale < 0) /* and stride > 0 */
2069 // I < (low_limit-(offset+1))/scale
2070 // else /* scale > 0 and stride < 0 */
2071 // I > (low_limit-(offset+1))/scale
2072 // )
2073
2074 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
2075 }
2076 }
2077
2078
2079 //------------------------------is_scaled_iv---------------------------------
2080 // Return true if exp is a constant times an induction var
2081 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
2082 if (exp == iv) {
2083 if (p_scale != NULL) {
2084 *p_scale = 1;
2085 }
2086 return true;
2087 }
2088 int opc = exp->Opcode();
2089 if (opc == Op_MulI) {
2090 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
2091 if (p_scale != NULL) {
2092 *p_scale = exp->in(2)->get_int();
2093 }
2094 return true;
2095 }
2096 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
2097 if (p_scale != NULL) {
2098 *p_scale = exp->in(1)->get_int();
2099 }
2100 return true;
2101 }
2102 } else if (opc == Op_LShiftI) {
2103 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
2104 if (p_scale != NULL) {
2105 *p_scale = 1 << exp->in(2)->get_int();
2106 }
2107 return true;
2108 }
2109 }
2110 return false;
2111 }
2112
2113 //-----------------------------is_scaled_iv_plus_offset------------------------------
2114 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
2115 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
2116 if (is_scaled_iv(exp, iv, p_scale)) {
2117 if (p_offset != NULL) {
2118 Node *zero = _igvn.intcon(0);
2119 set_ctrl(zero, C->root());
2120 *p_offset = zero;
2121 }
2122 return true;
2123 }
2124 int opc = exp->Opcode();
2125 if (opc == Op_AddI) {
2126 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
2127 if (p_offset != NULL) {
2128 *p_offset = exp->in(2);
2129 }
2130 return true;
2131 }
2132 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
2133 if (p_offset != NULL) {
2134 *p_offset = exp->in(1);
2135 }
2136 return true;
2137 }
2138 if (exp->in(2)->is_Con()) {
2139 Node* offset2 = NULL;
2140 if (depth < 2 &&
2141 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
2142 p_offset != NULL ? &offset2 : NULL, depth+1)) {
2143 if (p_offset != NULL) {
2144 Node *ctrl_off2 = get_ctrl(offset2);
2145 Node* offset = new AddINode(offset2, exp->in(2));
2146 register_new_node(offset, ctrl_off2);
2147 *p_offset = offset;
2148 }
2149 return true;
2150 }
2151 }
2152 } else if (opc == Op_SubI) {
2153 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
2154 if (p_offset != NULL) {
2155 Node *zero = _igvn.intcon(0);
2156 set_ctrl(zero, C->root());
2157 Node *ctrl_off = get_ctrl(exp->in(2));
2158 Node* offset = new SubINode(zero, exp->in(2));
2159 register_new_node(offset, ctrl_off);
2160 *p_offset = offset;
2161 }
2162 return true;
2163 }
2164 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
2165 if (p_offset != NULL) {
2166 *p_scale *= -1;
2167 *p_offset = exp->in(1);
2168 }
2169 return true;
2170 }
2171 }
2172 return false;
2173 }
2174
2175 // Same as PhaseIdealLoop::duplicate_predicates() but for range checks
2176 // eliminated by iteration splitting.
2177 Node* PhaseIdealLoop::add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl,
2178 Node* predicate_proj, int scale_con, Node* offset,
2179 Node* limit, jint stride_con) {
2180 bool overflow = false;
2181 BoolNode* bol = rc_predicate(loop, predicate_proj, scale_con, offset, cl->init_trip(), NULL, stride_con, limit, (stride_con > 0) != (scale_con > 0), overflow);
2182 Node* opaque_bol = new Opaque4Node(C, bol, _igvn.intcon(1));
2183 register_new_node(opaque_bol, predicate_proj);
2184 IfNode* new_iff = NULL;
2185 if (overflow) {
2186 new_iff = new IfNode(predicate_proj, bol, PROB_MAX, COUNT_UNKNOWN);
2187 } else {
2188 new_iff = new RangeCheckNode(predicate_proj, bol, PROB_MAX, COUNT_UNKNOWN);
2189 }
2190 register_control(new_iff, loop->_parent, predicate_proj);
2191 Node* iffalse = new IfFalseNode(new_iff);
2192 register_control(iffalse, _ltree_root, new_iff);
2193 ProjNode* iftrue = new IfTrueNode(new_iff);
2194 register_control(iftrue, loop->_parent, new_iff);
2195 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
2196 register_new_node(frame, C->start());
2197 Node* halt = new HaltNode(iffalse, frame);
2198 register_control(halt, _ltree_root, iffalse);
2199 C->root()->add_req(halt);
2200 return iftrue;
2201 }
2202
2203 //------------------------------do_range_check---------------------------------
2204 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
2205 int PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
2206 #ifndef PRODUCT
2207 if (PrintOpto && VerifyLoopOptimizations) {
2208 tty->print("Range Check Elimination ");
2209 loop->dump_head();
2210 } else if (TraceLoopOpts) {
2211 tty->print("RangeCheck ");
2212 loop->dump_head();
2213 }
2214 #endif
2215 assert(RangeCheckElimination, "");
2216 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2217 // If we fail before trying to eliminate range checks, set multiversion state
2218 int closed_range_checks = 1;
2219
2220 // protect against stride not being a constant
2221 if (!cl->stride_is_con())
2222 return closed_range_checks;
2223
2224 // Find the trip counter; we are iteration splitting based on it
2225 Node *trip_counter = cl->phi();
2226 // Find the main loop limit; we will trim it's iterations
2227 // to not ever trip end tests
2228 Node *main_limit = cl->limit();
2229
2230 // Check graph shape. Cannot optimize a loop if zero-trip
2231 // Opaque1 node is optimized away and then another round
2232 // of loop opts attempted.
2233 if (!is_canonical_loop_entry(cl)) {
2234 return closed_range_checks;
2235 }
2236
2237 // Need to find the main-loop zero-trip guard
2238 Node *ctrl = cl->skip_predicates();
2239 Node *iffm = ctrl->in(0);
2240 Node *opqzm = iffm->in(1)->in(1)->in(2);
2241 assert(opqzm->in(1) == main_limit, "do not understand situation");
2242
2243 // Find the pre-loop limit; we will expand its iterations to
2244 // not ever trip low tests.
2245 Node *p_f = iffm->in(0);
2246 // pre loop may have been optimized out
2247 if (p_f->Opcode() != Op_IfFalse) {
2248 return closed_range_checks;
2249 }
2250 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2251 assert(pre_end->loopnode()->is_pre_loop(), "");
2252 Node *pre_opaq1 = pre_end->limit();
2253 // Occasionally it's possible for a pre-loop Opaque1 node to be
2254 // optimized away and then another round of loop opts attempted.
2255 // We can not optimize this particular loop in that case.
2256 if (pre_opaq1->Opcode() != Op_Opaque1)
2257 return closed_range_checks;
2258 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2259 Node *pre_limit = pre_opaq->in(1);
2260
2261 // Where do we put new limit calculations
2262 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2263
2264 // Ensure the original loop limit is available from the
2265 // pre-loop Opaque1 node.
2266 Node *orig_limit = pre_opaq->original_loop_limit();
2267 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
2268 return closed_range_checks;
2269
2270 // Must know if its a count-up or count-down loop
2271
2272 int stride_con = cl->stride_con();
2273 Node *zero = _igvn.intcon(0);
2274 Node *one = _igvn.intcon(1);
2275 // Use symmetrical int range [-max_jint,max_jint]
2276 Node *mini = _igvn.intcon(-max_jint);
2277 set_ctrl(zero, C->root());
2278 set_ctrl(one, C->root());
2279 set_ctrl(mini, C->root());
2280
2281 // Range checks that do not dominate the loop backedge (ie.
2282 // conditionally executed) can lengthen the pre loop limit beyond
2283 // the original loop limit. To prevent this, the pre limit is
2284 // (for stride > 0) MINed with the original loop limit (MAXed
2285 // stride < 0) when some range_check (rc) is conditionally
2286 // executed.
2287 bool conditional_rc = false;
2288
2289 // Count number of range checks and reduce by load range limits, if zero,
2290 // the loop is in canonical form to multiversion.
2291 closed_range_checks = 0;
2292
2293 Node* predicate_proj = cl->skip_strip_mined()->in(LoopNode::EntryControl);
2294 assert(predicate_proj->is_Proj() && predicate_proj->in(0)->is_If(), "if projection only");
2295 // Check loop body for tests of trip-counter plus loop-invariant vs loop-variant.
2296 for( uint i = 0; i < loop->_body.size(); i++ ) {
2297 Node *iff = loop->_body[i];
2298 if (iff->Opcode() == Op_If ||
2299 iff->Opcode() == Op_RangeCheck) { // Test?
2300 // Test is an IfNode, has 2 projections. If BOTH are in the loop
2301 // we need loop unswitching instead of iteration splitting.
2302 closed_range_checks++;
2303 Node *exit = loop->is_loop_exit(iff);
2304 if( !exit ) continue;
2305 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
2306
2307 // Get boolean condition to test
2308 Node *i1 = iff->in(1);
2309 if( !i1->is_Bool() ) continue;
2310 BoolNode *bol = i1->as_Bool();
2311 BoolTest b_test = bol->_test;
2312 // Flip sense of test if exit condition is flipped
2313 if( flip )
2314 b_test = b_test.negate();
2315
2316 // Get compare
2317 Node *cmp = bol->in(1);
2318
2319 // Look for trip_counter + offset vs limit
2320 Node *rc_exp = cmp->in(1);
2321 Node *limit = cmp->in(2);
2322 int scale_con= 1; // Assume trip counter not scaled
2323
2324 Node *limit_c = get_ctrl(limit);
2325 if( loop->is_member(get_loop(limit_c) ) ) {
2326 // Compare might have operands swapped; commute them
2327 b_test = b_test.commute();
2328 rc_exp = cmp->in(2);
2329 limit = cmp->in(1);
2330 limit_c = get_ctrl(limit);
2331 if( loop->is_member(get_loop(limit_c) ) )
2332 continue; // Both inputs are loop varying; cannot RCE
2333 }
2334 // Here we know 'limit' is loop invariant
2335
2336 // 'limit' maybe pinned below the zero trip test (probably from a
2337 // previous round of rce), in which case, it can't be used in the
2338 // zero trip test expression which must occur before the zero test's if.
2339 if (is_dominator(ctrl, limit_c)) {
2340 continue; // Don't rce this check but continue looking for other candidates.
2341 }
2342
2343 // Check for scaled induction variable plus an offset
2344 Node *offset = NULL;
2345
2346 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
2347 continue;
2348 }
2349
2350 Node *offset_c = get_ctrl(offset);
2351 if( loop->is_member( get_loop(offset_c) ) )
2352 continue; // Offset is not really loop invariant
2353 // Here we know 'offset' is loop invariant.
2354
2355 // As above for the 'limit', the 'offset' maybe pinned below the
2356 // zero trip test.
2357 if (is_dominator(ctrl, offset_c)) {
2358 continue; // Don't rce this check but continue looking for other candidates.
2359 }
2360 #ifdef ASSERT
2361 if (TraceRangeLimitCheck) {
2362 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
2363 bol->dump(2);
2364 }
2365 #endif
2366 // At this point we have the expression as:
2367 // scale_con * trip_counter + offset :: limit
2368 // where scale_con, offset and limit are loop invariant. Trip_counter
2369 // monotonically increases by stride_con, a constant. Both (or either)
2370 // stride_con and scale_con can be negative which will flip about the
2371 // sense of the test.
2372
2373 // Adjust pre and main loop limits to guard the correct iteration set
2374 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
2375 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
2376 // The underflow and overflow limits: 0 <= scale*I+offset < limit
2377 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
2378 // (0-offset)/scale could be outside of loop iterations range.
2379 conditional_rc = true;
2380 predicate_proj = add_range_check_predicate(loop, cl, predicate_proj, scale_con, offset, limit, stride_con);
2381 } else {
2382 if (PrintOpto) {
2383 tty->print_cr("missed RCE opportunity");
2384 }
2385 continue; // In release mode, ignore it
2386 }
2387 } else { // Otherwise work on normal compares
2388 switch( b_test._test ) {
2389 case BoolTest::gt:
2390 // Fall into GE case
2391 case BoolTest::ge:
2392 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
2393 scale_con = -scale_con;
2394 offset = new SubINode( zero, offset );
2395 register_new_node( offset, pre_ctrl );
2396 limit = new SubINode( zero, limit );
2397 register_new_node( limit, pre_ctrl );
2398 // Fall into LE case
2399 case BoolTest::le:
2400 if (b_test._test != BoolTest::gt) {
2401 // Convert X <= Y to X < Y+1
2402 limit = new AddINode( limit, one );
2403 register_new_node( limit, pre_ctrl );
2404 }
2405 // Fall into LT case
2406 case BoolTest::lt:
2407 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
2408 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
2409 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
2410 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
2411 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
2412 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
2413 // still be outside of loop range.
2414 conditional_rc = true;
2415 break;
2416 default:
2417 if (PrintOpto) {
2418 tty->print_cr("missed RCE opportunity");
2419 }
2420 continue; // Unhandled case
2421 }
2422 }
2423
2424 // Kill the eliminated test
2425 C->set_major_progress();
2426 Node *kill_con = _igvn.intcon( 1-flip );
2427 set_ctrl(kill_con, C->root());
2428 _igvn.replace_input_of(iff, 1, kill_con);
2429 // Find surviving projection
2430 assert(iff->is_If(), "");
2431 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2432 // Find loads off the surviving projection; remove their control edge
2433 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2434 Node* cd = dp->fast_out(i); // Control-dependent node
2435 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop
2436 // Allow the load to float around in the loop, or before it
2437 // but NOT before the pre-loop.
2438 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
2439 --i;
2440 --imax;
2441 }
2442 }
2443 if (limit->Opcode() == Op_LoadRange) {
2444 closed_range_checks--;
2445 }
2446
2447 } // End of is IF
2448
2449 }
2450 if (predicate_proj != cl->skip_strip_mined()->in(LoopNode::EntryControl)) {
2451 _igvn.replace_input_of(cl->skip_strip_mined(), LoopNode::EntryControl, predicate_proj);
2452 set_idom(cl->skip_strip_mined(), predicate_proj, dom_depth(cl->skip_strip_mined()));
2453 }
2454
2455 // Update loop limits
2456 if (conditional_rc) {
2457 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit)
2458 : (Node*)new MaxINode(pre_limit, orig_limit);
2459 register_new_node(pre_limit, pre_ctrl);
2460 }
2461 _igvn.replace_input_of(pre_opaq, 1, pre_limit);
2462
2463 // Note:: we are making the main loop limit no longer precise;
2464 // need to round up based on stride.
2465 cl->set_nonexact_trip_count();
2466 Node *main_cle = cl->loopexit();
2467 Node *main_bol = main_cle->in(1);
2468 // Hacking loop bounds; need private copies of exit test
2469 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2470 main_bol = main_bol->clone();// Clone a private BoolNode
2471 register_new_node( main_bol, main_cle->in(0) );
2472 _igvn.replace_input_of(main_cle, 1, main_bol);
2473 }
2474 Node *main_cmp = main_bol->in(1);
2475 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2476 main_cmp = main_cmp->clone();// Clone a private CmpNode
2477 register_new_node( main_cmp, main_cle->in(0) );
2478 _igvn.replace_input_of(main_bol, 1, main_cmp);
2479 }
2480 // Hack the now-private loop bounds
2481 _igvn.replace_input_of(main_cmp, 2, main_limit);
2482 // The OpaqueNode is unshared by design
2483 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2484 _igvn.replace_input_of(opqzm, 1, main_limit);
2485
2486 return closed_range_checks;
2487 }
2488
2489 //------------------------------has_range_checks-------------------------------
2490 // Check to see if RCE cleaned the current loop of range-checks.
2491 void PhaseIdealLoop::has_range_checks(IdealLoopTree *loop) {
2492 assert(RangeCheckElimination, "");
2493
2494 // skip if not a counted loop
2495 if (!loop->is_counted()) return;
2496
2497 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2498
2499 // skip this loop if it is already checked
2500 if (cl->has_been_range_checked()) return;
2501
2502 // Now check for existence of range checks
2503 for (uint i = 0; i < loop->_body.size(); i++) {
2504 Node *iff = loop->_body[i];
2505 int iff_opc = iff->Opcode();
2506 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
2507 cl->mark_has_range_checks();
2508 break;
2509 }
2510 }
2511 cl->set_has_been_range_checked();
2512 }
2513
2514 //-------------------------multi_version_post_loops----------------------------
2515 // Check the range checks that remain, if simple, use the bounds to guard
2516 // which version to a post loop we execute, one with range checks or one without
2517 bool PhaseIdealLoop::multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop) {
2518 bool multi_version_succeeded = false;
2519 assert(RangeCheckElimination, "");
2520 CountedLoopNode *legacy_cl = legacy_loop->_head->as_CountedLoop();
2521 assert(legacy_cl->is_post_loop(), "");
2522
2523 // Check for existence of range checks using the unique instance to make a guard with
2524 Unique_Node_List worklist;
2525 for (uint i = 0; i < legacy_loop->_body.size(); i++) {
2526 Node *iff = legacy_loop->_body[i];
2527 int iff_opc = iff->Opcode();
2528 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
2529 worklist.push(iff);
2530 }
2531 }
2532
2533 // Find RCE'd post loop so that we can stage its guard.
2534 if (!is_canonical_loop_entry(legacy_cl)) return multi_version_succeeded;
2535 Node* ctrl = legacy_cl->in(LoopNode::EntryControl);
2536 Node* iffm = ctrl->in(0);
2537
2538 // Now we test that both the post loops are connected
2539 Node* post_loop_region = iffm->in(0);
2540 if (post_loop_region == NULL) return multi_version_succeeded;
2541 if (!post_loop_region->is_Region()) return multi_version_succeeded;
2542 Node* covering_region = post_loop_region->in(RegionNode::Control+1);
2543 if (covering_region == NULL) return multi_version_succeeded;
2544 if (!covering_region->is_Region()) return multi_version_succeeded;
2545 Node* p_f = covering_region->in(RegionNode::Control);
2546 if (p_f == NULL) return multi_version_succeeded;
2547 if (!p_f->is_IfFalse()) return multi_version_succeeded;
2548 if (!p_f->in(0)->is_CountedLoopEnd()) return multi_version_succeeded;
2549 CountedLoopEndNode* rce_loop_end = p_f->in(0)->as_CountedLoopEnd();
2550 if (rce_loop_end == NULL) return multi_version_succeeded;
2551 CountedLoopNode* rce_cl = rce_loop_end->loopnode();
2552 if (rce_cl == NULL || !rce_cl->is_post_loop()) return multi_version_succeeded;
2553 CountedLoopNode *known_rce_cl = rce_loop->_head->as_CountedLoop();
2554 if (rce_cl != known_rce_cl) return multi_version_succeeded;
2555
2556 // Then we fetch the cover entry test
2557 ctrl = rce_cl->in(LoopNode::EntryControl);
2558 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return multi_version_succeeded;
2559
2560 #ifndef PRODUCT
2561 if (TraceLoopOpts) {
2562 tty->print("PostMultiVersion\n");
2563 rce_loop->dump_head();
2564 legacy_loop->dump_head();
2565 }
2566 #endif
2567
2568 // Now fetch the limit we want to compare against
2569 Node *limit = rce_cl->limit();
2570 bool first_time = true;
2571
2572 // If we got this far, we identified the post loop which has been RCE'd and
2573 // we have a work list. Now we will try to transform the if guard to cause
2574 // the loop pair to be multi version executed with the determination left to runtime
2575 // or the optimizer if full information is known about the given arrays at compile time.
2576 Node *last_min = NULL;
2577 multi_version_succeeded = true;
2578 while (worklist.size()) {
2579 Node* rc_iffm = worklist.pop();
2580 if (rc_iffm->is_If()) {
2581 Node *rc_bolzm = rc_iffm->in(1);
2582 if (rc_bolzm->is_Bool()) {
2583 Node *rc_cmpzm = rc_bolzm->in(1);
2584 if (rc_cmpzm->is_Cmp()) {
2585 Node *rc_left = rc_cmpzm->in(2);
2586 if (rc_left->Opcode() != Op_LoadRange) {
2587 multi_version_succeeded = false;
2588 break;
2589 }
2590 if (first_time) {
2591 last_min = rc_left;
2592 first_time = false;
2593 } else {
2594 Node *cur_min = new MinINode(last_min, rc_left);
2595 last_min = cur_min;
2596 _igvn.register_new_node_with_optimizer(last_min);
2597 }
2598 }
2599 }
2600 }
2601 }
2602
2603 // All we have to do is update the limit of the rce loop
2604 // with the min of our expression and the current limit.
2605 // We will use this expression to replace the current limit.
2606 if (last_min && multi_version_succeeded) {
2607 Node *cur_min = new MinINode(last_min, limit);
2608 _igvn.register_new_node_with_optimizer(cur_min);
2609 Node *cmp_node = rce_loop_end->cmp_node();
2610 _igvn.replace_input_of(cmp_node, 2, cur_min);
2611 set_ctrl(cur_min, ctrl);
2612 set_loop(cur_min, rce_loop->_parent);
2613
2614 legacy_cl->mark_is_multiversioned();
2615 rce_cl->mark_is_multiversioned();
2616 multi_version_succeeded = true;
2617
2618 C->set_major_progress();
2619 }
2620
2621 return multi_version_succeeded;
2622 }
2623
2624 //-------------------------poison_rce_post_loop--------------------------------
2625 // Causes the rce'd post loop to be optimized away if multiversioning fails
2626 void PhaseIdealLoop::poison_rce_post_loop(IdealLoopTree *rce_loop) {
2627 CountedLoopNode *rce_cl = rce_loop->_head->as_CountedLoop();
2628 Node* ctrl = rce_cl->in(LoopNode::EntryControl);
2629 if (ctrl->is_IfTrue() || ctrl->is_IfFalse()) {
2630 Node* iffm = ctrl->in(0);
2631 if (iffm->is_If()) {
2632 Node* cur_bool = iffm->in(1);
2633 if (cur_bool->is_Bool()) {
2634 Node* cur_cmp = cur_bool->in(1);
2635 if (cur_cmp->is_Cmp()) {
2636 BoolTest::mask new_test = BoolTest::gt;
2637 BoolNode *new_bool = new BoolNode(cur_cmp, new_test);
2638 _igvn.replace_node(cur_bool, new_bool);
2639 _igvn._worklist.push(new_bool);
2640 Node* left_op = cur_cmp->in(1);
2641 _igvn.replace_input_of(cur_cmp, 2, left_op);
2642 C->set_major_progress();
2643 }
2644 }
2645 }
2646 }
2647 }
2648
2649 //------------------------------DCE_loop_body----------------------------------
2650 // Remove simplistic dead code from loop body
2651 void IdealLoopTree::DCE_loop_body() {
2652 for( uint i = 0; i < _body.size(); i++ )
2653 if( _body.at(i)->outcnt() == 0 )
2654 _body.map( i--, _body.pop() );
2655 }
2656
2657
2658 //------------------------------adjust_loop_exit_prob--------------------------
2659 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2660 // Replace with a 1-in-10 exit guess.
2661 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2662 Node *test = tail();
2663 while( test != _head ) {
2664 uint top = test->Opcode();
2665 if( top == Op_IfTrue || top == Op_IfFalse ) {
2666 int test_con = ((ProjNode*)test)->_con;
2667 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2668 IfNode *iff = test->in(0)->as_If();
2669 if( iff->outcnt() == 2 ) { // Ignore dead tests
2670 Node *bol = iff->in(1);
2671 if( bol && bol->req() > 1 && bol->in(1) &&
2672 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2673 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2674 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2675 (bol->in(1)->Opcode() == Op_CompareAndExchangeB ) ||
2676 (bol->in(1)->Opcode() == Op_CompareAndExchangeS ) ||
2677 (bol->in(1)->Opcode() == Op_CompareAndExchangeI ) ||
2678 (bol->in(1)->Opcode() == Op_CompareAndExchangeL ) ||
2679 (bol->in(1)->Opcode() == Op_CompareAndExchangeP ) ||
2680 (bol->in(1)->Opcode() == Op_CompareAndExchangeN ) ||
2681 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapB ) ||
2682 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapS ) ||
2683 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapI ) ||
2684 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapL ) ||
2685 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapP ) ||
2686 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapN ) ||
2687 (bol->in(1)->Opcode() == Op_CompareAndSwapB ) ||
2688 (bol->in(1)->Opcode() == Op_CompareAndSwapS ) ||
2689 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2690 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2691 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2692 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2693 return; // Allocation loops RARELY take backedge
2694 // Find the OTHER exit path from the IF
2695 Node* ex = iff->proj_out(1-test_con);
2696 float p = iff->_prob;
2697 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2698 if( top == Op_IfTrue ) {
2699 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2700 iff->_prob = PROB_STATIC_FREQUENT;
2701 }
2702 } else {
2703 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2704 iff->_prob = PROB_STATIC_INFREQUENT;
2705 }
2706 }
2707 }
2708 }
2709 }
2710 test = phase->idom(test);
2711 }
2712 }
2713
2714 #ifdef ASSERT
2715 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) {
2716 Node *ctrl = cl->skip_predicates();
2717 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
2718 Node *iffm = ctrl->in(0);
2719 assert(iffm->Opcode() == Op_If, "");
2720 Node *p_f = iffm->in(0);
2721 assert(p_f->Opcode() == Op_IfFalse, "");
2722 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2723 assert(pre_end->loopnode()->is_pre_loop(), "");
2724 return pre_end->loopnode();
2725 }
2726 #endif
2727
2728 // Remove the main and post loops and make the pre loop execute all
2729 // iterations. Useful when the pre loop is found empty.
2730 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) {
2731 CountedLoopEndNode* pre_end = cl->loopexit();
2732 Node* pre_cmp = pre_end->cmp_node();
2733 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) {
2734 // Only safe to remove the main loop if the compiler optimized it
2735 // out based on an unknown number of iterations
2736 return;
2737 }
2738
2739 // Can we find the main loop?
2740 if (_next == NULL) {
2741 return;
2742 }
2743
2744 Node* next_head = _next->_head;
2745 if (!next_head->is_CountedLoop()) {
2746 return;
2747 }
2748
2749 CountedLoopNode* main_head = next_head->as_CountedLoop();
2750 if (!main_head->is_main_loop()) {
2751 return;
2752 }
2753
2754 assert(locate_pre_from_main(main_head) == cl, "bad main loop");
2755 Node* main_iff = main_head->skip_predicates()->in(0);
2756
2757 // Remove the Opaque1Node of the pre loop and make it execute all iterations
2758 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2));
2759 // Remove the Opaque1Node of the main loop so it can be optimized out
2760 Node* main_cmp = main_iff->in(1)->in(1);
2761 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?");
2762 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1));
2763 }
2764
2765 //------------------------------policy_do_remove_empty_loop--------------------
2766 // Micro-benchmark spamming. Policy is to always remove empty loops.
2767 // The 'DO' part is to replace the trip counter with the value it will
2768 // have on the last iteration. This will break the loop.
2769 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2770 // Minimum size must be empty loop
2771 if (_body.size() > EMPTY_LOOP_SIZE)
2772 return false;
2773
2774 if (!_head->is_CountedLoop())
2775 return false; // Dead loop
2776 CountedLoopNode *cl = _head->as_CountedLoop();
2777 if (!cl->is_valid_counted_loop())
2778 return false; // Malformed loop
2779 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2780 return false; // Infinite loop
2781
2782 if (cl->is_pre_loop()) {
2783 // If the loop we are removing is a pre-loop then the main and
2784 // post loop can be removed as well
2785 remove_main_post_loops(cl, phase);
2786 }
2787
2788 #ifdef ASSERT
2789 // Ensure only one phi which is the iv.
2790 Node* iv = NULL;
2791 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2792 Node* n = cl->fast_out(i);
2793 if (n->Opcode() == Op_Phi) {
2794 assert(iv == NULL, "Too many phis" );
2795 iv = n;
2796 }
2797 }
2798 assert(iv == cl->phi(), "Wrong phi" );
2799 #endif
2800
2801 // main and post loops have explicitly created zero trip guard
2802 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2803 if (needs_guard) {
2804 // Skip guard if values not overlap.
2805 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2806 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2807 int stride_con = cl->stride_con();
2808 if (stride_con > 0) {
2809 needs_guard = (init_t->_hi >= limit_t->_lo);
2810 } else {
2811 needs_guard = (init_t->_lo <= limit_t->_hi);
2812 }
2813 }
2814 if (needs_guard) {
2815 // Check for an obvious zero trip guard.
2816 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->skip_predicates());
2817 if (inctrl->Opcode() == Op_IfTrue || inctrl->Opcode() == Op_IfFalse) {
2818 bool maybe_swapped = (inctrl->Opcode() == Op_IfFalse);
2819 // The test should look like just the backedge of a CountedLoop
2820 Node* iff = inctrl->in(0);
2821 if (iff->is_If()) {
2822 Node* bol = iff->in(1);
2823 if (bol->is_Bool()) {
2824 BoolTest test = bol->as_Bool()->_test;
2825 if (maybe_swapped) {
2826 test._test = test.commute();
2827 test._test = test.negate();
2828 }
2829 if (test._test == cl->loopexit()->test_trip()) {
2830 Node* cmp = bol->in(1);
2831 int init_idx = maybe_swapped ? 2 : 1;
2832 int limit_idx = maybe_swapped ? 1 : 2;
2833 if (cmp->is_Cmp() && cmp->in(init_idx) == cl->init_trip() && cmp->in(limit_idx) == cl->limit()) {
2834 needs_guard = false;
2835 }
2836 }
2837 }
2838 }
2839 }
2840 }
2841
2842 #ifndef PRODUCT
2843 if (PrintOpto) {
2844 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2845 this->dump_head();
2846 } else if (TraceLoopOpts) {
2847 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2848 this->dump_head();
2849 }
2850 #endif
2851
2852 if (needs_guard) {
2853 // Peel the loop to ensure there's a zero trip guard
2854 Node_List old_new;
2855 phase->do_peeling(this, old_new);
2856 }
2857
2858 // Replace the phi at loop head with the final value of the last
2859 // iteration. Then the CountedLoopEnd will collapse (backedge never
2860 // taken) and all loop-invariant uses of the exit values will be correct.
2861 Node *phi = cl->phi();
2862 Node *exact_limit = phase->exact_limit(this);
2863 if (exact_limit != cl->limit()) {
2864 // We also need to replace the original limit to collapse loop exit.
2865 Node* cmp = cl->loopexit()->cmp_node();
2866 assert(cl->limit() == cmp->in(2), "sanity");
2867 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2868 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2869 }
2870 // Note: the final value after increment should not overflow since
2871 // counted loop has limit check predicate.
2872 Node *final = new SubINode( exact_limit, cl->stride() );
2873 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2874 phase->_igvn.replace_node(phi,final);
2875 phase->C->set_major_progress();
2876 return true;
2877 }
2878
2879 //------------------------------policy_do_one_iteration_loop-------------------
2880 // Convert one iteration loop into normal code.
2881 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2882 if (!_head->as_Loop()->is_valid_counted_loop())
2883 return false; // Only for counted loop
2884
2885 CountedLoopNode *cl = _head->as_CountedLoop();
2886 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2887 return false;
2888 }
2889
2890 #ifndef PRODUCT
2891 if(TraceLoopOpts) {
2892 tty->print("OneIteration ");
2893 this->dump_head();
2894 }
2895 #endif
2896
2897 Node *init_n = cl->init_trip();
2898 #ifdef ASSERT
2899 // Loop boundaries should be constant since trip count is exact.
2900 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2901 #endif
2902 // Replace the phi at loop head with the value of the init_trip.
2903 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2904 // and all loop-invariant uses of the exit values will be correct.
2905 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2906 phase->C->set_major_progress();
2907 return true;
2908 }
2909
2910 //=============================================================================
2911 //------------------------------iteration_split_impl---------------------------
2912 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2913 // Compute loop trip count if possible.
2914 compute_trip_count(phase);
2915
2916 // Convert one iteration loop into normal code.
2917 if (policy_do_one_iteration_loop(phase))
2918 return true;
2919
2920 // Check and remove empty loops (spam micro-benchmarks)
2921 if (policy_do_remove_empty_loop(phase))
2922 return true; // Here we removed an empty loop
2923
2924 bool should_peel = policy_peeling(phase); // Should we peel?
2925
2926 bool should_unswitch = policy_unswitching(phase);
2927
2928 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2929 // This removes loop-invariant tests (usually null checks).
2930 if (!_head->is_CountedLoop()) { // Non-counted loop
2931 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2932 // Partial peel succeeded so terminate this round of loop opts
2933 return false;
2934 }
2935 if (should_peel) { // Should we peel?
2936 if (PrintOpto) { tty->print_cr("should_peel"); }
2937 phase->do_peeling(this,old_new);
2938 } else if (should_unswitch) {
2939 phase->do_unswitching(this, old_new);
2940 }
2941 return true;
2942 }
2943 CountedLoopNode *cl = _head->as_CountedLoop();
2944
2945 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2946
2947 // Do nothing special to pre- and post- loops
2948 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2949
2950 // Compute loop trip count from profile data
2951 compute_profile_trip_cnt(phase);
2952
2953 // Before attempting fancy unrolling, RCE or alignment, see if we want
2954 // to completely unroll this loop or do loop unswitching.
2955 if (cl->is_normal_loop()) {
2956 if (should_unswitch) {
2957 phase->do_unswitching(this, old_new);
2958 return true;
2959 }
2960 bool should_maximally_unroll = policy_maximally_unroll(phase);
2961 if (should_maximally_unroll) {
2962 // Here we did some unrolling and peeling. Eventually we will
2963 // completely unroll this loop and it will no longer be a loop.
2964 phase->do_maximally_unroll(this,old_new);
2965 return true;
2966 }
2967 }
2968
2969 // Skip next optimizations if running low on nodes. Note that
2970 // policy_unswitching and policy_maximally_unroll have this check.
2971 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes();
2972 if ((int)(2 * _body.size()) > nodes_left) {
2973 return true;
2974 }
2975
2976 // Counted loops may be peeled, may need some iterations run up
2977 // front for RCE, and may want to align loop refs to a cache
2978 // line. Thus we clone a full loop up front whose trip count is
2979 // at least 1 (if peeling), but may be several more.
2980
2981 // The main loop will start cache-line aligned with at least 1
2982 // iteration of the unrolled body (zero-trip test required) and
2983 // will have some range checks removed.
2984
2985 // A post-loop will finish any odd iterations (leftover after
2986 // unrolling), plus any needed for RCE purposes.
2987
2988 bool should_unroll = policy_unroll(phase);
2989
2990 bool should_rce = policy_range_check(phase);
2991
2992 bool should_align = policy_align(phase);
2993
2994 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2995 // need a pre-loop. We may still need to peel an initial iteration but
2996 // we will not be needing an unknown number of pre-iterations.
2997 //
2998 // Basically, if may_rce_align reports FALSE first time through,
2999 // we will not be able to later do RCE or Aligning on this loop.
3000 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
3001
3002 // If we have any of these conditions (RCE, alignment, unrolling) met, then
3003 // we switch to the pre-/main-/post-loop model. This model also covers
3004 // peeling.
3005 if (should_rce || should_align || should_unroll) {
3006 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
3007 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
3008
3009 // Adjust the pre- and main-loop limits to let the pre and post loops run
3010 // with full checks, but the main-loop with no checks. Remove said
3011 // checks from the main body.
3012 if (should_rce) {
3013 if (phase->do_range_check(this, old_new) != 0) {
3014 cl->mark_has_range_checks();
3015 }
3016 } else if (PostLoopMultiversioning) {
3017 phase->has_range_checks(this);
3018 }
3019
3020 if (should_unroll && !should_peel && PostLoopMultiversioning) {
3021 // Try to setup multiversioning on main loops before they are unrolled
3022 if (cl->is_main_loop() && (cl->unrolled_count() == 1)) {
3023 phase->insert_scalar_rced_post_loop(this, old_new);
3024 }
3025 }
3026
3027 // Double loop body for unrolling. Adjust the minimum-trip test (will do
3028 // twice as many iterations as before) and the main body limit (only do
3029 // an even number of trips). If we are peeling, we might enable some RCE
3030 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
3031 // peeling.
3032 if (should_unroll && !should_peel) {
3033 if (SuperWordLoopUnrollAnalysis) {
3034 phase->insert_vector_post_loop(this, old_new);
3035 }
3036 phase->do_unroll(this, old_new, true);
3037 }
3038
3039 // Adjust the pre-loop limits to align the main body
3040 // iterations.
3041 if (should_align)
3042 Unimplemented();
3043
3044 } else { // Else we have an unchanged counted loop
3045 if (should_peel) // Might want to peel but do nothing else
3046 phase->do_peeling(this,old_new);
3047 }
3048 return true;
3049 }
3050
3051
3052 //=============================================================================
3053 //------------------------------iteration_split--------------------------------
3054 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
3055 // Recursively iteration split nested loops
3056 if (_child && !_child->iteration_split(phase, old_new))
3057 return false;
3058
3059 // Clean out prior deadwood
3060 DCE_loop_body();
3061
3062
3063 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
3064 // Replace with a 1-in-10 exit guess.
3065 if (_parent /*not the root loop*/ &&
3066 !_irreducible &&
3067 // Also ignore the occasional dead backedge
3068 !tail()->is_top()) {
3069 adjust_loop_exit_prob(phase);
3070 }
3071
3072 // Gate unrolling, RCE and peeling efforts.
3073 if (!_child && // If not an inner loop, do not split
3074 !_irreducible &&
3075 _allow_optimizations &&
3076 !tail()->is_top()) { // Also ignore the occasional dead backedge
3077 if (!_has_call) {
3078 if (!iteration_split_impl(phase, old_new)) {
3079 return false;
3080 }
3081 } else if (policy_unswitching(phase)) {
3082 phase->do_unswitching(this, old_new);
3083 }
3084 }
3085
3086 // Minor offset re-organization to remove loop-fallout uses of
3087 // trip counter when there was no major reshaping.
3088 phase->reorg_offsets(this);
3089
3090 if (_next && !_next->iteration_split(phase, old_new))
3091 return false;
3092 return true;
3093 }
3094
3095
3096 //=============================================================================
3097 // Process all the loops in the loop tree and replace any fill
3098 // patterns with an intrinsic version.
3099 bool PhaseIdealLoop::do_intrinsify_fill() {
3100 bool changed = false;
3101 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3102 IdealLoopTree* lpt = iter.current();
3103 changed |= intrinsify_fill(lpt);
3104 }
3105 return changed;
3106 }
3107
3108
3109 // Examine an inner loop looking for a a single store of an invariant
3110 // value in a unit stride loop,
3111 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
3112 Node*& shift, Node*& con) {
3113 const char* msg = NULL;
3114 Node* msg_node = NULL;
3115
3116 store_value = NULL;
3117 con = NULL;
3118 shift = NULL;
3119
3120 // Process the loop looking for stores. If there are multiple
3121 // stores or extra control flow give at this point.
3122 CountedLoopNode* head = lpt->_head->as_CountedLoop();
3123 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3124 Node* n = lpt->_body.at(i);
3125 if (n->outcnt() == 0) continue; // Ignore dead
3126 if (n->is_Store()) {
3127 if (store != NULL) {
3128 msg = "multiple stores";
3129 break;
3130 }
3131 int opc = n->Opcode();
3132 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
3133 msg = "oop fills not handled";
3134 break;
3135 }
3136 Node* value = n->in(MemNode::ValueIn);
3137 if (!lpt->is_invariant(value)) {
3138 msg = "variant store value";
3139 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
3140 msg = "not array address";
3141 }
3142 store = n;
3143 store_value = value;
3144 } else if (n->is_If() && n != head->loopexit_or_null()) {
3145 msg = "extra control flow";
3146 msg_node = n;
3147 }
3148 }
3149
3150 if (store == NULL) {
3151 // No store in loop
3152 return false;
3153 }
3154
3155 if (msg == NULL && head->stride_con() != 1) {
3156 // could handle negative strides too
3157 if (head->stride_con() < 0) {
3158 msg = "negative stride";
3159 } else {
3160 msg = "non-unit stride";
3161 }
3162 }
3163
3164 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
3165 msg = "can't handle store address";
3166 msg_node = store->in(MemNode::Address);
3167 }
3168
3169 if (msg == NULL &&
3170 (!store->in(MemNode::Memory)->is_Phi() ||
3171 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
3172 msg = "store memory isn't proper phi";
3173 msg_node = store->in(MemNode::Memory);
3174 }
3175
3176 // Make sure there is an appropriate fill routine
3177 BasicType t = store->as_Mem()->memory_type();
3178 const char* fill_name;
3179 if (msg == NULL &&
3180 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
3181 msg = "unsupported store";
3182 msg_node = store;
3183 }
3184
3185 if (msg != NULL) {
3186 #ifndef PRODUCT
3187 if (TraceOptimizeFill) {
3188 tty->print_cr("not fill intrinsic candidate: %s", msg);
3189 if (msg_node != NULL) msg_node->dump();
3190 }
3191 #endif
3192 return false;
3193 }
3194
3195 // Make sure the address expression can be handled. It should be
3196 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()).
3197 Node* elements[4];
3198 Node* cast = NULL;
3199 Node* conv = NULL;
3200 bool found_index = false;
3201 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
3202 for (int e = 0; e < count; e++) {
3203 Node* n = elements[e];
3204 if (n->is_Con() && con == NULL) {
3205 con = n;
3206 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
3207 Node* value = n->in(1);
3208 #ifdef _LP64
3209 if (value->Opcode() == Op_ConvI2L) {
3210 conv = value;
3211 value = value->in(1);
3212 }
3213 if (value->Opcode() == Op_CastII &&
3214 value->as_CastII()->has_range_check()) {
3215 // Skip range check dependent CastII nodes
3216 cast = value;
3217 value = value->in(1);
3218 }
3219 #endif
3220 if (value != head->phi()) {
3221 msg = "unhandled shift in address";
3222 } else {
3223 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
3224 msg = "scale doesn't match";
3225 } else {
3226 found_index = true;
3227 shift = n;
3228 }
3229 }
3230 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
3231 conv = n;
3232 n = n->in(1);
3233 if (n->Opcode() == Op_CastII &&
3234 n->as_CastII()->has_range_check()) {
3235 // Skip range check dependent CastII nodes
3236 cast = n;
3237 n = n->in(1);
3238 }
3239 if (n == head->phi()) {
3240 found_index = true;
3241 } else {
3242 msg = "unhandled input to ConvI2L";
3243 }
3244 } else if (n == head->phi()) {
3245 // no shift, check below for allowed cases
3246 found_index = true;
3247 } else {
3248 msg = "unhandled node in address";
3249 msg_node = n;
3250 }
3251 }
3252
3253 if (count == -1) {
3254 msg = "malformed address expression";
3255 msg_node = store;
3256 }
3257
3258 if (!found_index) {
3259 msg = "missing use of index";
3260 }
3261
3262 // byte sized items won't have a shift
3263 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
3264 msg = "can't find shift";
3265 msg_node = store;
3266 }
3267
3268 if (msg != NULL) {
3269 #ifndef PRODUCT
3270 if (TraceOptimizeFill) {
3271 tty->print_cr("not fill intrinsic: %s", msg);
3272 if (msg_node != NULL) msg_node->dump();
3273 }
3274 #endif
3275 return false;
3276 }
3277
3278 // No make sure all the other nodes in the loop can be handled
3279 VectorSet ok(Thread::current()->resource_area());
3280
3281 // store related values are ok
3282 ok.set(store->_idx);
3283 ok.set(store->in(MemNode::Memory)->_idx);
3284
3285 CountedLoopEndNode* loop_exit = head->loopexit();
3286
3287 // Loop structure is ok
3288 ok.set(head->_idx);
3289 ok.set(loop_exit->_idx);
3290 ok.set(head->phi()->_idx);
3291 ok.set(head->incr()->_idx);
3292 ok.set(loop_exit->cmp_node()->_idx);
3293 ok.set(loop_exit->in(1)->_idx);
3294
3295 // Address elements are ok
3296 if (con) ok.set(con->_idx);
3297 if (shift) ok.set(shift->_idx);
3298 if (cast) ok.set(cast->_idx);
3299 if (conv) ok.set(conv->_idx);
3300
3301 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3302 Node* n = lpt->_body.at(i);
3303 if (n->outcnt() == 0) continue; // Ignore dead
3304 if (ok.test(n->_idx)) continue;
3305 // Backedge projection is ok
3306 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
3307 if (!n->is_AddP()) {
3308 msg = "unhandled node";
3309 msg_node = n;
3310 break;
3311 }
3312 }
3313
3314 // Make sure no unexpected values are used outside the loop
3315 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3316 Node* n = lpt->_body.at(i);
3317 // These values can be replaced with other nodes if they are used
3318 // outside the loop.
3319 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
3320 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
3321 Node* use = iter.get();
3322 if (!lpt->_body.contains(use)) {
3323 msg = "node is used outside loop";
3324 // lpt->_body.dump();
3325 msg_node = n;
3326 break;
3327 }
3328 }
3329 }
3330
3331 #ifdef ASSERT
3332 if (TraceOptimizeFill) {
3333 if (msg != NULL) {
3334 tty->print_cr("no fill intrinsic: %s", msg);
3335 if (msg_node != NULL) msg_node->dump();
3336 } else {
3337 tty->print_cr("fill intrinsic for:");
3338 }
3339 store->dump();
3340 if (Verbose) {
3341 lpt->_body.dump();
3342 }
3343 }
3344 #endif
3345
3346 return msg == NULL;
3347 }
3348
3349
3350
3351 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
3352 // Only for counted inner loops
3353 if (!lpt->is_counted() || !lpt->is_inner()) {
3354 return false;
3355 }
3356
3357 // Must have constant stride
3358 CountedLoopNode* head = lpt->_head->as_CountedLoop();
3359 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
3360 return false;
3361 }
3362
3363 head->verify_strip_mined(1);
3364
3365 // Check that the body only contains a store of a loop invariant
3366 // value that is indexed by the loop phi.
3367 Node* store = NULL;
3368 Node* store_value = NULL;
3369 Node* shift = NULL;
3370 Node* offset = NULL;
3371 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
3372 return false;
3373 }
3374
3375 Node* exit = head->loopexit()->proj_out_or_null(0);
3376 if (exit == NULL) {
3377 return false;
3378 }
3379
3380 #ifndef PRODUCT
3381 if (TraceLoopOpts) {
3382 tty->print("ArrayFill ");
3383 lpt->dump_head();
3384 }
3385 #endif
3386
3387 // Now replace the whole loop body by a call to a fill routine that
3388 // covers the same region as the loop.
3389 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
3390
3391 // Build an expression for the beginning of the copy region
3392 Node* index = head->init_trip();
3393 #ifdef _LP64
3394 index = new ConvI2LNode(index);
3395 _igvn.register_new_node_with_optimizer(index);
3396 #endif
3397 if (shift != NULL) {
3398 // byte arrays don't require a shift but others do.
3399 index = new LShiftXNode(index, shift->in(2));
3400 _igvn.register_new_node_with_optimizer(index);
3401 }
3402 index = new AddPNode(base, base, index);
3403 _igvn.register_new_node_with_optimizer(index);
3404 Node* from = new AddPNode(base, index, offset);
3405 _igvn.register_new_node_with_optimizer(from);
3406 // Compute the number of elements to copy
3407 Node* len = new SubINode(head->limit(), head->init_trip());
3408 _igvn.register_new_node_with_optimizer(len);
3409
3410 BasicType t = store->as_Mem()->memory_type();
3411 bool aligned = false;
3412 if (offset != NULL && head->init_trip()->is_Con()) {
3413 int element_size = type2aelembytes(t);
3414 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
3415 }
3416
3417 // Build a call to the fill routine
3418 const char* fill_name;
3419 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
3420 assert(fill != NULL, "what?");
3421
3422 // Convert float/double to int/long for fill routines
3423 if (t == T_FLOAT) {
3424 store_value = new MoveF2INode(store_value);
3425 _igvn.register_new_node_with_optimizer(store_value);
3426 } else if (t == T_DOUBLE) {
3427 store_value = new MoveD2LNode(store_value);
3428 _igvn.register_new_node_with_optimizer(store_value);
3429 }
3430
3431 Node* mem_phi = store->in(MemNode::Memory);
3432 Node* result_ctrl;
3433 Node* result_mem;
3434 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
3435 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill,
3436 fill_name, TypeAryPtr::get_array_body_type(t));
3437 uint cnt = 0;
3438 call->init_req(TypeFunc::Parms + cnt++, from);
3439 call->init_req(TypeFunc::Parms + cnt++, store_value);
3440 #ifdef _LP64
3441 len = new ConvI2LNode(len);
3442 _igvn.register_new_node_with_optimizer(len);
3443 #endif
3444 call->init_req(TypeFunc::Parms + cnt++, len);
3445 #ifdef _LP64
3446 call->init_req(TypeFunc::Parms + cnt++, C->top());
3447 #endif
3448 call->init_req(TypeFunc::Control, head->init_control());
3449 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O.
3450 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl));
3451 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out_or_null(TypeFunc::ReturnAdr));
3452 call->init_req(TypeFunc::FramePtr, C->start()->proj_out_or_null(TypeFunc::FramePtr));
3453 _igvn.register_new_node_with_optimizer(call);
3454 result_ctrl = new ProjNode(call,TypeFunc::Control);
3455 _igvn.register_new_node_with_optimizer(result_ctrl);
3456 result_mem = new ProjNode(call,TypeFunc::Memory);
3457 _igvn.register_new_node_with_optimizer(result_mem);
3458
3459 /* Disable following optimization until proper fix (add missing checks).
3460
3461 // If this fill is tightly coupled to an allocation and overwrites
3462 // the whole body, allow it to take over the zeroing.
3463 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
3464 if (alloc != NULL && alloc->is_AllocateArray()) {
3465 Node* length = alloc->as_AllocateArray()->Ideal_length();
3466 if (head->limit() == length &&
3467 head->init_trip() == _igvn.intcon(0)) {
3468 if (TraceOptimizeFill) {
3469 tty->print_cr("Eliminated zeroing in allocation");
3470 }
3471 alloc->maybe_set_complete(&_igvn);
3472 } else {
3473 #ifdef ASSERT
3474 if (TraceOptimizeFill) {
3475 tty->print_cr("filling array but bounds don't match");
3476 alloc->dump();
3477 head->init_trip()->dump();
3478 head->limit()->dump();
3479 length->dump();
3480 }
3481 #endif
3482 }
3483 }
3484 */
3485
3486 if (head->is_strip_mined()) {
3487 // Inner strip mined loop goes away so get rid of outer strip
3488 // mined loop
3489 Node* outer_sfpt = head->outer_safepoint();
3490 Node* in = outer_sfpt->in(0);
3491 Node* outer_out = head->outer_loop_exit();
3492 lazy_replace(outer_out, in);
3493 _igvn.replace_input_of(outer_sfpt, 0, C->top());
3494 }
3495
3496 // Redirect the old control and memory edges that are outside the loop.
3497 // Sometimes the memory phi of the head is used as the outgoing
3498 // state of the loop. It's safe in this case to replace it with the
3499 // result_mem.
3500 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
3501 lazy_replace(exit, result_ctrl);
3502 _igvn.replace_node(store, result_mem);
3503 // Any uses the increment outside of the loop become the loop limit.
3504 _igvn.replace_node(head->incr(), head->limit());
3505
3506 // Disconnect the head from the loop.
3507 for (uint i = 0; i < lpt->_body.size(); i++) {
3508 Node* n = lpt->_body.at(i);
3509 _igvn.replace_node(n, C->top());
3510 }
3511
3512 return true;
3513 }