Mon, 04 Apr 2011 19:02:36 -0700
7004547: regular loop unroll should not unroll more than max unrolling
Summary: Take into account that after unroll conjoined heads and tails will fold.
Reviewed-by: never
1 /*
2 * Copyright (c) 2000, 2010, 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 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/connode.hpp"
31 #include "opto/divnode.hpp"
32 #include "opto/loopnode.hpp"
33 #include "opto/mulnode.hpp"
34 #include "opto/rootnode.hpp"
35 #include "opto/runtime.hpp"
36 #include "opto/subnode.hpp"
38 //------------------------------is_loop_exit-----------------------------------
39 // Given an IfNode, return the loop-exiting projection or NULL if both
40 // arms remain in the loop.
41 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
42 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
43 PhaseIdealLoop *phase = _phase;
44 // Test is an IfNode, has 2 projections. If BOTH are in the loop
45 // we need loop unswitching instead of peeling.
46 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
47 return iff->raw_out(0);
48 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
49 return iff->raw_out(1);
50 return NULL;
51 }
54 //=============================================================================
57 //------------------------------record_for_igvn----------------------------
58 // Put loop body on igvn work list
59 void IdealLoopTree::record_for_igvn() {
60 for( uint i = 0; i < _body.size(); i++ ) {
61 Node *n = _body.at(i);
62 _phase->_igvn._worklist.push(n);
63 }
64 }
66 //------------------------------compute_profile_trip_cnt----------------------------
67 // Compute loop trip count from profile data as
68 // (backedge_count + loop_exit_count) / loop_exit_count
69 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
70 if (!_head->is_CountedLoop()) {
71 return;
72 }
73 CountedLoopNode* head = _head->as_CountedLoop();
74 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
75 return; // Already computed
76 }
77 float trip_cnt = (float)max_jint; // default is big
79 Node* back = head->in(LoopNode::LoopBackControl);
80 while (back != head) {
81 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
82 back->in(0) &&
83 back->in(0)->is_If() &&
84 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
85 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
86 break;
87 }
88 back = phase->idom(back);
89 }
90 if (back != head) {
91 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
92 back->in(0), "if-projection exists");
93 IfNode* back_if = back->in(0)->as_If();
94 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
96 // Now compute a loop exit count
97 float loop_exit_cnt = 0.0f;
98 for( uint i = 0; i < _body.size(); i++ ) {
99 Node *n = _body[i];
100 if( n->is_If() ) {
101 IfNode *iff = n->as_If();
102 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
103 Node *exit = is_loop_exit(iff);
104 if( exit ) {
105 float exit_prob = iff->_prob;
106 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
107 if (exit_prob > PROB_MIN) {
108 float exit_cnt = iff->_fcnt * exit_prob;
109 loop_exit_cnt += exit_cnt;
110 }
111 }
112 }
113 }
114 }
115 if (loop_exit_cnt > 0.0f) {
116 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
117 } else {
118 // No exit count so use
119 trip_cnt = loop_back_cnt;
120 }
121 }
122 #ifndef PRODUCT
123 if (TraceProfileTripCount) {
124 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
125 }
126 #endif
127 head->set_profile_trip_cnt(trip_cnt);
128 }
130 //---------------------is_invariant_addition-----------------------------
131 // Return nonzero index of invariant operand for an Add or Sub
132 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
133 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
134 int op = n->Opcode();
135 if (op == Op_AddI || op == Op_SubI) {
136 bool in1_invar = this->is_invariant(n->in(1));
137 bool in2_invar = this->is_invariant(n->in(2));
138 if (in1_invar && !in2_invar) return 1;
139 if (!in1_invar && in2_invar) return 2;
140 }
141 return 0;
142 }
144 //---------------------reassociate_add_sub-----------------------------
145 // Reassociate invariant add and subtract expressions:
146 //
147 // inv1 + (x + inv2) => ( inv1 + inv2) + x
148 // (x + inv2) + inv1 => ( inv1 + inv2) + x
149 // inv1 + (x - inv2) => ( inv1 - inv2) + x
150 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
151 // (x + inv2) - inv1 => (-inv1 + inv2) + x
152 // (x - inv2) + inv1 => ( inv1 - inv2) + x
153 // (x - inv2) - inv1 => (-inv1 - inv2) + x
154 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
155 // inv1 - (x - inv2) => ( inv1 + inv2) - x
156 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
157 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
158 // inv1 - (x + inv2) => ( inv1 - inv2) - x
159 //
160 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
161 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
162 if (is_invariant(n1)) return NULL;
163 int inv1_idx = is_invariant_addition(n1, phase);
164 if (!inv1_idx) return NULL;
165 // Don't mess with add of constant (igvn moves them to expression tree root.)
166 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
167 Node* inv1 = n1->in(inv1_idx);
168 Node* n2 = n1->in(3 - inv1_idx);
169 int inv2_idx = is_invariant_addition(n2, phase);
170 if (!inv2_idx) return NULL;
171 Node* x = n2->in(3 - inv2_idx);
172 Node* inv2 = n2->in(inv2_idx);
174 bool neg_x = n2->is_Sub() && inv2_idx == 1;
175 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
176 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
177 if (n1->is_Sub() && inv1_idx == 1) {
178 neg_x = !neg_x;
179 neg_inv2 = !neg_inv2;
180 }
181 Node* inv1_c = phase->get_ctrl(inv1);
182 Node* inv2_c = phase->get_ctrl(inv2);
183 Node* n_inv1;
184 if (neg_inv1) {
185 Node *zero = phase->_igvn.intcon(0);
186 phase->set_ctrl(zero, phase->C->root());
187 n_inv1 = new (phase->C, 3) SubINode(zero, inv1);
188 phase->register_new_node(n_inv1, inv1_c);
189 } else {
190 n_inv1 = inv1;
191 }
192 Node* inv;
193 if (neg_inv2) {
194 inv = new (phase->C, 3) SubINode(n_inv1, inv2);
195 } else {
196 inv = new (phase->C, 3) AddINode(n_inv1, inv2);
197 }
198 phase->register_new_node(inv, phase->get_early_ctrl(inv));
200 Node* addx;
201 if (neg_x) {
202 addx = new (phase->C, 3) SubINode(inv, x);
203 } else {
204 addx = new (phase->C, 3) AddINode(x, inv);
205 }
206 phase->register_new_node(addx, phase->get_ctrl(x));
207 phase->_igvn.replace_node(n1, addx);
208 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
209 _body.yank(n1);
210 return addx;
211 }
213 //---------------------reassociate_invariants-----------------------------
214 // Reassociate invariant expressions:
215 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
216 for (int i = _body.size() - 1; i >= 0; i--) {
217 Node *n = _body.at(i);
218 for (int j = 0; j < 5; j++) {
219 Node* nn = reassociate_add_sub(n, phase);
220 if (nn == NULL) break;
221 n = nn; // again
222 };
223 }
224 }
226 //------------------------------policy_peeling---------------------------------
227 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
228 // make some loop-invariant test (usually a null-check) happen before the loop.
229 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
230 Node *test = ((IdealLoopTree*)this)->tail();
231 int body_size = ((IdealLoopTree*)this)->_body.size();
232 int uniq = phase->C->unique();
233 // Peeling does loop cloning which can result in O(N^2) node construction
234 if( body_size > 255 /* Prevent overflow for large body_size */
235 || (body_size * body_size + uniq > MaxNodeLimit) ) {
236 return false; // too large to safely clone
237 }
238 while( test != _head ) { // Scan till run off top of loop
239 if( test->is_If() ) { // Test?
240 Node *ctrl = phase->get_ctrl(test->in(1));
241 if (ctrl->is_top())
242 return false; // Found dead test on live IF? No peeling!
243 // Standard IF only has one input value to check for loop invariance
244 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
245 // Condition is not a member of this loop?
246 if( !is_member(phase->get_loop(ctrl)) &&
247 is_loop_exit(test) )
248 return true; // Found reason to peel!
249 }
250 // Walk up dominators to loop _head looking for test which is
251 // executed on every path thru loop.
252 test = phase->idom(test);
253 }
254 return false;
255 }
257 //------------------------------peeled_dom_test_elim---------------------------
258 // If we got the effect of peeling, either by actually peeling or by making
259 // a pre-loop which must execute at least once, we can remove all
260 // loop-invariant dominated tests in the main body.
261 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
262 bool progress = true;
263 while( progress ) {
264 progress = false; // Reset for next iteration
265 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
266 Node *test = prev->in(0);
267 while( test != loop->_head ) { // Scan till run off top of loop
269 int p_op = prev->Opcode();
270 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
271 test->is_If() && // Test?
272 !test->in(1)->is_Con() && // And not already obvious?
273 // Condition is not a member of this loop?
274 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
275 // Walk loop body looking for instances of this test
276 for( uint i = 0; i < loop->_body.size(); i++ ) {
277 Node *n = loop->_body.at(i);
278 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
279 // IfNode was dominated by version in peeled loop body
280 progress = true;
281 dominated_by( old_new[prev->_idx], n );
282 }
283 }
284 }
285 prev = test;
286 test = idom(test);
287 } // End of scan tests in loop
289 } // End of while( progress )
290 }
292 //------------------------------do_peeling-------------------------------------
293 // Peel the first iteration of the given loop.
294 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
295 // The pre-loop illegally has 2 control users (old & new loops).
296 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
297 // Do this by making the old-loop fall-in edges act as if they came
298 // around the loopback from the prior iteration (follow the old-loop
299 // backedges) and then map to the new peeled iteration. This leaves
300 // the pre-loop with only 1 user (the new peeled iteration), but the
301 // peeled-loop backedge has 2 users.
302 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
303 // extra backedge user.
304 //
305 // orig
306 //
307 // stmt1
308 // |
309 // v
310 // loop predicate
311 // |
312 // v
313 // loop<----+
314 // | |
315 // stmt2 |
316 // | |
317 // v |
318 // if ^
319 // / \ |
320 // / \ |
321 // v v |
322 // false true |
323 // / \ |
324 // / ----+
325 // |
326 // v
327 // exit
328 //
329 //
330 // after clone loop
331 //
332 // stmt1
333 // |
334 // v
335 // loop predicate
336 // / \
337 // clone / \ orig
338 // / \
339 // / \
340 // v v
341 // +---->loop clone loop<----+
342 // | | | |
343 // | stmt2 clone stmt2 |
344 // | | | |
345 // | v v |
346 // ^ if clone If ^
347 // | / \ / \ |
348 // | / \ / \ |
349 // | v v v v |
350 // | true false false true |
351 // | / \ / \ |
352 // +---- \ / ----+
353 // \ /
354 // 1v v2
355 // region
356 // |
357 // v
358 // exit
359 //
360 //
361 // after peel and predicate move
362 //
363 // stmt1
364 // /
365 // /
366 // clone / orig
367 // /
368 // / +----------+
369 // / | |
370 // / loop predicate |
371 // / | |
372 // v v |
373 // TOP-->loop clone loop<----+ |
374 // | | | |
375 // stmt2 clone stmt2 | |
376 // | | | ^
377 // v v | |
378 // if clone If ^ |
379 // / \ / \ | |
380 // / \ / \ | |
381 // v v v v | |
382 // true false false true | |
383 // | \ / \ | |
384 // | \ / ----+ ^
385 // | \ / |
386 // | 1v v2 |
387 // v region |
388 // | | |
389 // | v |
390 // | exit |
391 // | |
392 // +--------------->-----------------+
393 //
394 //
395 // final graph
396 //
397 // stmt1
398 // |
399 // v
400 // stmt2 clone
401 // |
402 // v
403 // if clone
404 // / |
405 // / |
406 // v v
407 // false true
408 // | |
409 // | v
410 // | loop predicate
411 // | |
412 // | v
413 // | loop<----+
414 // | | |
415 // | stmt2 |
416 // | | |
417 // | v |
418 // v if ^
419 // | / \ |
420 // | / \ |
421 // | v v |
422 // | false true |
423 // | | \ |
424 // v v --+
425 // region
426 // |
427 // v
428 // exit
429 //
430 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
432 C->set_major_progress();
433 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
434 // 'pre' loop from the main and the 'pre' can no longer have it's
435 // iterations adjusted. Therefore, we need to declare this loop as
436 // no longer a 'main' loop; it will need new pre and post loops before
437 // we can do further RCE.
438 #ifndef PRODUCT
439 if (TraceLoopOpts) {
440 tty->print("Peel ");
441 loop->dump_head();
442 }
443 #endif
444 Node* head = loop->_head;
445 bool counted_loop = head->is_CountedLoop();
446 if (counted_loop) {
447 CountedLoopNode *cl = head->as_CountedLoop();
448 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
449 cl->set_trip_count(cl->trip_count() - 1);
450 if (cl->is_main_loop()) {
451 cl->set_normal_loop();
452 #ifndef PRODUCT
453 if (PrintOpto && VerifyLoopOptimizations) {
454 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
455 loop->dump_head();
456 }
457 #endif
458 }
459 }
460 Node* entry = head->in(LoopNode::EntryControl);
462 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
463 // The pre-loop illegally has 2 control users (old & new loops).
464 clone_loop( loop, old_new, dom_depth(head) );
466 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
467 // Do this by making the old-loop fall-in edges act as if they came
468 // around the loopback from the prior iteration (follow the old-loop
469 // backedges) and then map to the new peeled iteration. This leaves
470 // the pre-loop with only 1 user (the new peeled iteration), but the
471 // peeled-loop backedge has 2 users.
472 Node* new_exit_value = old_new[head->in(LoopNode::LoopBackControl)->_idx];
473 new_exit_value = move_loop_predicates(entry, new_exit_value);
474 _igvn.hash_delete(head);
475 head->set_req(LoopNode::EntryControl, new_exit_value);
476 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
477 Node* old = head->fast_out(j);
478 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
479 new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
480 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
481 // Then loop body backedge value remains the same.
482 new_exit_value = old->in(LoopNode::LoopBackControl);
483 _igvn.hash_delete(old);
484 old->set_req(LoopNode::EntryControl, new_exit_value);
485 }
486 }
489 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
490 // extra backedge user.
491 Node* new_head = old_new[head->_idx];
492 _igvn.hash_delete(new_head);
493 new_head->set_req(LoopNode::LoopBackControl, C->top());
494 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
495 Node* use = new_head->fast_out(j2);
496 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
497 _igvn.hash_delete(use);
498 use->set_req(LoopNode::LoopBackControl, C->top());
499 }
500 }
503 // Step 4: Correct dom-depth info. Set to loop-head depth.
504 int dd = dom_depth(head);
505 set_idom(head, head->in(1), dd);
506 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
507 Node *old = loop->_body.at(j3);
508 Node *nnn = old_new[old->_idx];
509 if (!has_ctrl(nnn))
510 set_idom(nnn, idom(nnn), dd-1);
511 // While we're at it, remove any SafePoints from the peeled code
512 if (old->Opcode() == Op_SafePoint) {
513 Node *nnn = old_new[old->_idx];
514 lazy_replace(nnn,nnn->in(TypeFunc::Control));
515 }
516 }
518 // Now force out all loop-invariant dominating tests. The optimizer
519 // finds some, but we _know_ they are all useless.
520 peeled_dom_test_elim(loop,old_new);
522 loop->record_for_igvn();
523 }
525 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
527 //------------------------------policy_maximally_unroll------------------------
528 // Calculate exact loop trip count and return true if loop can be maximally
529 // unrolled.
530 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
531 CountedLoopNode *cl = _head->as_CountedLoop();
532 assert(cl->is_normal_loop(), "");
533 if (!cl->is_valid_counted_loop())
534 return false; // Malformed counted loop
536 Node *init_n = cl->init_trip();
537 Node *limit_n = cl->limit();
539 // Non-constant bounds
540 if (init_n == NULL || !init_n->is_Con() ||
541 limit_n == NULL || !limit_n->is_Con()) {
542 return false;
543 }
544 // Use longs to avoid integer overflow.
545 int stride_con = cl->stride_con();
546 long init_con = cl->init_trip()->get_int();
547 long limit_con = cl->limit()->get_int();
548 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
549 long trip_cnt = (limit_con - init_con + stride_m)/stride_con;
551 // Note, max_jint is used to indicate unknown trip count.
552 if (trip_cnt <= 0 || trip_cnt >= (long)max_jint) {
553 // return a false (no maximally unroll) and the regular unroll/peel
554 // route will make a small mess which CCP will fold away.
555 return false;
556 }
557 uint trip_count = (uint)trip_cnt;
558 cl->set_trip_count(trip_count);
560 // Real policy: if we maximally unroll, does it get too big?
561 // Allow the unrolled mess to get larger than standard loop
562 // size. After all, it will no longer be a loop.
563 uint body_size = _body.size();
564 uint unroll_limit = (uint)LoopUnrollLimit * 4;
565 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
566 if (trip_count > unroll_limit || body_size > unroll_limit) {
567 return false;
568 }
570 // Take into account that after unroll conjoined heads and tails will fold,
571 // otherwise policy_unroll() may allow more unrolling than max unrolling.
572 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
573 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
574 if (body_size != tst_body_size) // Check for int overflow
575 return false;
576 if (new_body_size > unroll_limit ||
577 // Unrolling can result in a large amount of node construction
578 new_body_size >= MaxNodeLimit - phase->C->unique()) {
579 return false;
580 }
582 // Currently we don't have policy to optimize one iteration loops.
583 // Maximally unrolling transformation is used for that:
584 // it is peeled and the original loop become non reachable (dead).
585 // Also fully unroll a loop with few iterations regardless next
586 // conditions since following loop optimizations will split
587 // such loop anyway (pre-main-post).
588 if (trip_count <= 3)
589 return true;
591 // Do not unroll a loop with String intrinsics code.
592 // String intrinsics are large and have loops.
593 for (uint k = 0; k < _body.size(); k++) {
594 Node* n = _body.at(k);
595 switch (n->Opcode()) {
596 case Op_StrComp:
597 case Op_StrEquals:
598 case Op_StrIndexOf:
599 case Op_AryEq: {
600 return false;
601 }
602 } // switch
603 }
605 return true; // Do maximally unroll
606 }
609 //------------------------------policy_unroll----------------------------------
610 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
611 // the loop is a CountedLoop and the body is small enough.
612 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
614 CountedLoopNode *cl = _head->as_CountedLoop();
615 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
617 if (!cl->is_valid_counted_loop())
618 return false; // Malformed counted loop
620 // protect against over-unrolling
621 if (cl->trip_count() <= 1) return false;
623 // Check for stride being a small enough constant
624 if (abs(cl->stride_con()) > (1<<3)) return false;
626 int future_unroll_ct = cl->unrolled_count() * 2;
628 // Don't unroll if the next round of unrolling would push us
629 // over the expected trip count of the loop. One is subtracted
630 // from the expected trip count because the pre-loop normally
631 // executes 1 iteration.
632 if (UnrollLimitForProfileCheck > 0 &&
633 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
634 future_unroll_ct > UnrollLimitForProfileCheck &&
635 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
636 return false;
637 }
639 // When unroll count is greater than LoopUnrollMin, don't unroll if:
640 // the residual iterations are more than 10% of the trip count
641 // and rounds of "unroll,optimize" are not making significant progress
642 // Progress defined as current size less than 20% larger than previous size.
643 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
644 future_unroll_ct > LoopUnrollMin &&
645 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
646 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
647 return false;
648 }
650 Node *init_n = cl->init_trip();
651 Node *limit_n = cl->limit();
652 // Non-constant bounds.
653 // Protect against over-unrolling when init or/and limit are not constant
654 // (so that trip_count's init value is maxint) but iv range is known.
655 if (init_n == NULL || !init_n->is_Con() ||
656 limit_n == NULL || !limit_n->is_Con()) {
657 Node* phi = cl->phi();
658 if (phi != NULL) {
659 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
660 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
661 int next_stride = cl->stride_con() * 2; // stride after this unroll
662 if (next_stride > 0) {
663 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
664 iv_type->_lo + next_stride > iv_type->_hi) {
665 return false; // over-unrolling
666 }
667 } else if (next_stride < 0) {
668 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
669 iv_type->_hi + next_stride < iv_type->_lo) {
670 return false; // over-unrolling
671 }
672 }
673 }
674 }
676 // Adjust body_size to determine if we unroll or not
677 uint body_size = _body.size();
678 // Key test to unroll CaffeineMark's Logic test
679 int xors_in_loop = 0;
680 // Also count ModL, DivL and MulL which expand mightly
681 for (uint k = 0; k < _body.size(); k++) {
682 Node* n = _body.at(k);
683 switch (n->Opcode()) {
684 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test
685 case Op_ModL: body_size += 30; break;
686 case Op_DivL: body_size += 30; break;
687 case Op_MulL: body_size += 10; break;
688 case Op_StrComp:
689 case Op_StrEquals:
690 case Op_StrIndexOf:
691 case Op_AryEq: {
692 // Do not unroll a loop with String intrinsics code.
693 // String intrinsics are large and have loops.
694 return false;
695 }
696 } // switch
697 }
699 // Check for being too big
700 if (body_size > (uint)LoopUnrollLimit) {
701 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
702 // Normal case: loop too big
703 return false;
704 }
706 // Unroll once! (Each trip will soon do double iterations)
707 return true;
708 }
710 //------------------------------policy_align-----------------------------------
711 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
712 // expression that does the alignment. Note that only one array base can be
713 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
714 // if we vectorize short memory ops into longer memory ops, we may want to
715 // increase alignment.
716 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
717 return false;
718 }
720 //------------------------------policy_range_check-----------------------------
721 // Return TRUE or FALSE if the loop should be range-check-eliminated.
722 // Actually we do iteration-splitting, a more powerful form of RCE.
723 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
724 if( !RangeCheckElimination ) return false;
726 CountedLoopNode *cl = _head->as_CountedLoop();
727 // If we unrolled with no intention of doing RCE and we later
728 // changed our minds, we got no pre-loop. Either we need to
729 // make a new pre-loop, or we gotta disallow RCE.
730 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now.
731 Node *trip_counter = cl->phi();
733 // Check loop body for tests of trip-counter plus loop-invariant vs
734 // loop-invariant.
735 for( uint i = 0; i < _body.size(); i++ ) {
736 Node *iff = _body[i];
737 if( iff->Opcode() == Op_If ) { // Test?
739 // Comparing trip+off vs limit
740 Node *bol = iff->in(1);
741 if( bol->req() != 2 ) continue; // dead constant test
742 if (!bol->is_Bool()) {
743 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
744 continue;
745 }
746 Node *cmp = bol->in(1);
748 Node *rc_exp = cmp->in(1);
749 Node *limit = cmp->in(2);
751 Node *limit_c = phase->get_ctrl(limit);
752 if( limit_c == phase->C->top() )
753 return false; // Found dead test on live IF? No RCE!
754 if( is_member(phase->get_loop(limit_c) ) ) {
755 // Compare might have operands swapped; commute them
756 rc_exp = cmp->in(2);
757 limit = cmp->in(1);
758 limit_c = phase->get_ctrl(limit);
759 if( is_member(phase->get_loop(limit_c) ) )
760 continue; // Both inputs are loop varying; cannot RCE
761 }
763 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
764 continue;
765 }
766 // Yeah! Found a test like 'trip+off vs limit'
767 // Test is an IfNode, has 2 projections. If BOTH are in the loop
768 // we need loop unswitching instead of iteration splitting.
769 if( is_loop_exit(iff) )
770 return true; // Found reason to split iterations
771 } // End of is IF
772 }
774 return false;
775 }
777 //------------------------------policy_peel_only-------------------------------
778 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
779 // for unrolling loops with NO array accesses.
780 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
782 for( uint i = 0; i < _body.size(); i++ )
783 if( _body[i]->is_Mem() )
784 return false;
786 // No memory accesses at all!
787 return true;
788 }
790 //------------------------------clone_up_backedge_goo--------------------------
791 // If Node n lives in the back_ctrl block and cannot float, we clone a private
792 // version of n in preheader_ctrl block and return that, otherwise return n.
793 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) {
794 if( get_ctrl(n) != back_ctrl ) return n;
796 Node *x = NULL; // If required, a clone of 'n'
797 // Check for 'n' being pinned in the backedge.
798 if( n->in(0) && n->in(0) == back_ctrl ) {
799 x = n->clone(); // Clone a copy of 'n' to preheader
800 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
801 }
803 // Recursive fixup any other input edges into x.
804 // If there are no changes we can just return 'n', otherwise
805 // we need to clone a private copy and change it.
806 for( uint i = 1; i < n->req(); i++ ) {
807 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) );
808 if( g != n->in(i) ) {
809 if( !x )
810 x = n->clone();
811 x->set_req(i, g);
812 }
813 }
814 if( x ) { // x can legally float to pre-header location
815 register_new_node( x, preheader_ctrl );
816 return x;
817 } else { // raise n to cover LCA of uses
818 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
819 }
820 return n;
821 }
823 //------------------------------insert_pre_post_loops--------------------------
824 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
825 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
826 // alignment. Useful to unroll loops that do no array accesses.
827 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
829 #ifndef PRODUCT
830 if (TraceLoopOpts) {
831 if (peel_only)
832 tty->print("PeelMainPost ");
833 else
834 tty->print("PreMainPost ");
835 loop->dump_head();
836 }
837 #endif
838 C->set_major_progress();
840 // Find common pieces of the loop being guarded with pre & post loops
841 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
842 assert( main_head->is_normal_loop(), "" );
843 CountedLoopEndNode *main_end = main_head->loopexit();
844 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
845 uint dd_main_head = dom_depth(main_head);
846 uint max = main_head->outcnt();
848 Node *pre_header= main_head->in(LoopNode::EntryControl);
849 Node *init = main_head->init_trip();
850 Node *incr = main_end ->incr();
851 Node *limit = main_end ->limit();
852 Node *stride = main_end ->stride();
853 Node *cmp = main_end ->cmp_node();
854 BoolTest::mask b_test = main_end->test_trip();
856 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
857 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
858 if( bol->outcnt() != 1 ) {
859 bol = bol->clone();
860 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
861 _igvn.hash_delete(main_end);
862 main_end->set_req(CountedLoopEndNode::TestValue, bol);
863 }
864 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
865 if( cmp->outcnt() != 1 ) {
866 cmp = cmp->clone();
867 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
868 _igvn.hash_delete(bol);
869 bol->set_req(1, cmp);
870 }
872 //------------------------------
873 // Step A: Create Post-Loop.
874 Node* main_exit = main_end->proj_out(false);
875 assert( main_exit->Opcode() == Op_IfFalse, "" );
876 int dd_main_exit = dom_depth(main_exit);
878 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
879 // loop pre-header illegally has 2 control users (old & new loops).
880 clone_loop( loop, old_new, dd_main_exit );
881 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
882 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
883 post_head->set_post_loop(main_head);
885 // Reduce the post-loop trip count.
886 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
887 post_end->_prob = PROB_FAIR;
889 // Build the main-loop normal exit.
890 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
891 _igvn.register_new_node_with_optimizer( new_main_exit );
892 set_idom(new_main_exit, main_end, dd_main_exit );
893 set_loop(new_main_exit, loop->_parent);
895 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
896 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
897 // (the main-loop trip-counter exit value) because we will be changing
898 // the exit value (via unrolling) so we cannot constant-fold away the zero
899 // trip guard until all unrolling is done.
900 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr);
901 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
902 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
903 register_new_node( zer_opaq, new_main_exit );
904 register_new_node( zer_cmp , new_main_exit );
905 register_new_node( zer_bol , new_main_exit );
907 // Build the IfNode
908 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
909 _igvn.register_new_node_with_optimizer( zer_iff );
910 set_idom(zer_iff, new_main_exit, dd_main_exit);
911 set_loop(zer_iff, loop->_parent);
913 // Plug in the false-path, taken if we need to skip post-loop
914 _igvn.hash_delete( main_exit );
915 main_exit->set_req(0, zer_iff);
916 _igvn._worklist.push(main_exit);
917 set_idom(main_exit, zer_iff, dd_main_exit);
918 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
919 // Make the true-path, must enter the post loop
920 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
921 _igvn.register_new_node_with_optimizer( zer_taken );
922 set_idom(zer_taken, zer_iff, dd_main_exit);
923 set_loop(zer_taken, loop->_parent);
924 // Plug in the true path
925 _igvn.hash_delete( post_head );
926 post_head->set_req(LoopNode::EntryControl, zer_taken);
927 set_idom(post_head, zer_taken, dd_main_exit);
929 // Step A3: Make the fall-in values to the post-loop come from the
930 // fall-out values of the main-loop.
931 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
932 Node* main_phi = main_head->fast_out(i);
933 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
934 Node *post_phi = old_new[main_phi->_idx];
935 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
936 post_head->init_control(),
937 main_phi->in(LoopNode::LoopBackControl));
938 _igvn.hash_delete(post_phi);
939 post_phi->set_req( LoopNode::EntryControl, fallmain );
940 }
941 }
943 // Update local caches for next stanza
944 main_exit = new_main_exit;
947 //------------------------------
948 // Step B: Create Pre-Loop.
950 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
951 // loop pre-header illegally has 2 control users (old & new loops).
952 clone_loop( loop, old_new, dd_main_head );
953 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
954 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
955 pre_head->set_pre_loop(main_head);
956 Node *pre_incr = old_new[incr->_idx];
958 // Reduce the pre-loop trip count.
959 pre_end->_prob = PROB_FAIR;
961 // Find the pre-loop normal exit.
962 Node* pre_exit = pre_end->proj_out(false);
963 assert( pre_exit->Opcode() == Op_IfFalse, "" );
964 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
965 _igvn.register_new_node_with_optimizer( new_pre_exit );
966 set_idom(new_pre_exit, pre_end, dd_main_head);
967 set_loop(new_pre_exit, loop->_parent);
969 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
970 // pre-loop, the main-loop may not execute at all. Later in life this
971 // zero-trip guard will become the minimum-trip guard when we unroll
972 // the main-loop.
973 Node *min_opaq = new (C, 2) Opaque1Node(C, limit);
974 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
975 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
976 register_new_node( min_opaq, new_pre_exit );
977 register_new_node( min_cmp , new_pre_exit );
978 register_new_node( min_bol , new_pre_exit );
980 // Build the IfNode (assume the main-loop is executed always).
981 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
982 _igvn.register_new_node_with_optimizer( min_iff );
983 set_idom(min_iff, new_pre_exit, dd_main_head);
984 set_loop(min_iff, loop->_parent);
986 // Plug in the false-path, taken if we need to skip main-loop
987 _igvn.hash_delete( pre_exit );
988 pre_exit->set_req(0, min_iff);
989 set_idom(pre_exit, min_iff, dd_main_head);
990 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
991 // Make the true-path, must enter the main loop
992 Node *min_taken = new (C, 1) IfTrueNode( min_iff );
993 _igvn.register_new_node_with_optimizer( min_taken );
994 set_idom(min_taken, min_iff, dd_main_head);
995 set_loop(min_taken, loop->_parent);
996 // Plug in the true path
997 _igvn.hash_delete( main_head );
998 main_head->set_req(LoopNode::EntryControl, min_taken);
999 set_idom(main_head, min_taken, dd_main_head);
1001 // Step B3: Make the fall-in values to the main-loop come from the
1002 // fall-out values of the pre-loop.
1003 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1004 Node* main_phi = main_head->fast_out(i2);
1005 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1006 Node *pre_phi = old_new[main_phi->_idx];
1007 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1008 main_head->init_control(),
1009 pre_phi->in(LoopNode::LoopBackControl));
1010 _igvn.hash_delete(main_phi);
1011 main_phi->set_req( LoopNode::EntryControl, fallpre );
1012 }
1013 }
1015 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1016 // RCE and alignment may change this later.
1017 Node *cmp_end = pre_end->cmp_node();
1018 assert( cmp_end->in(2) == limit, "" );
1019 Node *pre_limit = new (C, 3) AddINode( init, stride );
1021 // Save the original loop limit in this Opaque1 node for
1022 // use by range check elimination.
1023 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit);
1025 register_new_node( pre_limit, pre_head->in(0) );
1026 register_new_node( pre_opaq , pre_head->in(0) );
1028 // Since no other users of pre-loop compare, I can hack limit directly
1029 assert( cmp_end->outcnt() == 1, "no other users" );
1030 _igvn.hash_delete(cmp_end);
1031 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1033 // Special case for not-equal loop bounds:
1034 // Change pre loop test, main loop test, and the
1035 // main loop guard test to use lt or gt depending on stride
1036 // direction:
1037 // positive stride use <
1038 // negative stride use >
1040 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1042 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1043 // Modify pre loop end condition
1044 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1045 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
1046 register_new_node( new_bol0, pre_head->in(0) );
1047 _igvn.hash_delete(pre_end);
1048 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
1049 // Modify main loop guard condition
1050 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1051 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
1052 register_new_node( new_bol1, new_pre_exit );
1053 _igvn.hash_delete(min_iff);
1054 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1055 // Modify main loop end condition
1056 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1057 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
1058 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1059 _igvn.hash_delete(main_end);
1060 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
1061 }
1063 // Flag main loop
1064 main_head->set_main_loop();
1065 if( peel_only ) main_head->set_main_no_pre_loop();
1067 // It's difficult to be precise about the trip-counts
1068 // for the pre/post loops. They are usually very short,
1069 // so guess that 4 trips is a reasonable value.
1070 post_head->set_profile_trip_cnt(4.0);
1071 pre_head->set_profile_trip_cnt(4.0);
1073 // Now force out all loop-invariant dominating tests. The optimizer
1074 // finds some, but we _know_ they are all useless.
1075 peeled_dom_test_elim(loop,old_new);
1076 }
1078 //------------------------------is_invariant-----------------------------
1079 // Return true if n is invariant
1080 bool IdealLoopTree::is_invariant(Node* n) const {
1081 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1082 if (n_c->is_top()) return false;
1083 return !is_member(_phase->get_loop(n_c));
1084 }
1087 //------------------------------do_unroll--------------------------------------
1088 // Unroll the loop body one step - make each trip do 2 iterations.
1089 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1090 assert(LoopUnrollLimit, "");
1091 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1092 CountedLoopEndNode *loop_end = loop_head->loopexit();
1093 assert(loop_end, "");
1094 #ifndef PRODUCT
1095 if (PrintOpto && VerifyLoopOptimizations) {
1096 tty->print("Unrolling ");
1097 loop->dump_head();
1098 } else if (TraceLoopOpts) {
1099 if (loop_head->trip_count() < LoopUnrollLimit) {
1100 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1101 } else {
1102 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1103 }
1104 loop->dump_head();
1105 }
1106 #endif
1108 // Remember loop node count before unrolling to detect
1109 // if rounds of unroll,optimize are making progress
1110 loop_head->set_node_count_before_unroll(loop->_body.size());
1112 Node *ctrl = loop_head->in(LoopNode::EntryControl);
1113 Node *limit = loop_head->limit();
1114 Node *init = loop_head->init_trip();
1115 Node *stride = loop_head->stride();
1117 Node *opaq = NULL;
1118 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment
1119 assert( loop_head->is_main_loop(), "" );
1120 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1121 Node *iff = ctrl->in(0);
1122 assert( iff->Opcode() == Op_If, "" );
1123 Node *bol = iff->in(1);
1124 assert( bol->Opcode() == Op_Bool, "" );
1125 Node *cmp = bol->in(1);
1126 assert( cmp->Opcode() == Op_CmpI, "" );
1127 opaq = cmp->in(2);
1128 // Occasionally it's possible for a pre-loop Opaque1 node to be
1129 // optimized away and then another round of loop opts attempted.
1130 // We can not optimize this particular loop in that case.
1131 if( opaq->Opcode() != Op_Opaque1 )
1132 return; // Cannot find pre-loop! Bail out!
1133 }
1135 C->set_major_progress();
1137 // Adjust max trip count. The trip count is intentionally rounded
1138 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1139 // the main, unrolled, part of the loop will never execute as it is protected
1140 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1141 // and later determined that part of the unrolled loop was dead.
1142 loop_head->set_trip_count(loop_head->trip_count() / 2);
1144 // Double the count of original iterations in the unrolled loop body.
1145 loop_head->double_unrolled_count();
1147 // -----------
1148 // Step 2: Cut back the trip counter for an unroll amount of 2.
1149 // Loop will normally trip (limit - init)/stride_con. Since it's a
1150 // CountedLoop this is exact (stride divides limit-init exactly).
1151 // We are going to double the loop body, so we want to knock off any
1152 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
1153 Node *span = new (C, 3) SubINode( limit, init );
1154 register_new_node( span, ctrl );
1155 Node *trip = new (C, 3) DivINode( 0, span, stride );
1156 register_new_node( trip, ctrl );
1157 Node *mtwo = _igvn.intcon(-2);
1158 set_ctrl(mtwo, C->root());
1159 Node *rond = new (C, 3) AndINode( trip, mtwo );
1160 register_new_node( rond, ctrl );
1161 Node *spn2 = new (C, 3) MulINode( rond, stride );
1162 register_new_node( spn2, ctrl );
1163 Node *lim2 = new (C, 3) AddINode( spn2, init );
1164 register_new_node( lim2, ctrl );
1166 // Hammer in the new limit
1167 Node *ctrl2 = loop_end->in(0);
1168 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 );
1169 register_new_node( cmp2, ctrl2 );
1170 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
1171 register_new_node( bol2, ctrl2 );
1172 _igvn.hash_delete(loop_end);
1173 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
1175 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1176 // Make it a 1-trip test (means at least 2 trips).
1177 if( adjust_min_trip ) {
1178 // Guard test uses an 'opaque' node which is not shared. Hence I
1179 // can edit it's inputs directly. Hammer in the new limit for the
1180 // minimum-trip guard.
1181 assert( opaq->outcnt() == 1, "" );
1182 _igvn.hash_delete(opaq);
1183 opaq->set_req(1, lim2);
1184 }
1186 // ---------
1187 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1188 // represents the odd iterations; since the loop trips an even number of
1189 // times its backedge is never taken. Kill the backedge.
1190 uint dd = dom_depth(loop_head);
1191 clone_loop( loop, old_new, dd );
1193 // Make backedges of the clone equal to backedges of the original.
1194 // Make the fall-in from the original come from the fall-out of the clone.
1195 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1196 Node* phi = loop_head->fast_out(j);
1197 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1198 Node *newphi = old_new[phi->_idx];
1199 _igvn.hash_delete( phi );
1200 _igvn.hash_delete( newphi );
1202 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1203 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1204 phi ->set_req(LoopNode::LoopBackControl, C->top());
1205 }
1206 }
1207 Node *clone_head = old_new[loop_head->_idx];
1208 _igvn.hash_delete( clone_head );
1209 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1210 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1211 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1212 loop->_head = clone_head; // New loop header
1214 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1215 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1217 // Kill the clone's backedge
1218 Node *newcle = old_new[loop_end->_idx];
1219 _igvn.hash_delete( newcle );
1220 Node *one = _igvn.intcon(1);
1221 set_ctrl(one, C->root());
1222 newcle->set_req(1, one);
1223 // Force clone into same loop body
1224 uint max = loop->_body.size();
1225 for( uint k = 0; k < max; k++ ) {
1226 Node *old = loop->_body.at(k);
1227 Node *nnn = old_new[old->_idx];
1228 loop->_body.push(nnn);
1229 if (!has_ctrl(old))
1230 set_loop(nnn, loop);
1231 }
1233 loop->record_for_igvn();
1234 }
1236 //------------------------------do_maximally_unroll----------------------------
1238 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1239 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1240 assert(cl->trip_count() > 0, "");
1241 #ifndef PRODUCT
1242 if (TraceLoopOpts) {
1243 tty->print("MaxUnroll %d ", cl->trip_count());
1244 loop->dump_head();
1245 }
1246 #endif
1248 // If loop is tripping an odd number of times, peel odd iteration
1249 if ((cl->trip_count() & 1) == 1) {
1250 do_peeling(loop, old_new);
1251 }
1253 // Now its tripping an even number of times remaining. Double loop body.
1254 // Do not adjust pre-guards; they are not needed and do not exist.
1255 if (cl->trip_count() > 0) {
1256 do_unroll(loop, old_new, false);
1257 }
1258 }
1260 //------------------------------dominates_backedge---------------------------------
1261 // Returns true if ctrl is executed on every complete iteration
1262 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1263 assert(ctrl->is_CFG(), "must be control");
1264 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1265 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1266 }
1268 //------------------------------add_constraint---------------------------------
1269 // Constrain the main loop iterations so the condition:
1270 // scale_con * I + offset < limit
1271 // always holds true. That is, either increase the number of iterations in
1272 // the pre-loop or the post-loop until the condition holds true in the main
1273 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1274 // stride and scale are constants (offset and limit often are).
1275 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
1277 // Compute "I :: (limit-offset)/scale_con"
1278 Node *con = new (C, 3) SubINode( limit, offset );
1279 register_new_node( con, pre_ctrl );
1280 Node *scale = _igvn.intcon(scale_con);
1281 set_ctrl(scale, C->root());
1282 Node *X = new (C, 3) DivINode( 0, con, scale );
1283 register_new_node( X, pre_ctrl );
1285 // For positive stride, the pre-loop limit always uses a MAX function
1286 // and the main loop a MIN function. For negative stride these are
1287 // reversed.
1289 // Also for positive stride*scale the affine function is increasing, so the
1290 // pre-loop must check for underflow and the post-loop for overflow.
1291 // Negative stride*scale reverses this; pre-loop checks for overflow and
1292 // post-loop for underflow.
1293 if( stride_con*scale_con > 0 ) {
1294 // Compute I < (limit-offset)/scale_con
1295 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X)
1296 *main_limit = (stride_con > 0)
1297 ? (Node*)(new (C, 3) MinINode( *main_limit, X ))
1298 : (Node*)(new (C, 3) MaxINode( *main_limit, X ));
1299 register_new_node( *main_limit, pre_ctrl );
1301 } else {
1302 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I
1303 // Add the negation of the main-loop constraint to the pre-loop.
1304 // See footnote [++] below for a derivation of the limit expression.
1305 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1);
1306 set_ctrl(incr, C->root());
1307 Node *adj = new (C, 3) AddINode( X, incr );
1308 register_new_node( adj, pre_ctrl );
1309 *pre_limit = (scale_con > 0)
1310 ? (Node*)new (C, 3) MinINode( *pre_limit, adj )
1311 : (Node*)new (C, 3) MaxINode( *pre_limit, adj );
1312 register_new_node( *pre_limit, pre_ctrl );
1314 // [++] Here's the algebra that justifies the pre-loop limit expression:
1315 //
1316 // NOT( scale_con * I + offset < limit )
1317 // ==
1318 // scale_con * I + offset >= limit
1319 // ==
1320 // SGN(scale_con) * I >= (limit-offset)/|scale_con|
1321 // ==
1322 // (limit-offset)/|scale_con| <= I * SGN(scale_con)
1323 // ==
1324 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con)
1325 // ==
1326 // ( if (scale_con > 0) /*common case*/
1327 // (limit-offset)/scale_con - 1 < I
1328 // else
1329 // (limit-offset)/scale_con + 1 > I
1330 // )
1331 // ( if (scale_con > 0) /*common case*/
1332 // (limit-offset)/scale_con + SGN(-scale_con) < I
1333 // else
1334 // (limit-offset)/scale_con + SGN(-scale_con) > I
1335 }
1336 }
1339 //------------------------------is_scaled_iv---------------------------------
1340 // Return true if exp is a constant times an induction var
1341 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1342 if (exp == iv) {
1343 if (p_scale != NULL) {
1344 *p_scale = 1;
1345 }
1346 return true;
1347 }
1348 int opc = exp->Opcode();
1349 if (opc == Op_MulI) {
1350 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1351 if (p_scale != NULL) {
1352 *p_scale = exp->in(2)->get_int();
1353 }
1354 return true;
1355 }
1356 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1357 if (p_scale != NULL) {
1358 *p_scale = exp->in(1)->get_int();
1359 }
1360 return true;
1361 }
1362 } else if (opc == Op_LShiftI) {
1363 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1364 if (p_scale != NULL) {
1365 *p_scale = 1 << exp->in(2)->get_int();
1366 }
1367 return true;
1368 }
1369 }
1370 return false;
1371 }
1373 //-----------------------------is_scaled_iv_plus_offset------------------------------
1374 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1375 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1376 if (is_scaled_iv(exp, iv, p_scale)) {
1377 if (p_offset != NULL) {
1378 Node *zero = _igvn.intcon(0);
1379 set_ctrl(zero, C->root());
1380 *p_offset = zero;
1381 }
1382 return true;
1383 }
1384 int opc = exp->Opcode();
1385 if (opc == Op_AddI) {
1386 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1387 if (p_offset != NULL) {
1388 *p_offset = exp->in(2);
1389 }
1390 return true;
1391 }
1392 if (exp->in(2)->is_Con()) {
1393 Node* offset2 = NULL;
1394 if (depth < 2 &&
1395 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1396 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1397 if (p_offset != NULL) {
1398 Node *ctrl_off2 = get_ctrl(offset2);
1399 Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
1400 register_new_node(offset, ctrl_off2);
1401 *p_offset = offset;
1402 }
1403 return true;
1404 }
1405 }
1406 } else if (opc == Op_SubI) {
1407 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1408 if (p_offset != NULL) {
1409 Node *zero = _igvn.intcon(0);
1410 set_ctrl(zero, C->root());
1411 Node *ctrl_off = get_ctrl(exp->in(2));
1412 Node* offset = new (C, 3) SubINode(zero, exp->in(2));
1413 register_new_node(offset, ctrl_off);
1414 *p_offset = offset;
1415 }
1416 return true;
1417 }
1418 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1419 if (p_offset != NULL) {
1420 *p_scale *= -1;
1421 *p_offset = exp->in(1);
1422 }
1423 return true;
1424 }
1425 }
1426 return false;
1427 }
1429 //------------------------------do_range_check---------------------------------
1430 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1431 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1432 #ifndef PRODUCT
1433 if (PrintOpto && VerifyLoopOptimizations) {
1434 tty->print("Range Check Elimination ");
1435 loop->dump_head();
1436 } else if (TraceLoopOpts) {
1437 tty->print("RangeCheck ");
1438 loop->dump_head();
1439 }
1440 #endif
1441 assert(RangeCheckElimination, "");
1442 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1443 assert(cl->is_main_loop(), "");
1445 // protect against stride not being a constant
1446 if (!cl->stride_is_con())
1447 return;
1449 // Find the trip counter; we are iteration splitting based on it
1450 Node *trip_counter = cl->phi();
1451 // Find the main loop limit; we will trim it's iterations
1452 // to not ever trip end tests
1453 Node *main_limit = cl->limit();
1455 // Need to find the main-loop zero-trip guard
1456 Node *ctrl = cl->in(LoopNode::EntryControl);
1457 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1458 Node *iffm = ctrl->in(0);
1459 assert(iffm->Opcode() == Op_If, "");
1460 Node *bolzm = iffm->in(1);
1461 assert(bolzm->Opcode() == Op_Bool, "");
1462 Node *cmpzm = bolzm->in(1);
1463 assert(cmpzm->is_Cmp(), "");
1464 Node *opqzm = cmpzm->in(2);
1465 // Can not optimize a loop if pre-loop Opaque1 node is optimized
1466 // away and then another round of loop opts attempted.
1467 if (opqzm->Opcode() != Op_Opaque1)
1468 return;
1469 assert(opqzm->in(1) == main_limit, "do not understand situation");
1471 // Find the pre-loop limit; we will expand it's iterations to
1472 // not ever trip low tests.
1473 Node *p_f = iffm->in(0);
1474 assert(p_f->Opcode() == Op_IfFalse, "");
1475 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1476 assert(pre_end->loopnode()->is_pre_loop(), "");
1477 Node *pre_opaq1 = pre_end->limit();
1478 // Occasionally it's possible for a pre-loop Opaque1 node to be
1479 // optimized away and then another round of loop opts attempted.
1480 // We can not optimize this particular loop in that case.
1481 if (pre_opaq1->Opcode() != Op_Opaque1)
1482 return;
1483 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1484 Node *pre_limit = pre_opaq->in(1);
1486 // Where do we put new limit calculations
1487 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1489 // Ensure the original loop limit is available from the
1490 // pre-loop Opaque1 node.
1491 Node *orig_limit = pre_opaq->original_loop_limit();
1492 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1493 return;
1495 // Must know if its a count-up or count-down loop
1497 int stride_con = cl->stride_con();
1498 Node *zero = _igvn.intcon(0);
1499 Node *one = _igvn.intcon(1);
1500 set_ctrl(zero, C->root());
1501 set_ctrl(one, C->root());
1503 // Range checks that do not dominate the loop backedge (ie.
1504 // conditionally executed) can lengthen the pre loop limit beyond
1505 // the original loop limit. To prevent this, the pre limit is
1506 // (for stride > 0) MINed with the original loop limit (MAXed
1507 // stride < 0) when some range_check (rc) is conditionally
1508 // executed.
1509 bool conditional_rc = false;
1511 // Check loop body for tests of trip-counter plus loop-invariant vs
1512 // loop-invariant.
1513 for( uint i = 0; i < loop->_body.size(); i++ ) {
1514 Node *iff = loop->_body[i];
1515 if( iff->Opcode() == Op_If ) { // Test?
1517 // Test is an IfNode, has 2 projections. If BOTH are in the loop
1518 // we need loop unswitching instead of iteration splitting.
1519 Node *exit = loop->is_loop_exit(iff);
1520 if( !exit ) continue;
1521 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1523 // Get boolean condition to test
1524 Node *i1 = iff->in(1);
1525 if( !i1->is_Bool() ) continue;
1526 BoolNode *bol = i1->as_Bool();
1527 BoolTest b_test = bol->_test;
1528 // Flip sense of test if exit condition is flipped
1529 if( flip )
1530 b_test = b_test.negate();
1532 // Get compare
1533 Node *cmp = bol->in(1);
1535 // Look for trip_counter + offset vs limit
1536 Node *rc_exp = cmp->in(1);
1537 Node *limit = cmp->in(2);
1538 jint scale_con= 1; // Assume trip counter not scaled
1540 Node *limit_c = get_ctrl(limit);
1541 if( loop->is_member(get_loop(limit_c) ) ) {
1542 // Compare might have operands swapped; commute them
1543 b_test = b_test.commute();
1544 rc_exp = cmp->in(2);
1545 limit = cmp->in(1);
1546 limit_c = get_ctrl(limit);
1547 if( loop->is_member(get_loop(limit_c) ) )
1548 continue; // Both inputs are loop varying; cannot RCE
1549 }
1550 // Here we know 'limit' is loop invariant
1552 // 'limit' maybe pinned below the zero trip test (probably from a
1553 // previous round of rce), in which case, it can't be used in the
1554 // zero trip test expression which must occur before the zero test's if.
1555 if( limit_c == ctrl ) {
1556 continue; // Don't rce this check but continue looking for other candidates.
1557 }
1559 // Check for scaled induction variable plus an offset
1560 Node *offset = NULL;
1562 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1563 continue;
1564 }
1566 Node *offset_c = get_ctrl(offset);
1567 if( loop->is_member( get_loop(offset_c) ) )
1568 continue; // Offset is not really loop invariant
1569 // Here we know 'offset' is loop invariant.
1571 // As above for the 'limit', the 'offset' maybe pinned below the
1572 // zero trip test.
1573 if( offset_c == ctrl ) {
1574 continue; // Don't rce this check but continue looking for other candidates.
1575 }
1577 // At this point we have the expression as:
1578 // scale_con * trip_counter + offset :: limit
1579 // where scale_con, offset and limit are loop invariant. Trip_counter
1580 // monotonically increases by stride_con, a constant. Both (or either)
1581 // stride_con and scale_con can be negative which will flip about the
1582 // sense of the test.
1584 // Adjust pre and main loop limits to guard the correct iteration set
1585 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1586 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1587 // The overflow limit: scale*I+offset < limit
1588 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
1589 // The underflow limit: 0 <= scale*I+offset.
1590 // Some math yields: -scale*I-(offset+1) < 0
1591 Node *plus_one = new (C, 3) AddINode( offset, one );
1592 register_new_node( plus_one, pre_ctrl );
1593 Node *neg_offset = new (C, 3) SubINode( zero, plus_one );
1594 register_new_node( neg_offset, pre_ctrl );
1595 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit );
1596 if (!conditional_rc) {
1597 conditional_rc = !loop->dominates_backedge(iff);
1598 }
1599 } else {
1600 #ifndef PRODUCT
1601 if( PrintOpto )
1602 tty->print_cr("missed RCE opportunity");
1603 #endif
1604 continue; // In release mode, ignore it
1605 }
1606 } else { // Otherwise work on normal compares
1607 switch( b_test._test ) {
1608 case BoolTest::ge: // Convert X >= Y to -X <= -Y
1609 scale_con = -scale_con;
1610 offset = new (C, 3) SubINode( zero, offset );
1611 register_new_node( offset, pre_ctrl );
1612 limit = new (C, 3) SubINode( zero, limit );
1613 register_new_node( limit, pre_ctrl );
1614 // Fall into LE case
1615 case BoolTest::le: // Convert X <= Y to X < Y+1
1616 limit = new (C, 3) AddINode( limit, one );
1617 register_new_node( limit, pre_ctrl );
1618 // Fall into LT case
1619 case BoolTest::lt:
1620 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
1621 if (!conditional_rc) {
1622 conditional_rc = !loop->dominates_backedge(iff);
1623 }
1624 break;
1625 default:
1626 #ifndef PRODUCT
1627 if( PrintOpto )
1628 tty->print_cr("missed RCE opportunity");
1629 #endif
1630 continue; // Unhandled case
1631 }
1632 }
1634 // Kill the eliminated test
1635 C->set_major_progress();
1636 Node *kill_con = _igvn.intcon( 1-flip );
1637 set_ctrl(kill_con, C->root());
1638 _igvn.hash_delete(iff);
1639 iff->set_req(1, kill_con);
1640 _igvn._worklist.push(iff);
1641 // Find surviving projection
1642 assert(iff->is_If(), "");
1643 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
1644 // Find loads off the surviving projection; remove their control edge
1645 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1646 Node* cd = dp->fast_out(i); // Control-dependent node
1647 if( cd->is_Load() ) { // Loads can now float around in the loop
1648 _igvn.hash_delete(cd);
1649 // Allow the load to float around in the loop, or before it
1650 // but NOT before the pre-loop.
1651 cd->set_req(0, ctrl); // ctrl, not NULL
1652 _igvn._worklist.push(cd);
1653 --i;
1654 --imax;
1655 }
1656 }
1658 } // End of is IF
1660 }
1662 // Update loop limits
1663 if (conditional_rc) {
1664 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
1665 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
1666 register_new_node(pre_limit, pre_ctrl);
1667 }
1668 _igvn.hash_delete(pre_opaq);
1669 pre_opaq->set_req(1, pre_limit);
1671 // Note:: we are making the main loop limit no longer precise;
1672 // need to round up based on stride.
1673 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case
1674 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
1675 // Hopefully, compiler will optimize for powers of 2.
1676 Node *ctrl = get_ctrl(main_limit);
1677 Node *stride = cl->stride();
1678 Node *init = cl->init_trip();
1679 Node *span = new (C, 3) SubINode(main_limit,init);
1680 register_new_node(span,ctrl);
1681 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
1682 Node *add = new (C, 3) AddINode(span,rndup);
1683 register_new_node(add,ctrl);
1684 Node *div = new (C, 3) DivINode(0,add,stride);
1685 register_new_node(div,ctrl);
1686 Node *mul = new (C, 3) MulINode(div,stride);
1687 register_new_node(mul,ctrl);
1688 Node *newlim = new (C, 3) AddINode(mul,init);
1689 register_new_node(newlim,ctrl);
1690 main_limit = newlim;
1691 }
1693 Node *main_cle = cl->loopexit();
1694 Node *main_bol = main_cle->in(1);
1695 // Hacking loop bounds; need private copies of exit test
1696 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
1697 _igvn.hash_delete(main_cle);
1698 main_bol = main_bol->clone();// Clone a private BoolNode
1699 register_new_node( main_bol, main_cle->in(0) );
1700 main_cle->set_req(1,main_bol);
1701 }
1702 Node *main_cmp = main_bol->in(1);
1703 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
1704 _igvn.hash_delete(main_bol);
1705 main_cmp = main_cmp->clone();// Clone a private CmpNode
1706 register_new_node( main_cmp, main_cle->in(0) );
1707 main_bol->set_req(1,main_cmp);
1708 }
1709 // Hack the now-private loop bounds
1710 _igvn.hash_delete(main_cmp);
1711 main_cmp->set_req(2, main_limit);
1712 _igvn._worklist.push(main_cmp);
1713 // The OpaqueNode is unshared by design
1714 _igvn.hash_delete(opqzm);
1715 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
1716 opqzm->set_req(1,main_limit);
1717 _igvn._worklist.push(opqzm);
1718 }
1720 //------------------------------DCE_loop_body----------------------------------
1721 // Remove simplistic dead code from loop body
1722 void IdealLoopTree::DCE_loop_body() {
1723 for( uint i = 0; i < _body.size(); i++ )
1724 if( _body.at(i)->outcnt() == 0 )
1725 _body.map( i--, _body.pop() );
1726 }
1729 //------------------------------adjust_loop_exit_prob--------------------------
1730 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
1731 // Replace with a 1-in-10 exit guess.
1732 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
1733 Node *test = tail();
1734 while( test != _head ) {
1735 uint top = test->Opcode();
1736 if( top == Op_IfTrue || top == Op_IfFalse ) {
1737 int test_con = ((ProjNode*)test)->_con;
1738 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
1739 IfNode *iff = test->in(0)->as_If();
1740 if( iff->outcnt() == 2 ) { // Ignore dead tests
1741 Node *bol = iff->in(1);
1742 if( bol && bol->req() > 1 && bol->in(1) &&
1743 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
1744 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
1745 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
1746 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
1747 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
1748 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
1749 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
1750 return; // Allocation loops RARELY take backedge
1751 // Find the OTHER exit path from the IF
1752 Node* ex = iff->proj_out(1-test_con);
1753 float p = iff->_prob;
1754 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
1755 if( top == Op_IfTrue ) {
1756 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
1757 iff->_prob = PROB_STATIC_FREQUENT;
1758 }
1759 } else {
1760 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
1761 iff->_prob = PROB_STATIC_INFREQUENT;
1762 }
1763 }
1764 }
1765 }
1766 }
1767 test = phase->idom(test);
1768 }
1769 }
1772 //------------------------------policy_do_remove_empty_loop--------------------
1773 // Micro-benchmark spamming. Policy is to always remove empty loops.
1774 // The 'DO' part is to replace the trip counter with the value it will
1775 // have on the last iteration. This will break the loop.
1776 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
1777 // Minimum size must be empty loop
1778 if (_body.size() > EMPTY_LOOP_SIZE)
1779 return false;
1781 if (!_head->is_CountedLoop())
1782 return false; // Dead loop
1783 CountedLoopNode *cl = _head->as_CountedLoop();
1784 if (!cl->loopexit())
1785 return false; // Malformed loop
1786 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
1787 return false; // Infinite loop
1789 #ifdef ASSERT
1790 // Ensure only one phi which is the iv.
1791 Node* iv = NULL;
1792 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
1793 Node* n = cl->fast_out(i);
1794 if (n->Opcode() == Op_Phi) {
1795 assert(iv == NULL, "Too many phis" );
1796 iv = n;
1797 }
1798 }
1799 assert(iv == cl->phi(), "Wrong phi" );
1800 #endif
1802 // main and post loops have explicitly created zero trip guard
1803 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
1804 if (needs_guard) {
1805 // Check for an obvious zero trip guard.
1806 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
1807 if (inctrl->Opcode() == Op_IfTrue) {
1808 // The test should look like just the backedge of a CountedLoop
1809 Node* iff = inctrl->in(0);
1810 if (iff->is_If()) {
1811 Node* bol = iff->in(1);
1812 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
1813 Node* cmp = bol->in(1);
1814 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
1815 needs_guard = false;
1816 }
1817 }
1818 }
1819 }
1820 }
1822 #ifndef PRODUCT
1823 if (PrintOpto) {
1824 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
1825 this->dump_head();
1826 } else if (TraceLoopOpts) {
1827 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
1828 this->dump_head();
1829 }
1830 #endif
1832 if (needs_guard) {
1833 // Peel the loop to ensure there's a zero trip guard
1834 Node_List old_new;
1835 phase->do_peeling(this, old_new);
1836 }
1838 // Replace the phi at loop head with the final value of the last
1839 // iteration. Then the CountedLoopEnd will collapse (backedge never
1840 // taken) and all loop-invariant uses of the exit values will be correct.
1841 Node *phi = cl->phi();
1842 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() );
1843 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
1844 phase->_igvn.replace_node(phi,final);
1845 phase->C->set_major_progress();
1846 return true;
1847 }
1850 //=============================================================================
1851 //------------------------------iteration_split_impl---------------------------
1852 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
1853 // Check and remove empty loops (spam micro-benchmarks)
1854 if( policy_do_remove_empty_loop(phase) )
1855 return true; // Here we removed an empty loop
1857 bool should_peel = policy_peeling(phase); // Should we peel?
1859 bool should_unswitch = policy_unswitching(phase);
1861 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
1862 // This removes loop-invariant tests (usually null checks).
1863 if( !_head->is_CountedLoop() ) { // Non-counted loop
1864 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
1865 // Partial peel succeeded so terminate this round of loop opts
1866 return false;
1867 }
1868 if( should_peel ) { // Should we peel?
1869 #ifndef PRODUCT
1870 if (PrintOpto) tty->print_cr("should_peel");
1871 #endif
1872 phase->do_peeling(this,old_new);
1873 } else if( should_unswitch ) {
1874 phase->do_unswitching(this, old_new);
1875 }
1876 return true;
1877 }
1878 CountedLoopNode *cl = _head->as_CountedLoop();
1880 if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops
1882 // Do nothing special to pre- and post- loops
1883 if( cl->is_pre_loop() || cl->is_post_loop() ) return true;
1885 // Compute loop trip count from profile data
1886 compute_profile_trip_cnt(phase);
1888 // Before attempting fancy unrolling, RCE or alignment, see if we want
1889 // to completely unroll this loop or do loop unswitching.
1890 if( cl->is_normal_loop() ) {
1891 if (should_unswitch) {
1892 phase->do_unswitching(this, old_new);
1893 return true;
1894 }
1895 bool should_maximally_unroll = policy_maximally_unroll(phase);
1896 if( should_maximally_unroll ) {
1897 // Here we did some unrolling and peeling. Eventually we will
1898 // completely unroll this loop and it will no longer be a loop.
1899 phase->do_maximally_unroll(this,old_new);
1900 return true;
1901 }
1902 }
1904 // Skip next optimizations if running low on nodes. Note that
1905 // policy_unswitching and policy_maximally_unroll have this check.
1906 uint nodes_left = MaxNodeLimit - phase->C->unique();
1907 if ((2 * _body.size()) > nodes_left) {
1908 return true;
1909 }
1911 // Counted loops may be peeled, may need some iterations run up
1912 // front for RCE, and may want to align loop refs to a cache
1913 // line. Thus we clone a full loop up front whose trip count is
1914 // at least 1 (if peeling), but may be several more.
1916 // The main loop will start cache-line aligned with at least 1
1917 // iteration of the unrolled body (zero-trip test required) and
1918 // will have some range checks removed.
1920 // A post-loop will finish any odd iterations (leftover after
1921 // unrolling), plus any needed for RCE purposes.
1923 bool should_unroll = policy_unroll(phase);
1925 bool should_rce = policy_range_check(phase);
1927 bool should_align = policy_align(phase);
1929 // If not RCE'ing (iteration splitting) or Aligning, then we do not
1930 // need a pre-loop. We may still need to peel an initial iteration but
1931 // we will not be needing an unknown number of pre-iterations.
1932 //
1933 // Basically, if may_rce_align reports FALSE first time through,
1934 // we will not be able to later do RCE or Aligning on this loop.
1935 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
1937 // If we have any of these conditions (RCE, alignment, unrolling) met, then
1938 // we switch to the pre-/main-/post-loop model. This model also covers
1939 // peeling.
1940 if( should_rce || should_align || should_unroll ) {
1941 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops
1942 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
1944 // Adjust the pre- and main-loop limits to let the pre and post loops run
1945 // with full checks, but the main-loop with no checks. Remove said
1946 // checks from the main body.
1947 if( should_rce )
1948 phase->do_range_check(this,old_new);
1950 // Double loop body for unrolling. Adjust the minimum-trip test (will do
1951 // twice as many iterations as before) and the main body limit (only do
1952 // an even number of trips). If we are peeling, we might enable some RCE
1953 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
1954 // peeling.
1955 if( should_unroll && !should_peel )
1956 phase->do_unroll(this,old_new, true);
1958 // Adjust the pre-loop limits to align the main body
1959 // iterations.
1960 if( should_align )
1961 Unimplemented();
1963 } else { // Else we have an unchanged counted loop
1964 if( should_peel ) // Might want to peel but do nothing else
1965 phase->do_peeling(this,old_new);
1966 }
1967 return true;
1968 }
1971 //=============================================================================
1972 //------------------------------iteration_split--------------------------------
1973 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
1974 // Recursively iteration split nested loops
1975 if (_child && !_child->iteration_split(phase, old_new))
1976 return false;
1978 // Clean out prior deadwood
1979 DCE_loop_body();
1982 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
1983 // Replace with a 1-in-10 exit guess.
1984 if (_parent /*not the root loop*/ &&
1985 !_irreducible &&
1986 // Also ignore the occasional dead backedge
1987 !tail()->is_top()) {
1988 adjust_loop_exit_prob(phase);
1989 }
1991 // Gate unrolling, RCE and peeling efforts.
1992 if (!_child && // If not an inner loop, do not split
1993 !_irreducible &&
1994 _allow_optimizations &&
1995 !tail()->is_top()) { // Also ignore the occasional dead backedge
1996 if (!_has_call) {
1997 if (!iteration_split_impl(phase, old_new)) {
1998 return false;
1999 }
2000 } else if (policy_unswitching(phase)) {
2001 phase->do_unswitching(this, old_new);
2002 }
2003 }
2005 // Minor offset re-organization to remove loop-fallout uses of
2006 // trip counter when there was no major reshaping.
2007 phase->reorg_offsets(this);
2009 if (_next && !_next->iteration_split(phase, old_new))
2010 return false;
2011 return true;
2012 }
2015 //=============================================================================
2016 // Process all the loops in the loop tree and replace any fill
2017 // patterns with an intrisc version.
2018 bool PhaseIdealLoop::do_intrinsify_fill() {
2019 bool changed = false;
2020 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2021 IdealLoopTree* lpt = iter.current();
2022 changed |= intrinsify_fill(lpt);
2023 }
2024 return changed;
2025 }
2028 // Examine an inner loop looking for a a single store of an invariant
2029 // value in a unit stride loop,
2030 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2031 Node*& shift, Node*& con) {
2032 const char* msg = NULL;
2033 Node* msg_node = NULL;
2035 store_value = NULL;
2036 con = NULL;
2037 shift = NULL;
2039 // Process the loop looking for stores. If there are multiple
2040 // stores or extra control flow give at this point.
2041 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2042 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2043 Node* n = lpt->_body.at(i);
2044 if (n->outcnt() == 0) continue; // Ignore dead
2045 if (n->is_Store()) {
2046 if (store != NULL) {
2047 msg = "multiple stores";
2048 break;
2049 }
2050 int opc = n->Opcode();
2051 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) {
2052 msg = "oop fills not handled";
2053 break;
2054 }
2055 Node* value = n->in(MemNode::ValueIn);
2056 if (!lpt->is_invariant(value)) {
2057 msg = "variant store value";
2058 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2059 msg = "not array address";
2060 }
2061 store = n;
2062 store_value = value;
2063 } else if (n->is_If() && n != head->loopexit()) {
2064 msg = "extra control flow";
2065 msg_node = n;
2066 }
2067 }
2069 if (store == NULL) {
2070 // No store in loop
2071 return false;
2072 }
2074 if (msg == NULL && head->stride_con() != 1) {
2075 // could handle negative strides too
2076 if (head->stride_con() < 0) {
2077 msg = "negative stride";
2078 } else {
2079 msg = "non-unit stride";
2080 }
2081 }
2083 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2084 msg = "can't handle store address";
2085 msg_node = store->in(MemNode::Address);
2086 }
2088 if (msg == NULL &&
2089 (!store->in(MemNode::Memory)->is_Phi() ||
2090 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2091 msg = "store memory isn't proper phi";
2092 msg_node = store->in(MemNode::Memory);
2093 }
2095 // Make sure there is an appropriate fill routine
2096 BasicType t = store->as_Mem()->memory_type();
2097 const char* fill_name;
2098 if (msg == NULL &&
2099 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2100 msg = "unsupported store";
2101 msg_node = store;
2102 }
2104 if (msg != NULL) {
2105 #ifndef PRODUCT
2106 if (TraceOptimizeFill) {
2107 tty->print_cr("not fill intrinsic candidate: %s", msg);
2108 if (msg_node != NULL) msg_node->dump();
2109 }
2110 #endif
2111 return false;
2112 }
2114 // Make sure the address expression can be handled. It should be
2115 // head->phi * elsize + con. head->phi might have a ConvI2L.
2116 Node* elements[4];
2117 Node* conv = NULL;
2118 bool found_index = false;
2119 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2120 for (int e = 0; e < count; e++) {
2121 Node* n = elements[e];
2122 if (n->is_Con() && con == NULL) {
2123 con = n;
2124 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2125 Node* value = n->in(1);
2126 #ifdef _LP64
2127 if (value->Opcode() == Op_ConvI2L) {
2128 conv = value;
2129 value = value->in(1);
2130 }
2131 #endif
2132 if (value != head->phi()) {
2133 msg = "unhandled shift in address";
2134 } else {
2135 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2136 msg = "scale doesn't match";
2137 } else {
2138 found_index = true;
2139 shift = n;
2140 }
2141 }
2142 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2143 if (n->in(1) == head->phi()) {
2144 found_index = true;
2145 conv = n;
2146 } else {
2147 msg = "unhandled input to ConvI2L";
2148 }
2149 } else if (n == head->phi()) {
2150 // no shift, check below for allowed cases
2151 found_index = true;
2152 } else {
2153 msg = "unhandled node in address";
2154 msg_node = n;
2155 }
2156 }
2158 if (count == -1) {
2159 msg = "malformed address expression";
2160 msg_node = store;
2161 }
2163 if (!found_index) {
2164 msg = "missing use of index";
2165 }
2167 // byte sized items won't have a shift
2168 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2169 msg = "can't find shift";
2170 msg_node = store;
2171 }
2173 if (msg != NULL) {
2174 #ifndef PRODUCT
2175 if (TraceOptimizeFill) {
2176 tty->print_cr("not fill intrinsic: %s", msg);
2177 if (msg_node != NULL) msg_node->dump();
2178 }
2179 #endif
2180 return false;
2181 }
2183 // No make sure all the other nodes in the loop can be handled
2184 VectorSet ok(Thread::current()->resource_area());
2186 // store related values are ok
2187 ok.set(store->_idx);
2188 ok.set(store->in(MemNode::Memory)->_idx);
2190 // Loop structure is ok
2191 ok.set(head->_idx);
2192 ok.set(head->loopexit()->_idx);
2193 ok.set(head->phi()->_idx);
2194 ok.set(head->incr()->_idx);
2195 ok.set(head->loopexit()->cmp_node()->_idx);
2196 ok.set(head->loopexit()->in(1)->_idx);
2198 // Address elements are ok
2199 if (con) ok.set(con->_idx);
2200 if (shift) ok.set(shift->_idx);
2201 if (conv) ok.set(conv->_idx);
2203 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2204 Node* n = lpt->_body.at(i);
2205 if (n->outcnt() == 0) continue; // Ignore dead
2206 if (ok.test(n->_idx)) continue;
2207 // Backedge projection is ok
2208 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue;
2209 if (!n->is_AddP()) {
2210 msg = "unhandled node";
2211 msg_node = n;
2212 break;
2213 }
2214 }
2216 // Make sure no unexpected values are used outside the loop
2217 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2218 Node* n = lpt->_body.at(i);
2219 // These values can be replaced with other nodes if they are used
2220 // outside the loop.
2221 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2222 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2223 Node* use = iter.get();
2224 if (!lpt->_body.contains(use)) {
2225 msg = "node is used outside loop";
2226 // lpt->_body.dump();
2227 msg_node = n;
2228 break;
2229 }
2230 }
2231 }
2233 #ifdef ASSERT
2234 if (TraceOptimizeFill) {
2235 if (msg != NULL) {
2236 tty->print_cr("no fill intrinsic: %s", msg);
2237 if (msg_node != NULL) msg_node->dump();
2238 } else {
2239 tty->print_cr("fill intrinsic for:");
2240 }
2241 store->dump();
2242 if (Verbose) {
2243 lpt->_body.dump();
2244 }
2245 }
2246 #endif
2248 return msg == NULL;
2249 }
2253 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2254 // Only for counted inner loops
2255 if (!lpt->is_counted() || !lpt->is_inner()) {
2256 return false;
2257 }
2259 // Must have constant stride
2260 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2261 if (!head->stride_is_con() || !head->is_normal_loop()) {
2262 return false;
2263 }
2265 // Check that the body only contains a store of a loop invariant
2266 // value that is indexed by the loop phi.
2267 Node* store = NULL;
2268 Node* store_value = NULL;
2269 Node* shift = NULL;
2270 Node* offset = NULL;
2271 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2272 return false;
2273 }
2275 #ifndef PRODUCT
2276 if (TraceLoopOpts) {
2277 tty->print("ArrayFill ");
2278 lpt->dump_head();
2279 }
2280 #endif
2282 // Now replace the whole loop body by a call to a fill routine that
2283 // covers the same region as the loop.
2284 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2286 // Build an expression for the beginning of the copy region
2287 Node* index = head->init_trip();
2288 #ifdef _LP64
2289 index = new (C, 2) ConvI2LNode(index);
2290 _igvn.register_new_node_with_optimizer(index);
2291 #endif
2292 if (shift != NULL) {
2293 // byte arrays don't require a shift but others do.
2294 index = new (C, 3) LShiftXNode(index, shift->in(2));
2295 _igvn.register_new_node_with_optimizer(index);
2296 }
2297 index = new (C, 4) AddPNode(base, base, index);
2298 _igvn.register_new_node_with_optimizer(index);
2299 Node* from = new (C, 4) AddPNode(base, index, offset);
2300 _igvn.register_new_node_with_optimizer(from);
2301 // Compute the number of elements to copy
2302 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip());
2303 _igvn.register_new_node_with_optimizer(len);
2305 BasicType t = store->as_Mem()->memory_type();
2306 bool aligned = false;
2307 if (offset != NULL && head->init_trip()->is_Con()) {
2308 int element_size = type2aelembytes(t);
2309 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2310 }
2312 // Build a call to the fill routine
2313 const char* fill_name;
2314 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2315 assert(fill != NULL, "what?");
2317 // Convert float/double to int/long for fill routines
2318 if (t == T_FLOAT) {
2319 store_value = new (C, 2) MoveF2INode(store_value);
2320 _igvn.register_new_node_with_optimizer(store_value);
2321 } else if (t == T_DOUBLE) {
2322 store_value = new (C, 2) MoveD2LNode(store_value);
2323 _igvn.register_new_node_with_optimizer(store_value);
2324 }
2326 Node* mem_phi = store->in(MemNode::Memory);
2327 Node* result_ctrl;
2328 Node* result_mem;
2329 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2330 int size = call_type->domain()->cnt();
2331 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill,
2332 fill_name, TypeAryPtr::get_array_body_type(t));
2333 call->init_req(TypeFunc::Parms+0, from);
2334 call->init_req(TypeFunc::Parms+1, store_value);
2335 #ifdef _LP64
2336 len = new (C, 2) ConvI2LNode(len);
2337 _igvn.register_new_node_with_optimizer(len);
2338 #endif
2339 call->init_req(TypeFunc::Parms+2, len);
2340 #ifdef _LP64
2341 call->init_req(TypeFunc::Parms+3, C->top());
2342 #endif
2343 call->init_req( TypeFunc::Control, head->init_control());
2344 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o
2345 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) );
2346 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) );
2347 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) );
2348 _igvn.register_new_node_with_optimizer(call);
2349 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
2350 _igvn.register_new_node_with_optimizer(result_ctrl);
2351 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory);
2352 _igvn.register_new_node_with_optimizer(result_mem);
2354 // If this fill is tightly coupled to an allocation and overwrites
2355 // the whole body, allow it to take over the zeroing.
2356 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2357 if (alloc != NULL && alloc->is_AllocateArray()) {
2358 Node* length = alloc->as_AllocateArray()->Ideal_length();
2359 if (head->limit() == length &&
2360 head->init_trip() == _igvn.intcon(0)) {
2361 if (TraceOptimizeFill) {
2362 tty->print_cr("Eliminated zeroing in allocation");
2363 }
2364 alloc->maybe_set_complete(&_igvn);
2365 } else {
2366 #ifdef ASSERT
2367 if (TraceOptimizeFill) {
2368 tty->print_cr("filling array but bounds don't match");
2369 alloc->dump();
2370 head->init_trip()->dump();
2371 head->limit()->dump();
2372 length->dump();
2373 }
2374 #endif
2375 }
2376 }
2378 // Redirect the old control and memory edges that are outside the loop.
2379 Node* exit = head->loopexit()->proj_out(0);
2380 // Sometimes the memory phi of the head is used as the outgoing
2381 // state of the loop. It's safe in this case to replace it with the
2382 // result_mem.
2383 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
2384 _igvn.replace_node(exit, result_ctrl);
2385 _igvn.replace_node(store, result_mem);
2386 // Any uses the increment outside of the loop become the loop limit.
2387 _igvn.replace_node(head->incr(), head->limit());
2389 // Disconnect the head from the loop.
2390 for (uint i = 0; i < lpt->_body.size(); i++) {
2391 Node* n = lpt->_body.at(i);
2392 _igvn.replace_node(n, C->top());
2393 }
2395 return true;
2396 }