1 /*
2 * Copyright (c) 1998, 2020, 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 "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/superword.hpp"
39
40 //=============================================================================
41 //------------------------------is_loop_iv-------------------------------------
42 // Determine if a node is Counted loop induction variable.
43 // The method is declared in node.hpp.
44 const Node* Node::is_loop_iv() const {
45 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
46 this->as_Phi()->region()->is_CountedLoop() &&
47 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
48 return this;
49 } else {
50 return NULL;
51 }
52 }
53
54 //=============================================================================
55 //------------------------------dump_spec--------------------------------------
56 // Dump special per-node info
57 #ifndef PRODUCT
58 void LoopNode::dump_spec(outputStream *st) const {
59 if (is_inner_loop()) st->print( "inner " );
60 if (is_partial_peel_loop()) st->print( "partial_peel " );
61 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
62 }
63 #endif
64
65 //------------------------------is_valid_counted_loop-------------------------
66 bool LoopNode::is_valid_counted_loop() const {
67 if (is_CountedLoop()) {
68 CountedLoopNode* l = as_CountedLoop();
69 CountedLoopEndNode* le = l->loopexit();
70 if (le != NULL &&
71 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
72 Node* phi = l->phi();
73 Node* exit = le->proj_out(0 /* false */);
74 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
75 phi != NULL && phi->is_Phi() &&
76 phi->in(LoopNode::LoopBackControl) == l->incr() &&
77 le->loopnode() == l && le->stride_is_con()) {
78 return true;
79 }
80 }
81 }
82 return false;
83 }
84
85 //------------------------------get_early_ctrl---------------------------------
86 // Compute earliest legal control
87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
88 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
89 uint i;
90 Node *early;
91 if (n->in(0) && !n->is_expensive()) {
92 early = n->in(0);
93 if (!early->is_CFG()) // Might be a non-CFG multi-def
94 early = get_ctrl(early); // So treat input as a straight data input
95 i = 1;
96 } else {
97 early = get_ctrl(n->in(1));
98 i = 2;
99 }
100 uint e_d = dom_depth(early);
101 assert( early, "" );
102 for (; i < n->req(); i++) {
103 Node *cin = get_ctrl(n->in(i));
104 assert( cin, "" );
105 // Keep deepest dominator depth
106 uint c_d = dom_depth(cin);
107 if (c_d > e_d) { // Deeper guy?
108 early = cin; // Keep deepest found so far
109 e_d = c_d;
110 } else if (c_d == e_d && // Same depth?
111 early != cin) { // If not equal, must use slower algorithm
112 // If same depth but not equal, one _must_ dominate the other
113 // and we want the deeper (i.e., dominated) guy.
114 Node *n1 = early;
115 Node *n2 = cin;
116 while (1) {
117 n1 = idom(n1); // Walk up until break cycle
118 n2 = idom(n2);
119 if (n1 == cin || // Walked early up to cin
120 dom_depth(n2) < c_d)
121 break; // early is deeper; keep him
122 if (n2 == early || // Walked cin up to early
123 dom_depth(n1) < c_d) {
124 early = cin; // cin is deeper; keep him
125 break;
126 }
127 }
128 e_d = dom_depth(early); // Reset depth register cache
129 }
130 }
131
132 // Return earliest legal location
133 assert(early == find_non_split_ctrl(early), "unexpected early control");
134
135 if (n->is_expensive()) {
136 assert(n->in(0), "should have control input");
137 early = get_early_ctrl_for_expensive(n, early);
138 }
139
140 return early;
141 }
142
143 //------------------------------get_early_ctrl_for_expensive---------------------------------
144 // Move node up the dominator tree as high as legal while still beneficial
145 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
146 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
147 assert(OptimizeExpensiveOps, "optimization off?");
148
149 Node* ctl = n->in(0);
150 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
151 uint min_dom_depth = dom_depth(earliest);
152 #ifdef ASSERT
153 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
154 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
155 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
156 }
157 #endif
158 if (dom_depth(ctl) < min_dom_depth) {
159 return earliest;
160 }
161
162 while (1) {
163 Node *next = ctl;
164 // Moving the node out of a loop on the projection of a If
165 // confuses loop predication. So once we hit a Loop in a If branch
166 // that doesn't branch to an UNC, we stop. The code that process
167 // expensive nodes will notice the loop and skip over it to try to
168 // move the node further up.
169 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
170 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
171 break;
172 }
173 next = idom(ctl->in(1)->in(0));
174 } else if (ctl->is_Proj()) {
175 // We only move it up along a projection if the projection is
176 // the single control projection for its parent: same code path,
177 // if it's a If with UNC or fallthrough of a call.
178 Node* parent_ctl = ctl->in(0);
179 if (parent_ctl == NULL) {
180 break;
181 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
182 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
183 } else if (parent_ctl->is_If()) {
184 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
185 break;
186 }
187 assert(idom(ctl) == parent_ctl, "strange");
188 next = idom(parent_ctl);
189 } else if (ctl->is_CatchProj()) {
190 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
191 break;
192 }
193 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
194 next = parent_ctl->in(0)->in(0)->in(0);
195 } else {
196 // Check if parent control has a single projection (this
197 // control is the only possible successor of the parent
198 // control). If so, we can try to move the node above the
199 // parent control.
200 int nb_ctl_proj = 0;
201 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
202 Node *p = parent_ctl->fast_out(i);
203 if (p->is_Proj() && p->is_CFG()) {
204 nb_ctl_proj++;
205 if (nb_ctl_proj > 1) {
206 break;
207 }
208 }
209 }
210
211 if (nb_ctl_proj > 1) {
212 break;
213 }
214 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node");
215 assert(idom(ctl) == parent_ctl, "strange");
216 next = idom(parent_ctl);
217 }
218 } else {
219 next = idom(ctl);
220 }
221 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
222 break;
223 }
224 ctl = next;
225 }
226
227 if (ctl != n->in(0)) {
228 _igvn.hash_delete(n);
229 n->set_req(0, ctl);
230 _igvn.hash_insert(n);
231 }
232
233 return ctl;
234 }
235
236
237 //------------------------------set_early_ctrl---------------------------------
238 // Set earliest legal control
239 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
240 Node *early = get_early_ctrl(n);
241
242 // Record earliest legal location
243 set_ctrl(n, early);
244 }
245
246 //------------------------------set_subtree_ctrl-------------------------------
247 // set missing _ctrl entries on new nodes
248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
249 // Already set? Get out.
250 if( _nodes[n->_idx] ) return;
251 // Recursively set _nodes array to indicate where the Node goes
252 uint i;
253 for( i = 0; i < n->req(); ++i ) {
254 Node *m = n->in(i);
255 if( m && m != C->root() )
256 set_subtree_ctrl( m );
257 }
258
259 // Fixup self
260 set_early_ctrl( n );
261 }
262
263 //------------------------------is_counted_loop--------------------------------
264 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
265 PhaseGVN *gvn = &_igvn;
266
267 // Counted loop head must be a good RegionNode with only 3 not NULL
268 // control input edges: Self, Entry, LoopBack.
269 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
270 return false;
271 }
272 Node *init_control = x->in(LoopNode::EntryControl);
273 Node *back_control = x->in(LoopNode::LoopBackControl);
274 if (init_control == NULL || back_control == NULL) // Partially dead
275 return false;
276 // Must also check for TOP when looking for a dead loop
277 if (init_control->is_top() || back_control->is_top())
278 return false;
279
280 // Allow funny placement of Safepoint
281 if (back_control->Opcode() == Op_SafePoint) {
282 if (UseCountedLoopSafepoints) {
283 // Leaving the safepoint on the backedge and creating a
284 // CountedLoop will confuse optimizations. We can't move the
285 // safepoint around because its jvm state wouldn't match a new
286 // location. Give up on that loop.
287 return false;
288 }
289 back_control = back_control->in(TypeFunc::Control);
290 }
291
292 // Controlling test for loop
293 Node *iftrue = back_control;
294 uint iftrue_op = iftrue->Opcode();
295 if (iftrue_op != Op_IfTrue &&
296 iftrue_op != Op_IfFalse)
297 // I have a weird back-control. Probably the loop-exit test is in
298 // the middle of the loop and I am looking at some trailing control-flow
299 // merge point. To fix this I would have to partially peel the loop.
300 return false; // Obscure back-control
301
302 // Get boolean guarding loop-back test
303 Node *iff = iftrue->in(0);
304 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
305 return false;
306 BoolNode *test = iff->in(1)->as_Bool();
307 BoolTest::mask bt = test->_test._test;
308 float cl_prob = iff->as_If()->_prob;
309 if (iftrue_op == Op_IfFalse) {
310 bt = BoolTest(bt).negate();
311 cl_prob = 1.0 - cl_prob;
312 }
313 // Get backedge compare
314 Node *cmp = test->in(1);
315 int cmp_op = cmp->Opcode();
316 if (cmp_op != Op_CmpI)
317 return false; // Avoid pointer & float compares
318
319 // Find the trip-counter increment & limit. Limit must be loop invariant.
320 Node *incr = cmp->in(1);
321 Node *limit = cmp->in(2);
322
323 // ---------
324 // need 'loop()' test to tell if limit is loop invariant
325 // ---------
326
327 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
328 Node *tmp = incr; // Then reverse order into the CmpI
329 incr = limit;
330 limit = tmp;
331 bt = BoolTest(bt).commute(); // And commute the exit test
332 }
333 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
334 return false;
335 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
336 return false;
337
338 Node* phi_incr = NULL;
339 // Trip-counter increment must be commutative & associative.
340 if (incr->is_Phi()) {
341 if (incr->as_Phi()->region() != x || incr->req() != 3)
342 return false; // Not simple trip counter expression
343 phi_incr = incr;
344 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
345 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
346 return false;
347 }
348
349 Node* trunc1 = NULL;
350 Node* trunc2 = NULL;
351 const TypeInt* iv_trunc_t = NULL;
352 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
353 return false; // Funny increment opcode
354 }
355 assert(incr->Opcode() == Op_AddI, "wrong increment code");
356
357 // Get merge point
358 Node *xphi = incr->in(1);
359 Node *stride = incr->in(2);
360 if (!stride->is_Con()) { // Oops, swap these
361 if (!xphi->is_Con()) // Is the other guy a constant?
362 return false; // Nope, unknown stride, bail out
363 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
364 xphi = stride;
365 stride = tmp;
366 }
367 // Stride must be constant
368 int stride_con = stride->get_int();
369 if (stride_con == 0)
370 return false; // missed some peephole opt
371
372 if (!xphi->is_Phi())
373 return false; // Too much math on the trip counter
374 if (phi_incr != NULL && phi_incr != xphi)
375 return false;
376 PhiNode *phi = xphi->as_Phi();
377
378 // Phi must be of loop header; backedge must wrap to increment
379 if (phi->region() != x)
380 return false;
381 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
382 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
383 return false;
384 }
385 Node *init_trip = phi->in(LoopNode::EntryControl);
386
387 // If iv trunc type is smaller than int, check for possible wrap.
388 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
389 assert(trunc1 != NULL, "must have found some truncation");
390
391 // Get a better type for the phi (filtered thru if's)
392 const TypeInt* phi_ft = filtered_type(phi);
393
394 // Can iv take on a value that will wrap?
395 //
396 // Ensure iv's limit is not within "stride" of the wrap value.
397 //
398 // Example for "short" type
399 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
400 // If the stride is +10, then the last value of the induction
401 // variable before the increment (phi_ft->_hi) must be
402 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
403 // ensure no truncation occurs after the increment.
404
405 if (stride_con > 0) {
406 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
407 iv_trunc_t->_lo > phi_ft->_lo) {
408 return false; // truncation may occur
409 }
410 } else if (stride_con < 0) {
411 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
412 iv_trunc_t->_hi < phi_ft->_hi) {
413 return false; // truncation may occur
414 }
415 }
416 // No possibility of wrap so truncation can be discarded
417 // Promote iv type to Int
418 } else {
419 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
420 }
421
422 // If the condition is inverted and we will be rolling
423 // through MININT to MAXINT, then bail out.
424 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
425 // Odd stride
426 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
427 // Count down loop rolls through MAXINT
428 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
429 // Count up loop rolls through MININT
430 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
431 return false; // Bail out
432 }
433
434 const TypeInt* init_t = gvn->type(init_trip)->is_int();
435 const TypeInt* limit_t = gvn->type(limit)->is_int();
436
437 if (stride_con > 0) {
438 jlong init_p = (jlong)init_t->_lo + stride_con;
439 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
440 return false; // cyclic loop or this loop trips only once
441 } else {
442 jlong init_p = (jlong)init_t->_hi + stride_con;
443 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
444 return false; // cyclic loop or this loop trips only once
445 }
446
447 if (phi_incr != NULL) {
448 // check if there is a possiblity of IV overflowing after the first increment
449 if (stride_con > 0) {
450 if (init_t->_hi > max_jint - stride_con) {
451 return false;
452 }
453 } else {
454 if (init_t->_lo < min_jint - stride_con) {
455 return false;
456 }
457 }
458 }
459
460 // =================================================
461 // ---- SUCCESS! Found A Trip-Counted Loop! -----
462 //
463 assert(x->Opcode() == Op_Loop, "regular loops only");
464 C->print_method(PHASE_BEFORE_CLOOPS, 3);
465
466 Node *hook = new (C) Node(6);
467
468 if (LoopLimitCheck) {
469
470 // ===================================================
471 // Generate loop limit check to avoid integer overflow
472 // in cases like next (cyclic loops):
473 //
474 // for (i=0; i <= max_jint; i++) {}
475 // for (i=0; i < max_jint; i+=2) {}
476 //
477 //
478 // Limit check predicate depends on the loop test:
479 //
480 // for(;i != limit; i++) --> limit <= (max_jint)
481 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
482 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
483 //
484
485 // Check if limit is excluded to do more precise int overflow check.
486 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
487 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
488
489 // If compare points directly to the phi we need to adjust
490 // the compare so that it points to the incr. Limit have
491 // to be adjusted to keep trip count the same and the
492 // adjusted limit should be checked for int overflow.
493 if (phi_incr != NULL) {
494 stride_m += stride_con;
495 }
496
497 if (limit->is_Con()) {
498 int limit_con = limit->get_int();
499 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
500 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
501 // Bailout: it could be integer overflow.
502 return false;
503 }
504 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
505 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
506 // Limit's type may satisfy the condition, for example,
507 // when it is an array length.
508 } else {
509 // Generate loop's limit check.
510 // Loop limit check predicate should be near the loop.
511 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
512 if (!limit_check_proj) {
513 // The limit check predicate is not generated if this method trapped here before.
514 #ifdef ASSERT
515 if (TraceLoopLimitCheck) {
516 tty->print("missing loop limit check:");
517 loop->dump_head();
518 x->dump(1);
519 }
520 #endif
521 return false;
522 }
523
524 IfNode* check_iff = limit_check_proj->in(0)->as_If();
525 Node* cmp_limit;
526 Node* bol;
527
528 if (stride_con > 0) {
529 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
530 bol = new (C) BoolNode(cmp_limit, BoolTest::le);
531 } else {
532 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
533 bol = new (C) BoolNode(cmp_limit, BoolTest::ge);
534 }
535 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
536 bol = _igvn.register_new_node_with_optimizer(bol);
537 set_subtree_ctrl(bol);
538
539 // Replace condition in original predicate but preserve Opaque node
540 // so that previous predicates could be found.
541 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
542 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
543 Node* opq = check_iff->in(1)->in(1);
544 _igvn.hash_delete(opq);
545 opq->set_req(1, bol);
546 // Update ctrl.
547 set_ctrl(opq, check_iff->in(0));
548 set_ctrl(check_iff->in(1), check_iff->in(0));
549
550 #ifndef PRODUCT
551 // report that the loop predication has been actually performed
552 // for this loop
553 if (TraceLoopLimitCheck) {
554 tty->print_cr("Counted Loop Limit Check generated:");
555 debug_only( bol->dump(2); )
556 }
557 #endif
558 }
559
560 if (phi_incr != NULL) {
561 // If compare points directly to the phi we need to adjust
562 // the compare so that it points to the incr. Limit have
563 // to be adjusted to keep trip count the same and we
564 // should avoid int overflow.
565 //
566 // i = init; do {} while(i++ < limit);
567 // is converted to
568 // i = init; do {} while(++i < limit+1);
569 //
570 limit = gvn->transform(new (C) AddINode(limit, stride));
571 }
572
573 // Now we need to canonicalize loop condition.
574 if (bt == BoolTest::ne) {
575 assert(stride_con == 1 || stride_con == -1, "simple increment only");
576 // 'ne' can be replaced with 'lt' only when init < limit.
577 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
578 bt = BoolTest::lt;
579 // 'ne' can be replaced with 'gt' only when init > limit.
580 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
581 bt = BoolTest::gt;
582 }
583
584 if (incl_limit) {
585 // The limit check guaranties that 'limit <= (max_jint - stride)' so
586 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
587 //
588 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
589 limit = gvn->transform(new (C) AddINode(limit, one));
590 if (bt == BoolTest::le)
591 bt = BoolTest::lt;
592 else if (bt == BoolTest::ge)
593 bt = BoolTest::gt;
594 else
595 ShouldNotReachHere();
596 }
597 set_subtree_ctrl( limit );
598
599 } else { // LoopLimitCheck
600
601 // If compare points to incr, we are ok. Otherwise the compare
602 // can directly point to the phi; in this case adjust the compare so that
603 // it points to the incr by adjusting the limit.
604 if (cmp->in(1) == phi || cmp->in(2) == phi)
605 limit = gvn->transform(new (C) AddINode(limit,stride));
606
607 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
608 // Final value for iterator should be: trip_count * stride + init_trip.
609 Node *one_p = gvn->intcon( 1);
610 Node *one_m = gvn->intcon(-1);
611
612 Node *trip_count = NULL;
613 switch( bt ) {
614 case BoolTest::eq:
615 ShouldNotReachHere();
616 case BoolTest::ne: // Ahh, the case we desire
617 if (stride_con == 1)
618 trip_count = gvn->transform(new (C) SubINode(limit,init_trip));
619 else if (stride_con == -1)
620 trip_count = gvn->transform(new (C) SubINode(init_trip,limit));
621 else
622 ShouldNotReachHere();
623 set_subtree_ctrl(trip_count);
624 //_loop.map(trip_count->_idx,loop(limit));
625 break;
626 case BoolTest::le: // Maybe convert to '<' case
627 limit = gvn->transform(new (C) AddINode(limit,one_p));
628 set_subtree_ctrl( limit );
629 hook->init_req(4, limit);
630
631 bt = BoolTest::lt;
632 // Make the new limit be in the same loop nest as the old limit
633 //_loop.map(limit->_idx,limit_loop);
634 // Fall into next case
635 case BoolTest::lt: { // Maybe convert to '!=' case
636 if (stride_con < 0) // Count down loop rolls through MAXINT
637 ShouldNotReachHere();
638 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
639 set_subtree_ctrl( range );
640 hook->init_req(0, range);
641
642 Node *bias = gvn->transform(new (C) AddINode(range,stride));
643 set_subtree_ctrl( bias );
644 hook->init_req(1, bias);
645
646 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_m));
647 set_subtree_ctrl( bias1 );
648 hook->init_req(2, bias1);
649
650 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
651 set_subtree_ctrl( trip_count );
652 hook->init_req(3, trip_count);
653 break;
654 }
655
656 case BoolTest::ge: // Maybe convert to '>' case
657 limit = gvn->transform(new (C) AddINode(limit,one_m));
658 set_subtree_ctrl( limit );
659 hook->init_req(4 ,limit);
660
661 bt = BoolTest::gt;
662 // Make the new limit be in the same loop nest as the old limit
663 //_loop.map(limit->_idx,limit_loop);
664 // Fall into next case
665 case BoolTest::gt: { // Maybe convert to '!=' case
666 if (stride_con > 0) // count up loop rolls through MININT
667 ShouldNotReachHere();
668 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
669 set_subtree_ctrl( range );
670 hook->init_req(0, range);
671
672 Node *bias = gvn->transform(new (C) AddINode(range,stride));
673 set_subtree_ctrl( bias );
674 hook->init_req(1, bias);
675
676 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_p));
677 set_subtree_ctrl( bias1 );
678 hook->init_req(2, bias1);
679
680 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
681 set_subtree_ctrl( trip_count );
682 hook->init_req(3, trip_count);
683 break;
684 }
685 } // switch( bt )
686
687 Node *span = gvn->transform(new (C) MulINode(trip_count,stride));
688 set_subtree_ctrl( span );
689 hook->init_req(5, span);
690
691 limit = gvn->transform(new (C) AddINode(span,init_trip));
692 set_subtree_ctrl( limit );
693
694 } // LoopLimitCheck
695
696 if (!UseCountedLoopSafepoints) {
697 // Check for SafePoint on backedge and remove
698 Node *sfpt = x->in(LoopNode::LoopBackControl);
699 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
700 lazy_replace( sfpt, iftrue );
701 if (loop->_safepts != NULL) {
702 loop->_safepts->yank(sfpt);
703 }
704 loop->_tail = iftrue;
705 }
706 }
707
708 // Build a canonical trip test.
709 // Clone code, as old values may be in use.
710 incr = incr->clone();
711 incr->set_req(1,phi);
712 incr->set_req(2,stride);
713 incr = _igvn.register_new_node_with_optimizer(incr);
714 set_early_ctrl( incr );
715 _igvn.hash_delete(phi);
716 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
717
718 // If phi type is more restrictive than Int, raise to
719 // Int to prevent (almost) infinite recursion in igvn
720 // which can only handle integer types for constants or minint..maxint.
721 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
722 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
723 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
724 nphi = _igvn.register_new_node_with_optimizer(nphi);
725 set_ctrl(nphi, get_ctrl(phi));
726 _igvn.replace_node(phi, nphi);
727 phi = nphi->as_Phi();
728 }
729 cmp = cmp->clone();
730 cmp->set_req(1,incr);
731 cmp->set_req(2,limit);
732 cmp = _igvn.register_new_node_with_optimizer(cmp);
733 set_ctrl(cmp, iff->in(0));
734
735 test = test->clone()->as_Bool();
736 (*(BoolTest*)&test->_test)._test = bt;
737 test->set_req(1,cmp);
738 _igvn.register_new_node_with_optimizer(test);
739 set_ctrl(test, iff->in(0));
740
741 // Replace the old IfNode with a new LoopEndNode
742 Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
743 IfNode *le = lex->as_If();
744 uint dd = dom_depth(iff);
745 set_idom(le, le->in(0), dd); // Update dominance for loop exit
746 set_loop(le, loop);
747
748 // Get the loop-exit control
749 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
750
751 // Need to swap loop-exit and loop-back control?
752 if (iftrue_op == Op_IfFalse) {
753 Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le));
754 Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le));
755
756 loop->_tail = back_control = ift2;
757 set_loop(ift2, loop);
758 set_loop(iff2, get_loop(iffalse));
759
760 // Lazy update of 'get_ctrl' mechanism.
761 lazy_replace(iffalse, iff2);
762 lazy_replace(iftrue, ift2);
763
764 // Swap names
765 iffalse = iff2;
766 iftrue = ift2;
767 } else {
768 _igvn.hash_delete(iffalse);
769 _igvn.hash_delete(iftrue);
770 iffalse->set_req_X( 0, le, &_igvn );
771 iftrue ->set_req_X( 0, le, &_igvn );
772 }
773
774 set_idom(iftrue, le, dd+1);
775 set_idom(iffalse, le, dd+1);
776 assert(iff->outcnt() == 0, "should be dead now");
777 lazy_replace( iff, le ); // fix 'get_ctrl'
778
779 // Now setup a new CountedLoopNode to replace the existing LoopNode
780 CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control);
781 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
782 // The following assert is approximately true, and defines the intention
783 // of can_be_counted_loop. It fails, however, because phase->type
784 // is not yet initialized for this loop and its parts.
785 //assert(l->can_be_counted_loop(this), "sanity");
786 _igvn.register_new_node_with_optimizer(l);
787 set_loop(l, loop);
788 loop->_head = l;
789 // Fix all data nodes placed at the old loop head.
790 // Uses the lazy-update mechanism of 'get_ctrl'.
791 lazy_replace( x, l );
792 set_idom(l, init_control, dom_depth(x));
793
794 if (!UseCountedLoopSafepoints) {
795 // Check for immediately preceding SafePoint and remove
796 Node *sfpt2 = le->in(0);
797 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
798 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
799 if (loop->_safepts != NULL) {
800 loop->_safepts->yank(sfpt2);
801 }
802 }
803 }
804
805 // Free up intermediate goo
806 _igvn.remove_dead_node(hook);
807
808 #ifdef ASSERT
809 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
810 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
811 #endif
812 #ifndef PRODUCT
813 if (TraceLoopOpts) {
814 tty->print("Counted ");
815 loop->dump_head();
816 }
817 #endif
818
819 C->print_method(PHASE_AFTER_CLOOPS, 3);
820
821 return true;
822 }
823
824 //----------------------exact_limit-------------------------------------------
825 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
826 assert(loop->_head->is_CountedLoop(), "");
827 CountedLoopNode *cl = loop->_head->as_CountedLoop();
828 assert(cl->is_valid_counted_loop(), "");
829
830 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
831 cl->limit()->Opcode() == Op_LoopLimit) {
832 // Old code has exact limit (it could be incorrect in case of int overflow).
833 // Loop limit is exact with stride == 1. And loop may already have exact limit.
834 return cl->limit();
835 }
836 Node *limit = NULL;
837 #ifdef ASSERT
838 BoolTest::mask bt = cl->loopexit()->test_trip();
839 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
840 #endif
841 if (cl->has_exact_trip_count()) {
842 // Simple case: loop has constant boundaries.
843 // Use jlongs to avoid integer overflow.
844 int stride_con = cl->stride_con();
845 jlong init_con = cl->init_trip()->get_int();
846 jlong limit_con = cl->limit()->get_int();
847 julong trip_cnt = cl->trip_count();
848 jlong final_con = init_con + trip_cnt*stride_con;
849 int final_int = (int)final_con;
850 // The final value should be in integer range since the loop
851 // is counted and the limit was checked for overflow.
852 assert(final_con == (jlong)final_int, "final value should be integer");
853 limit = _igvn.intcon(final_int);
854 } else {
855 // Create new LoopLimit node to get exact limit (final iv value).
856 limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
857 register_new_node(limit, cl->in(LoopNode::EntryControl));
858 }
859 assert(limit != NULL, "sanity");
860 return limit;
861 }
862
863 //------------------------------Ideal------------------------------------------
864 // Return a node which is more "ideal" than the current node.
865 // Attempt to convert into a counted-loop.
866 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
867 if (!can_be_counted_loop(phase)) {
868 phase->C->set_major_progress();
869 }
870 return RegionNode::Ideal(phase, can_reshape);
871 }
872
873
874 //=============================================================================
875 //------------------------------Ideal------------------------------------------
876 // Return a node which is more "ideal" than the current node.
877 // Attempt to convert into a counted-loop.
878 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
879 return RegionNode::Ideal(phase, can_reshape);
880 }
881
882 //------------------------------dump_spec--------------------------------------
883 // Dump special per-node info
884 #ifndef PRODUCT
885 void CountedLoopNode::dump_spec(outputStream *st) const {
886 LoopNode::dump_spec(st);
887 if (stride_is_con()) {
888 st->print("stride: %d ",stride_con());
889 }
890 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
891 if (is_main_loop()) st->print("main of N%d", _idx);
892 if (is_post_loop()) st->print("post of N%d", _main_idx);
893 }
894 #endif
895
896 //=============================================================================
897 int CountedLoopEndNode::stride_con() const {
898 return stride()->bottom_type()->is_int()->get_con();
899 }
900
901 //=============================================================================
902 //------------------------------Value-----------------------------------------
903 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
904 const Type* init_t = phase->type(in(Init));
905 const Type* limit_t = phase->type(in(Limit));
906 const Type* stride_t = phase->type(in(Stride));
907 // Either input is TOP ==> the result is TOP
908 if (init_t == Type::TOP) return Type::TOP;
909 if (limit_t == Type::TOP) return Type::TOP;
910 if (stride_t == Type::TOP) return Type::TOP;
911
912 int stride_con = stride_t->is_int()->get_con();
913 if (stride_con == 1)
914 return NULL; // Identity
915
916 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
917 // Use jlongs to avoid integer overflow.
918 jlong init_con = init_t->is_int()->get_con();
919 jlong limit_con = limit_t->is_int()->get_con();
920 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
921 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
922 jlong final_con = init_con + stride_con*trip_count;
923 int final_int = (int)final_con;
924 // The final value should be in integer range since the loop
925 // is counted and the limit was checked for overflow.
926 assert(final_con == (jlong)final_int, "final value should be integer");
927 return TypeInt::make(final_int);
928 }
929
930 return bottom_type(); // TypeInt::INT
931 }
932
933 //------------------------------Ideal------------------------------------------
934 // Return a node which is more "ideal" than the current node.
935 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
936 if (phase->type(in(Init)) == Type::TOP ||
937 phase->type(in(Limit)) == Type::TOP ||
938 phase->type(in(Stride)) == Type::TOP)
939 return NULL; // Dead
940
941 int stride_con = phase->type(in(Stride))->is_int()->get_con();
942 if (stride_con == 1)
943 return NULL; // Identity
944
945 if (in(Init)->is_Con() && in(Limit)->is_Con())
946 return NULL; // Value
947
948 // Delay following optimizations until all loop optimizations
949 // done to keep Ideal graph simple.
950 if (!can_reshape || phase->C->major_progress())
951 return NULL;
952
953 const TypeInt* init_t = phase->type(in(Init) )->is_int();
954 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
955 int stride_p;
956 jlong lim, ini;
957 julong max;
958 if (stride_con > 0) {
959 stride_p = stride_con;
960 lim = limit_t->_hi;
961 ini = init_t->_lo;
962 max = (julong)max_jint;
963 } else {
964 stride_p = -stride_con;
965 lim = init_t->_hi;
966 ini = limit_t->_lo;
967 max = (julong)min_jint;
968 }
969 julong range = lim - ini + stride_p;
970 if (range <= max) {
971 // Convert to integer expression if it is not overflow.
972 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
973 Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init)));
974 Node *bias = phase->transform(new (phase->C) AddINode(range, stride_m));
975 Node *trip = phase->transform(new (phase->C) DivINode(0, bias, in(Stride)));
976 Node *span = phase->transform(new (phase->C) MulINode(trip, in(Stride)));
977 return new (phase->C) AddINode(span, in(Init)); // exact limit
978 }
979
980 if (is_power_of_2(stride_p) || // divisor is 2^n
981 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
982 // Convert to long expression to avoid integer overflow
983 // and let igvn optimizer convert this division.
984 //
985 Node* init = phase->transform( new (phase->C) ConvI2LNode(in(Init)));
986 Node* limit = phase->transform( new (phase->C) ConvI2LNode(in(Limit)));
987 Node* stride = phase->longcon(stride_con);
988 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
989
990 Node *range = phase->transform(new (phase->C) SubLNode(limit, init));
991 Node *bias = phase->transform(new (phase->C) AddLNode(range, stride_m));
992 Node *span;
993 if (stride_con > 0 && is_power_of_2(stride_p)) {
994 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
995 // and avoid generating rounding for division. Zero trip guard should
996 // guarantee that init < limit but sometimes the guard is missing and
997 // we can get situation when init > limit. Note, for the empty loop
998 // optimization zero trip guard is generated explicitly which leaves
999 // only RCE predicate where exact limit is used and the predicate
1000 // will simply fail forcing recompilation.
1001 Node* neg_stride = phase->longcon(-stride_con);
1002 span = phase->transform(new (phase->C) AndLNode(bias, neg_stride));
1003 } else {
1004 Node *trip = phase->transform(new (phase->C) DivLNode(0, bias, stride));
1005 span = phase->transform(new (phase->C) MulLNode(trip, stride));
1006 }
1007 // Convert back to int
1008 Node *span_int = phase->transform(new (phase->C) ConvL2INode(span));
1009 return new (phase->C) AddINode(span_int, in(Init)); // exact limit
1010 }
1011
1012 return NULL; // No progress
1013 }
1014
1015 //------------------------------Identity---------------------------------------
1016 // If stride == 1 return limit node.
1017 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
1018 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1019 if (stride_con == 1 || stride_con == -1)
1020 return in(Limit);
1021 return this;
1022 }
1023
1024 //=============================================================================
1025 //----------------------match_incr_with_optional_truncation--------------------
1026 // Match increment with optional truncation:
1027 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1028 // Return NULL for failure. Success returns the increment node.
1029 Node* CountedLoopNode::match_incr_with_optional_truncation(
1030 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1031 // Quick cutouts:
1032 if (expr == NULL || expr->req() != 3) return NULL;
1033
1034 Node *t1 = NULL;
1035 Node *t2 = NULL;
1036 const TypeInt* trunc_t = TypeInt::INT;
1037 Node* n1 = expr;
1038 int n1op = n1->Opcode();
1039
1040 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1041 if (n1op == Op_AndI &&
1042 n1->in(2)->is_Con() &&
1043 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1044 // %%% This check should match any mask of 2**K-1.
1045 t1 = n1;
1046 n1 = t1->in(1);
1047 n1op = n1->Opcode();
1048 trunc_t = TypeInt::CHAR;
1049 } else if (n1op == Op_RShiftI &&
1050 n1->in(1) != NULL &&
1051 n1->in(1)->Opcode() == Op_LShiftI &&
1052 n1->in(2) == n1->in(1)->in(2) &&
1053 n1->in(2)->is_Con()) {
1054 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1055 // %%% This check should match any shift in [1..31].
1056 if (shift == 16 || shift == 8) {
1057 t1 = n1;
1058 t2 = t1->in(1);
1059 n1 = t2->in(1);
1060 n1op = n1->Opcode();
1061 if (shift == 16) {
1062 trunc_t = TypeInt::SHORT;
1063 } else if (shift == 8) {
1064 trunc_t = TypeInt::BYTE;
1065 }
1066 }
1067 }
1068
1069 // If (maybe after stripping) it is an AddI, we won:
1070 if (n1op == Op_AddI) {
1071 *trunc1 = t1;
1072 *trunc2 = t2;
1073 *trunc_type = trunc_t;
1074 return n1;
1075 }
1076
1077 // failed
1078 return NULL;
1079 }
1080
1081
1082 //------------------------------filtered_type--------------------------------
1083 // Return a type based on condition control flow
1084 // A successful return will be a type that is restricted due
1085 // to a series of dominating if-tests, such as:
1086 // if (i < 10) {
1087 // if (i > 0) {
1088 // here: "i" type is [1..10)
1089 // }
1090 // }
1091 // or a control flow merge
1092 // if (i < 10) {
1093 // do {
1094 // phi( , ) -- at top of loop type is [min_int..10)
1095 // i = ?
1096 // } while ( i < 10)
1097 //
1098 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1099 assert(n && n->bottom_type()->is_int(), "must be int");
1100 const TypeInt* filtered_t = NULL;
1101 if (!n->is_Phi()) {
1102 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1103 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1104
1105 } else {
1106 Node* phi = n->as_Phi();
1107 Node* region = phi->in(0);
1108 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1109 if (region && region != C->top()) {
1110 for (uint i = 1; i < phi->req(); i++) {
1111 Node* val = phi->in(i);
1112 Node* use_c = region->in(i);
1113 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1114 if (val_t != NULL) {
1115 if (filtered_t == NULL) {
1116 filtered_t = val_t;
1117 } else {
1118 filtered_t = filtered_t->meet(val_t)->is_int();
1119 }
1120 }
1121 }
1122 }
1123 }
1124 const TypeInt* n_t = _igvn.type(n)->is_int();
1125 if (filtered_t != NULL) {
1126 n_t = n_t->join(filtered_t)->is_int();
1127 }
1128 return n_t;
1129 }
1130
1131
1132 //------------------------------filtered_type_from_dominators--------------------------------
1133 // Return a possibly more restrictive type for val based on condition control flow of dominators
1134 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1135 if (val->is_Con()) {
1136 return val->bottom_type()->is_int();
1137 }
1138 uint if_limit = 10; // Max number of dominating if's visited
1139 const TypeInt* rtn_t = NULL;
1140
1141 if (use_ctrl && use_ctrl != C->top()) {
1142 Node* val_ctrl = get_ctrl(val);
1143 uint val_dom_depth = dom_depth(val_ctrl);
1144 Node* pred = use_ctrl;
1145 uint if_cnt = 0;
1146 while (if_cnt < if_limit) {
1147 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1148 if_cnt++;
1149 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1150 if (if_t != NULL) {
1151 if (rtn_t == NULL) {
1152 rtn_t = if_t;
1153 } else {
1154 rtn_t = rtn_t->join(if_t)->is_int();
1155 }
1156 }
1157 }
1158 pred = idom(pred);
1159 if (pred == NULL || pred == C->top()) {
1160 break;
1161 }
1162 // Stop if going beyond definition block of val
1163 if (dom_depth(pred) < val_dom_depth) {
1164 break;
1165 }
1166 }
1167 }
1168 return rtn_t;
1169 }
1170
1171
1172 //------------------------------dump_spec--------------------------------------
1173 // Dump special per-node info
1174 #ifndef PRODUCT
1175 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1176 if( in(TestValue)->is_Bool() ) {
1177 BoolTest bt( test_trip()); // Added this for g++.
1178
1179 st->print("[");
1180 bt.dump_on(st);
1181 st->print("]");
1182 }
1183 st->print(" ");
1184 IfNode::dump_spec(st);
1185 }
1186 #endif
1187
1188 //=============================================================================
1189 //------------------------------is_member--------------------------------------
1190 // Is 'l' a member of 'this'?
1191 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1192 while( l->_nest > _nest ) l = l->_parent;
1193 return l == this;
1194 }
1195
1196 //------------------------------set_nest---------------------------------------
1197 // Set loop tree nesting depth. Accumulate _has_call bits.
1198 int IdealLoopTree::set_nest( uint depth ) {
1199 _nest = depth;
1200 int bits = _has_call;
1201 if( _child ) bits |= _child->set_nest(depth+1);
1202 if( bits ) _has_call = 1;
1203 if( _next ) bits |= _next ->set_nest(depth );
1204 return bits;
1205 }
1206
1207 //------------------------------split_fall_in----------------------------------
1208 // Split out multiple fall-in edges from the loop header. Move them to a
1209 // private RegionNode before the loop. This becomes the loop landing pad.
1210 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1211 PhaseIterGVN &igvn = phase->_igvn;
1212 uint i;
1213
1214 // Make a new RegionNode to be the landing pad.
1215 Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 );
1216 phase->set_loop(landing_pad,_parent);
1217 // Gather all the fall-in control paths into the landing pad
1218 uint icnt = fall_in_cnt;
1219 uint oreq = _head->req();
1220 for( i = oreq-1; i>0; i-- )
1221 if( !phase->is_member( this, _head->in(i) ) )
1222 landing_pad->set_req(icnt--,_head->in(i));
1223
1224 // Peel off PhiNode edges as well
1225 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1226 Node *oj = _head->fast_out(j);
1227 if( oj->is_Phi() ) {
1228 PhiNode* old_phi = oj->as_Phi();
1229 assert( old_phi->region() == _head, "" );
1230 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1231 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1232 uint icnt = fall_in_cnt;
1233 for( i = oreq-1; i>0; i-- ) {
1234 if( !phase->is_member( this, _head->in(i) ) ) {
1235 p->init_req(icnt--, old_phi->in(i));
1236 // Go ahead and clean out old edges from old phi
1237 old_phi->del_req(i);
1238 }
1239 }
1240 // Search for CSE's here, because ZKM.jar does a lot of
1241 // loop hackery and we need to be a little incremental
1242 // with the CSE to avoid O(N^2) node blow-up.
1243 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1244 if( p2 ) { // Found CSE
1245 p->destruct(); // Recover useless new node
1246 p = p2; // Use old node
1247 } else {
1248 igvn.register_new_node_with_optimizer(p, old_phi);
1249 }
1250 // Make old Phi refer to new Phi.
1251 old_phi->add_req(p);
1252 // Check for the special case of making the old phi useless and
1253 // disappear it. In JavaGrande I have a case where this useless
1254 // Phi is the loop limit and prevents recognizing a CountedLoop
1255 // which in turn prevents removing an empty loop.
1256 Node *id_old_phi = old_phi->Identity( &igvn );
1257 if( id_old_phi != old_phi ) { // Found a simple identity?
1258 // Note that I cannot call 'replace_node' here, because
1259 // that will yank the edge from old_phi to the Region and
1260 // I'm mid-iteration over the Region's uses.
1261 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1262 Node* use = old_phi->last_out(i);
1263 igvn.rehash_node_delayed(use);
1264 uint uses_found = 0;
1265 for (uint j = 0; j < use->len(); j++) {
1266 if (use->in(j) == old_phi) {
1267 if (j < use->req()) use->set_req (j, id_old_phi);
1268 else use->set_prec(j, id_old_phi);
1269 uses_found++;
1270 }
1271 }
1272 i -= uses_found; // we deleted 1 or more copies of this edge
1273 }
1274 }
1275 igvn._worklist.push(old_phi);
1276 }
1277 }
1278 // Finally clean out the fall-in edges from the RegionNode
1279 for( i = oreq-1; i>0; i-- ) {
1280 if( !phase->is_member( this, _head->in(i) ) ) {
1281 _head->del_req(i);
1282 }
1283 }
1284 // Transform landing pad
1285 igvn.register_new_node_with_optimizer(landing_pad, _head);
1286 // Insert landing pad into the header
1287 _head->add_req(landing_pad);
1288 }
1289
1290 //------------------------------split_outer_loop-------------------------------
1291 // Split out the outermost loop from this shared header.
1292 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1293 PhaseIterGVN &igvn = phase->_igvn;
1294
1295 // Find index of outermost loop; it should also be my tail.
1296 uint outer_idx = 1;
1297 while( _head->in(outer_idx) != _tail ) outer_idx++;
1298
1299 // Make a LoopNode for the outermost loop.
1300 Node *ctl = _head->in(LoopNode::EntryControl);
1301 Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) );
1302 outer = igvn.register_new_node_with_optimizer(outer, _head);
1303 phase->set_created_loop_node();
1304
1305 // Outermost loop falls into '_head' loop
1306 _head->set_req(LoopNode::EntryControl, outer);
1307 _head->del_req(outer_idx);
1308 // Split all the Phis up between '_head' loop and 'outer' loop.
1309 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1310 Node *out = _head->fast_out(j);
1311 if( out->is_Phi() ) {
1312 PhiNode *old_phi = out->as_Phi();
1313 assert( old_phi->region() == _head, "" );
1314 Node *phi = PhiNode::make_blank(outer, old_phi);
1315 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1316 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1317 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1318 // Make old Phi point to new Phi on the fall-in path
1319 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1320 old_phi->del_req(outer_idx);
1321 }
1322 }
1323
1324 // Use the new loop head instead of the old shared one
1325 _head = outer;
1326 phase->set_loop(_head, this);
1327 }
1328
1329 //------------------------------fix_parent-------------------------------------
1330 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1331 loop->_parent = parent;
1332 if( loop->_child ) fix_parent( loop->_child, loop );
1333 if( loop->_next ) fix_parent( loop->_next , parent );
1334 }
1335
1336 //------------------------------estimate_path_freq-----------------------------
1337 static float estimate_path_freq( Node *n ) {
1338 // Try to extract some path frequency info
1339 IfNode *iff;
1340 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1341 uint nop = n->Opcode();
1342 if( nop == Op_SafePoint ) { // Skip any safepoint
1343 n = n->in(0);
1344 continue;
1345 }
1346 if( nop == Op_CatchProj ) { // Get count from a prior call
1347 // Assume call does not always throw exceptions: means the call-site
1348 // count is also the frequency of the fall-through path.
1349 assert( n->is_CatchProj(), "" );
1350 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1351 return 0.0f; // Assume call exception path is rare
1352 Node *call = n->in(0)->in(0)->in(0);
1353 assert( call->is_Call(), "expect a call here" );
1354 const JVMState *jvms = ((CallNode*)call)->jvms();
1355 ciMethodData* methodData = jvms->method()->method_data();
1356 if (!methodData->is_mature()) return 0.0f; // No call-site data
1357 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1358 if ((data == NULL) || !data->is_CounterData()) {
1359 // no call profile available, try call's control input
1360 n = n->in(0);
1361 continue;
1362 }
1363 return data->as_CounterData()->count()/FreqCountInvocations;
1364 }
1365 // See if there's a gating IF test
1366 Node *n_c = n->in(0);
1367 if( !n_c->is_If() ) break; // No estimate available
1368 iff = n_c->as_If();
1369 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1370 // Compute how much count comes on this path
1371 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1372 // Have no count info. Skip dull uncommon-trap like branches.
1373 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1374 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1375 break;
1376 // Skip through never-taken branch; look for a real loop exit.
1377 n = iff->in(0);
1378 }
1379 return 0.0f; // No estimate available
1380 }
1381
1382 //------------------------------merge_many_backedges---------------------------
1383 // Merge all the backedges from the shared header into a private Region.
1384 // Feed that region as the one backedge to this loop.
1385 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1386 uint i;
1387
1388 // Scan for the top 2 hottest backedges
1389 float hotcnt = 0.0f;
1390 float warmcnt = 0.0f;
1391 uint hot_idx = 0;
1392 // Loop starts at 2 because slot 1 is the fall-in path
1393 for( i = 2; i < _head->req(); i++ ) {
1394 float cnt = estimate_path_freq(_head->in(i));
1395 if( cnt > hotcnt ) { // Grab hottest path
1396 warmcnt = hotcnt;
1397 hotcnt = cnt;
1398 hot_idx = i;
1399 } else if( cnt > warmcnt ) { // And 2nd hottest path
1400 warmcnt = cnt;
1401 }
1402 }
1403
1404 // See if the hottest backedge is worthy of being an inner loop
1405 // by being much hotter than the next hottest backedge.
1406 if( hotcnt <= 0.0001 ||
1407 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1408
1409 // Peel out the backedges into a private merge point; peel
1410 // them all except optionally hot_idx.
1411 PhaseIterGVN &igvn = phase->_igvn;
1412
1413 Node *hot_tail = NULL;
1414 // Make a Region for the merge point
1415 Node *r = new (phase->C) RegionNode(1);
1416 for( i = 2; i < _head->req(); i++ ) {
1417 if( i != hot_idx )
1418 r->add_req( _head->in(i) );
1419 else hot_tail = _head->in(i);
1420 }
1421 igvn.register_new_node_with_optimizer(r, _head);
1422 // Plug region into end of loop _head, followed by hot_tail
1423 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1424 _head->set_req(2, r);
1425 if( hot_idx ) _head->add_req(hot_tail);
1426
1427 // Split all the Phis up between '_head' loop and the Region 'r'
1428 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1429 Node *out = _head->fast_out(j);
1430 if( out->is_Phi() ) {
1431 PhiNode* n = out->as_Phi();
1432 igvn.hash_delete(n); // Delete from hash before hacking edges
1433 Node *hot_phi = NULL;
1434 Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type());
1435 // Check all inputs for the ones to peel out
1436 uint j = 1;
1437 for( uint i = 2; i < n->req(); i++ ) {
1438 if( i != hot_idx )
1439 phi->set_req( j++, n->in(i) );
1440 else hot_phi = n->in(i);
1441 }
1442 // Register the phi but do not transform until whole place transforms
1443 igvn.register_new_node_with_optimizer(phi, n);
1444 // Add the merge phi to the old Phi
1445 while( n->req() > 3 ) n->del_req( n->req()-1 );
1446 n->set_req(2, phi);
1447 if( hot_idx ) n->add_req(hot_phi);
1448 }
1449 }
1450
1451
1452 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1453 // of self loop tree. Turn self into a loop headed by _head and with
1454 // tail being the new merge point.
1455 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1456 phase->set_loop(_tail,ilt); // Adjust tail
1457 _tail = r; // Self's tail is new merge point
1458 phase->set_loop(r,this);
1459 ilt->_child = _child; // New guy has my children
1460 _child = ilt; // Self has new guy as only child
1461 ilt->_parent = this; // new guy has self for parent
1462 ilt->_nest = _nest; // Same nesting depth (for now)
1463
1464 // Starting with 'ilt', look for child loop trees using the same shared
1465 // header. Flatten these out; they will no longer be loops in the end.
1466 IdealLoopTree **pilt = &_child;
1467 while( ilt ) {
1468 if( ilt->_head == _head ) {
1469 uint i;
1470 for( i = 2; i < _head->req(); i++ )
1471 if( _head->in(i) == ilt->_tail )
1472 break; // Still a loop
1473 if( i == _head->req() ) { // No longer a loop
1474 // Flatten ilt. Hang ilt's "_next" list from the end of
1475 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1476 IdealLoopTree **cp = &ilt->_child;
1477 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1478 *cp = ilt->_next; // Hang next list at end of child list
1479 *pilt = ilt->_child; // Move child up to replace ilt
1480 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1481 ilt = ilt->_child; // Repeat using new ilt
1482 continue; // do not advance over ilt->_child
1483 }
1484 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1485 phase->set_loop(_head,ilt);
1486 }
1487 pilt = &ilt->_child; // Advance to next
1488 ilt = *pilt;
1489 }
1490
1491 if( _child ) fix_parent( _child, this );
1492 }
1493
1494 //------------------------------beautify_loops---------------------------------
1495 // Split shared headers and insert loop landing pads.
1496 // Insert a LoopNode to replace the RegionNode.
1497 // Return TRUE if loop tree is structurally changed.
1498 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1499 bool result = false;
1500 // Cache parts in locals for easy
1501 PhaseIterGVN &igvn = phase->_igvn;
1502
1503 igvn.hash_delete(_head); // Yank from hash before hacking edges
1504
1505 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1506 int fall_in_cnt = 0;
1507 for( uint i = 1; i < _head->req(); i++ )
1508 if( !phase->is_member( this, _head->in(i) ) )
1509 fall_in_cnt++;
1510 assert( fall_in_cnt, "at least 1 fall-in path" );
1511 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1512 split_fall_in( phase, fall_in_cnt );
1513
1514 // Swap inputs to the _head and all Phis to move the fall-in edge to
1515 // the left.
1516 fall_in_cnt = 1;
1517 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1518 fall_in_cnt++;
1519 if( fall_in_cnt > 1 ) {
1520 // Since I am just swapping inputs I do not need to update def-use info
1521 Node *tmp = _head->in(1);
1522 _head->set_req( 1, _head->in(fall_in_cnt) );
1523 _head->set_req( fall_in_cnt, tmp );
1524 // Swap also all Phis
1525 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1526 Node* phi = _head->fast_out(i);
1527 if( phi->is_Phi() ) {
1528 igvn.hash_delete(phi); // Yank from hash before hacking edges
1529 tmp = phi->in(1);
1530 phi->set_req( 1, phi->in(fall_in_cnt) );
1531 phi->set_req( fall_in_cnt, tmp );
1532 }
1533 }
1534 }
1535 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1536 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1537
1538 // If I am a shared header (multiple backedges), peel off the many
1539 // backedges into a private merge point and use the merge point as
1540 // the one true backedge.
1541 if (_head->req() > 3) {
1542 // Merge the many backedges into a single backedge but leave
1543 // the hottest backedge as separate edge for the following peel.
1544 if (!_irreducible) {
1545 merge_many_backedges( phase );
1546 }
1547
1548 // When recursively beautify my children, split_fall_in can change
1549 // loop tree structure when I am an irreducible loop. Then the head
1550 // of my children has a req() not bigger than 3. Here we need to set
1551 // result to true to catch that case in order to tell the caller to
1552 // rebuild loop tree. See issue JDK-8244407 for details.
1553 result = true;
1554 }
1555
1556 // If I have one hot backedge, peel off myself loop.
1557 // I better be the outermost loop.
1558 if (_head->req() > 3 && !_irreducible) {
1559 split_outer_loop( phase );
1560 result = true;
1561
1562 } else if (!_head->is_Loop() && !_irreducible) {
1563 // Make a new LoopNode to replace the old loop head
1564 Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) );
1565 l = igvn.register_new_node_with_optimizer(l, _head);
1566 phase->set_created_loop_node();
1567 // Go ahead and replace _head
1568 phase->_igvn.replace_node( _head, l );
1569 _head = l;
1570 phase->set_loop(_head, this);
1571 }
1572
1573 // Now recursively beautify nested loops
1574 if( _child ) result |= _child->beautify_loops( phase );
1575 if( _next ) result |= _next ->beautify_loops( phase );
1576 return result;
1577 }
1578
1579 //------------------------------allpaths_check_safepts----------------------------
1580 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1581 // encountered. Helper for check_safepts.
1582 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1583 assert(stack.size() == 0, "empty stack");
1584 stack.push(_tail);
1585 visited.Clear();
1586 visited.set(_tail->_idx);
1587 while (stack.size() > 0) {
1588 Node* n = stack.pop();
1589 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1590 // Terminate this path
1591 } else if (n->Opcode() == Op_SafePoint) {
1592 if (_phase->get_loop(n) != this) {
1593 if (_required_safept == NULL) _required_safept = new Node_List();
1594 _required_safept->push(n); // save the one closest to the tail
1595 }
1596 // Terminate this path
1597 } else {
1598 uint start = n->is_Region() ? 1 : 0;
1599 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1600 for (uint i = start; i < end; i++) {
1601 Node* in = n->in(i);
1602 assert(in->is_CFG(), "must be");
1603 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1604 stack.push(in);
1605 }
1606 }
1607 }
1608 }
1609 }
1610
1611 //------------------------------check_safepts----------------------------
1612 // Given dominators, try to find loops with calls that must always be
1613 // executed (call dominates loop tail). These loops do not need non-call
1614 // safepoints (ncsfpt).
1615 //
1616 // A complication is that a safepoint in a inner loop may be needed
1617 // by an outer loop. In the following, the inner loop sees it has a
1618 // call (block 3) on every path from the head (block 2) to the
1619 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1620 // in block 2, _but_ this leaves the outer loop without a safepoint.
1621 //
1622 // entry 0
1623 // |
1624 // v
1625 // outer 1,2 +->1
1626 // | |
1627 // | v
1628 // | 2<---+ ncsfpt in 2
1629 // |_/|\ |
1630 // | v |
1631 // inner 2,3 / 3 | call in 3
1632 // / | |
1633 // v +--+
1634 // exit 4
1635 //
1636 //
1637 // This method creates a list (_required_safept) of ncsfpt nodes that must
1638 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1639 // is first looked for in the lists for the outer loops of the current loop.
1640 //
1641 // The insights into the problem:
1642 // A) counted loops are okay
1643 // B) innermost loops are okay (only an inner loop can delete
1644 // a ncsfpt needed by an outer loop)
1645 // C) a loop is immune from an inner loop deleting a safepoint
1646 // if the loop has a call on the idom-path
1647 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1648 // idom-path that is not in a nested loop
1649 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1650 // loop needs to be prevented from deletion by an inner loop
1651 //
1652 // There are two analyses:
1653 // 1) The first, and cheaper one, scans the loop body from
1654 // tail to head following the idom (immediate dominator)
1655 // chain, looking for the cases (C,D,E) above.
1656 // Since inner loops are scanned before outer loops, there is summary
1657 // information about inner loops. Inner loops can be skipped over
1658 // when the tail of an inner loop is encountered.
1659 //
1660 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1661 // the idom path (which is rare), scans all predecessor control paths
1662 // from the tail to the head, terminating a path when a call or sfpt
1663 // is encountered, to find the ncsfpt's that are closest to the tail.
1664 //
1665 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1666 // Bottom up traversal
1667 IdealLoopTree* ch = _child;
1668 if (_child) _child->check_safepts(visited, stack);
1669 if (_next) _next ->check_safepts(visited, stack);
1670
1671 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1672 bool has_call = false; // call on dom-path
1673 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1674 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1675 // Scan the dom-path nodes from tail to head
1676 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1677 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1678 has_call = true;
1679 _has_sfpt = 1; // Then no need for a safept!
1680 break;
1681 } else if (n->Opcode() == Op_SafePoint) {
1682 if (_phase->get_loop(n) == this) {
1683 has_local_ncsfpt = true;
1684 break;
1685 }
1686 if (nonlocal_ncsfpt == NULL) {
1687 nonlocal_ncsfpt = n; // save the one closest to the tail
1688 }
1689 } else {
1690 IdealLoopTree* nlpt = _phase->get_loop(n);
1691 if (this != nlpt) {
1692 // If at an inner loop tail, see if the inner loop has already
1693 // recorded seeing a call on the dom-path (and stop.) If not,
1694 // jump to the head of the inner loop.
1695 assert(is_member(nlpt), "nested loop");
1696 Node* tail = nlpt->_tail;
1697 if (tail->in(0)->is_If()) tail = tail->in(0);
1698 if (n == tail) {
1699 // If inner loop has call on dom-path, so does outer loop
1700 if (nlpt->_has_sfpt) {
1701 has_call = true;
1702 _has_sfpt = 1;
1703 break;
1704 }
1705 // Skip to head of inner loop
1706 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1707 n = nlpt->_head;
1708 }
1709 }
1710 }
1711 }
1712 // Record safept's that this loop needs preserved when an
1713 // inner loop attempts to delete it's safepoints.
1714 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1715 if (nonlocal_ncsfpt != NULL) {
1716 if (_required_safept == NULL) _required_safept = new Node_List();
1717 _required_safept->push(nonlocal_ncsfpt);
1718 } else {
1719 // Failed to find a suitable safept on the dom-path. Now use
1720 // an all paths walk from tail to head, looking for safepoints to preserve.
1721 allpaths_check_safepts(visited, stack);
1722 }
1723 }
1724 }
1725 }
1726
1727 //---------------------------is_deleteable_safept----------------------------
1728 // Is safept not required by an outer loop?
1729 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1730 assert(sfpt->Opcode() == Op_SafePoint, "");
1731 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1732 while (lp != NULL) {
1733 Node_List* sfpts = lp->_required_safept;
1734 if (sfpts != NULL) {
1735 for (uint i = 0; i < sfpts->size(); i++) {
1736 if (sfpt == sfpts->at(i))
1737 return false;
1738 }
1739 }
1740 lp = lp->_parent;
1741 }
1742 return true;
1743 }
1744
1745 //---------------------------replace_parallel_iv-------------------------------
1746 // Replace parallel induction variable (parallel to trip counter)
1747 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1748 assert(loop->_head->is_CountedLoop(), "");
1749 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1750 if (!cl->is_valid_counted_loop())
1751 return; // skip malformed counted loop
1752 Node *incr = cl->incr();
1753 if (incr == NULL)
1754 return; // Dead loop?
1755 Node *init = cl->init_trip();
1756 Node *phi = cl->phi();
1757 int stride_con = cl->stride_con();
1758
1759 // Visit all children, looking for Phis
1760 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1761 Node *out = cl->out(i);
1762 // Look for other phis (secondary IVs). Skip dead ones
1763 if (!out->is_Phi() || out == phi || !has_node(out))
1764 continue;
1765 PhiNode* phi2 = out->as_Phi();
1766 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1767 // Look for induction variables of the form: X += constant
1768 if (phi2->region() != loop->_head ||
1769 incr2->req() != 3 ||
1770 incr2->in(1) != phi2 ||
1771 incr2 == incr ||
1772 incr2->Opcode() != Op_AddI ||
1773 !incr2->in(2)->is_Con())
1774 continue;
1775
1776 // Check for parallel induction variable (parallel to trip counter)
1777 // via an affine function. In particular, count-down loops with
1778 // count-up array indices are common. We only RCE references off
1779 // the trip-counter, so we need to convert all these to trip-counter
1780 // expressions.
1781 Node *init2 = phi2->in( LoopNode::EntryControl );
1782 int stride_con2 = incr2->in(2)->get_int();
1783
1784 // The ratio of the two strides cannot be represented as an int
1785 // if stride_con2 is min_int and stride_con is -1.
1786 if (stride_con2 == min_jint && stride_con == -1) {
1787 continue;
1788 }
1789
1790 // The general case here gets a little tricky. We want to find the
1791 // GCD of all possible parallel IV's and make a new IV using this
1792 // GCD for the loop. Then all possible IVs are simple multiples of
1793 // the GCD. In practice, this will cover very few extra loops.
1794 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1795 // where +/-1 is the common case, but other integer multiples are
1796 // also easy to handle.
1797 int ratio_con = stride_con2/stride_con;
1798
1799 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1800 #ifndef PRODUCT
1801 if (TraceLoopOpts) {
1802 tty->print("Parallel IV: %d ", phi2->_idx);
1803 loop->dump_head();
1804 }
1805 #endif
1806 // Convert to using the trip counter. The parallel induction
1807 // variable differs from the trip counter by a loop-invariant
1808 // amount, the difference between their respective initial values.
1809 // It is scaled by the 'ratio_con'.
1810 Node* ratio = _igvn.intcon(ratio_con);
1811 set_ctrl(ratio, C->root());
1812 Node* ratio_init = new (C) MulINode(init, ratio);
1813 _igvn.register_new_node_with_optimizer(ratio_init, init);
1814 set_early_ctrl(ratio_init);
1815 Node* diff = new (C) SubINode(init2, ratio_init);
1816 _igvn.register_new_node_with_optimizer(diff, init2);
1817 set_early_ctrl(diff);
1818 Node* ratio_idx = new (C) MulINode(phi, ratio);
1819 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1820 set_ctrl(ratio_idx, cl);
1821 Node* add = new (C) AddINode(ratio_idx, diff);
1822 _igvn.register_new_node_with_optimizer(add);
1823 set_ctrl(add, cl);
1824 _igvn.replace_node( phi2, add );
1825 // Sometimes an induction variable is unused
1826 if (add->outcnt() == 0) {
1827 _igvn.remove_dead_node(add);
1828 }
1829 --i; // deleted this phi; rescan starting with next position
1830 continue;
1831 }
1832 }
1833 }
1834
1835 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
1836 Node* keep = NULL;
1837 if (keep_one) {
1838 // Look for a safepoint on the idom-path.
1839 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
1840 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
1841 keep = i;
1842 break; // Found one
1843 }
1844 }
1845 }
1846
1847 // Don't remove any safepoints if it is requested to keep a single safepoint and
1848 // no safepoint was found on idom-path. It is not safe to remove any safepoint
1849 // in this case since there's no safepoint dominating all paths in the loop body.
1850 bool prune = !keep_one || keep != NULL;
1851
1852 // Delete other safepoints in this loop.
1853 Node_List* sfpts = _safepts;
1854 if (prune && sfpts != NULL) {
1855 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
1856 for (uint i = 0; i < sfpts->size(); i++) {
1857 Node* n = sfpts->at(i);
1858 assert(phase->get_loop(n) == this, "");
1859 if (n != keep && phase->is_deleteable_safept(n)) {
1860 phase->lazy_replace(n, n->in(TypeFunc::Control));
1861 }
1862 }
1863 }
1864 }
1865
1866 //------------------------------counted_loop-----------------------------------
1867 // Convert to counted loops where possible
1868 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1869
1870 // For grins, set the inner-loop flag here
1871 if (!_child) {
1872 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1873 }
1874
1875 if (_head->is_CountedLoop() ||
1876 phase->is_counted_loop(_head, this)) {
1877
1878 if (!UseCountedLoopSafepoints) {
1879 // Indicate we do not need a safepoint here
1880 _has_sfpt = 1;
1881 }
1882
1883 // Remove safepoints
1884 bool keep_one_sfpt = !(_has_call || _has_sfpt);
1885 remove_safepoints(phase, keep_one_sfpt);
1886
1887 // Look for induction variables
1888 phase->replace_parallel_iv(this);
1889
1890 } else if (_parent != NULL && !_irreducible) {
1891 // Not a counted loop. Keep one safepoint.
1892 bool keep_one_sfpt = true;
1893 remove_safepoints(phase, keep_one_sfpt);
1894 }
1895
1896 // Recursively
1897 if (_child) _child->counted_loop( phase );
1898 if (_next) _next ->counted_loop( phase );
1899 }
1900
1901 #ifndef PRODUCT
1902 //------------------------------dump_head--------------------------------------
1903 // Dump 1 liner for loop header info
1904 void IdealLoopTree::dump_head( ) const {
1905 for (uint i=0; i<_nest; i++)
1906 tty->print(" ");
1907 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1908 if (_irreducible) tty->print(" IRREDUCIBLE");
1909 Node* entry = _head->in(LoopNode::EntryControl);
1910 if (LoopLimitCheck) {
1911 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1912 if (predicate != NULL ) {
1913 tty->print(" limit_check");
1914 entry = entry->in(0)->in(0);
1915 }
1916 }
1917 if (UseLoopPredicate) {
1918 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1919 if (entry != NULL) {
1920 tty->print(" predicated");
1921 }
1922 }
1923 if (_head->is_CountedLoop()) {
1924 CountedLoopNode *cl = _head->as_CountedLoop();
1925 tty->print(" counted");
1926
1927 Node* init_n = cl->init_trip();
1928 if (init_n != NULL && init_n->is_Con())
1929 tty->print(" [%d,", cl->init_trip()->get_int());
1930 else
1931 tty->print(" [int,");
1932 Node* limit_n = cl->limit();
1933 if (limit_n != NULL && limit_n->is_Con())
1934 tty->print("%d),", cl->limit()->get_int());
1935 else
1936 tty->print("int),");
1937 int stride_con = cl->stride_con();
1938 if (stride_con > 0) tty->print("+");
1939 tty->print("%d", stride_con);
1940
1941 tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
1942
1943 if (cl->is_pre_loop ()) tty->print(" pre" );
1944 if (cl->is_main_loop()) tty->print(" main");
1945 if (cl->is_post_loop()) tty->print(" post");
1946 }
1947 if (_has_call) tty->print(" has_call");
1948 if (_has_sfpt) tty->print(" has_sfpt");
1949 if (_rce_candidate) tty->print(" rce");
1950 if (_safepts != NULL && _safepts->size() > 0) {
1951 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
1952 }
1953 if (_required_safept != NULL && _required_safept->size() > 0) {
1954 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
1955 }
1956 tty->cr();
1957 }
1958
1959 //------------------------------dump-------------------------------------------
1960 // Dump loops by loop tree
1961 void IdealLoopTree::dump( ) const {
1962 dump_head();
1963 if (_child) _child->dump();
1964 if (_next) _next ->dump();
1965 }
1966
1967 #endif
1968
1969 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1970 if (loop == root) {
1971 if (loop->_child != NULL) {
1972 log->begin_head("loop_tree");
1973 log->end_head();
1974 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1975 log->tail("loop_tree");
1976 assert(loop->_next == NULL, "what?");
1977 }
1978 } else {
1979 Node* head = loop->_head;
1980 log->begin_head("loop");
1981 log->print(" idx='%d' ", head->_idx);
1982 if (loop->_irreducible) log->print("irreducible='1' ");
1983 if (head->is_Loop()) {
1984 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1985 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1986 }
1987 if (head->is_CountedLoop()) {
1988 CountedLoopNode* cl = head->as_CountedLoop();
1989 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1990 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1991 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1992 }
1993 log->end_head();
1994 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1995 log->tail("loop");
1996 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1997 }
1998 }
1999
2000 //---------------------collect_potentially_useful_predicates-----------------------
2001 // Helper function to collect potentially useful predicates to prevent them from
2002 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2003 void PhaseIdealLoop::collect_potentially_useful_predicates(
2004 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2005 if (loop->_child) { // child
2006 collect_potentially_useful_predicates(loop->_child, useful_predicates);
2007 }
2008
2009 // self (only loops that we can apply loop predication may use their predicates)
2010 if (loop->_head->is_Loop() &&
2011 !loop->_irreducible &&
2012 !loop->tail()->is_top()) {
2013 LoopNode* lpn = loop->_head->as_Loop();
2014 Node* entry = lpn->in(LoopNode::EntryControl);
2015 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2016 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2017 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2018 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2019 entry = entry->in(0)->in(0);
2020 }
2021 predicate_proj = find_predicate(entry); // Predicate
2022 if (predicate_proj != NULL ) {
2023 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2024 }
2025 }
2026
2027 if (loop->_next) { // sibling
2028 collect_potentially_useful_predicates(loop->_next, useful_predicates);
2029 }
2030 }
2031
2032 //------------------------eliminate_useless_predicates-----------------------------
2033 // Eliminate all inserted predicates if they could not be used by loop predication.
2034 // Note: it will also eliminates loop limits check predicate since it also uses
2035 // Opaque1 node (see Parse::add_predicate()).
2036 void PhaseIdealLoop::eliminate_useless_predicates() {
2037 if (C->predicate_count() == 0)
2038 return; // no predicate left
2039
2040 Unique_Node_List useful_predicates; // to store useful predicates
2041 if (C->has_loops()) {
2042 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2043 }
2044
2045 for (int i = C->predicate_count(); i > 0; i--) {
2046 Node * n = C->predicate_opaque1_node(i-1);
2047 assert(n->Opcode() == Op_Opaque1, "must be");
2048 if (!useful_predicates.member(n)) { // not in the useful list
2049 _igvn.replace_node(n, n->in(1));
2050 }
2051 }
2052 }
2053
2054 //------------------------process_expensive_nodes-----------------------------
2055 // Expensive nodes have their control input set to prevent the GVN
2056 // from commoning them and as a result forcing the resulting node to
2057 // be in a more frequent path. Use CFG information here, to change the
2058 // control inputs so that some expensive nodes can be commoned while
2059 // not executed more frequently.
2060 bool PhaseIdealLoop::process_expensive_nodes() {
2061 assert(OptimizeExpensiveOps, "optimization off?");
2062
2063 // Sort nodes to bring similar nodes together
2064 C->sort_expensive_nodes();
2065
2066 bool progress = false;
2067
2068 for (int i = 0; i < C->expensive_count(); ) {
2069 Node* n = C->expensive_node(i);
2070 int start = i;
2071 // Find nodes similar to n
2072 i++;
2073 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2074 int end = i;
2075 // And compare them two by two
2076 for (int j = start; j < end; j++) {
2077 Node* n1 = C->expensive_node(j);
2078 if (is_node_unreachable(n1)) {
2079 continue;
2080 }
2081 for (int k = j+1; k < end; k++) {
2082 Node* n2 = C->expensive_node(k);
2083 if (is_node_unreachable(n2)) {
2084 continue;
2085 }
2086
2087 assert(n1 != n2, "should be pair of nodes");
2088
2089 Node* c1 = n1->in(0);
2090 Node* c2 = n2->in(0);
2091
2092 Node* parent_c1 = c1;
2093 Node* parent_c2 = c2;
2094
2095 // The call to get_early_ctrl_for_expensive() moves the
2096 // expensive nodes up but stops at loops that are in a if
2097 // branch. See whether we can exit the loop and move above the
2098 // If.
2099 if (c1->is_Loop()) {
2100 parent_c1 = c1->in(1);
2101 }
2102 if (c2->is_Loop()) {
2103 parent_c2 = c2->in(1);
2104 }
2105
2106 if (parent_c1 == parent_c2) {
2107 _igvn._worklist.push(n1);
2108 _igvn._worklist.push(n2);
2109 continue;
2110 }
2111
2112 // Look for identical expensive node up the dominator chain.
2113 if (is_dominator(c1, c2)) {
2114 c2 = c1;
2115 } else if (is_dominator(c2, c1)) {
2116 c1 = c2;
2117 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2118 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2119 // Both branches have the same expensive node so move it up
2120 // before the if.
2121 c1 = c2 = idom(parent_c1->in(0));
2122 }
2123 // Do the actual moves
2124 if (n1->in(0) != c1) {
2125 _igvn.hash_delete(n1);
2126 n1->set_req(0, c1);
2127 _igvn.hash_insert(n1);
2128 _igvn._worklist.push(n1);
2129 progress = true;
2130 }
2131 if (n2->in(0) != c2) {
2132 _igvn.hash_delete(n2);
2133 n2->set_req(0, c2);
2134 _igvn.hash_insert(n2);
2135 _igvn._worklist.push(n2);
2136 progress = true;
2137 }
2138 }
2139 }
2140 }
2141
2142 return progress;
2143 }
2144
2145
2146 //=============================================================================
2147 //----------------------------build_and_optimize-------------------------------
2148 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2149 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2150 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
2151 ResourceMark rm;
2152
2153 int old_progress = C->major_progress();
2154 uint orig_worklist_size = _igvn._worklist.size();
2155
2156 // Reset major-progress flag for the driver's heuristics
2157 C->clear_major_progress();
2158
2159 #ifndef PRODUCT
2160 // Capture for later assert
2161 uint unique = C->unique();
2162 _loop_invokes++;
2163 _loop_work += unique;
2164 #endif
2165
2166 // True if the method has at least 1 irreducible loop
2167 _has_irreducible_loops = false;
2168
2169 _created_loop_node = false;
2170
2171 Arena *a = Thread::current()->resource_area();
2172 VectorSet visited(a);
2173 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2174 _nodes.map(C->unique(), NULL);
2175 memset(_nodes.adr(), 0, wordSize * C->unique());
2176
2177 // Pre-build the top-level outermost loop tree entry
2178 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2179 // Do not need a safepoint at the top level
2180 _ltree_root->_has_sfpt = 1;
2181
2182 // Initialize Dominators.
2183 // Checked in clone_loop_predicate() during beautify_loops().
2184 _idom_size = 0;
2185 _idom = NULL;
2186 _dom_depth = NULL;
2187 _dom_stk = NULL;
2188
2189 // Empty pre-order array
2190 allocate_preorders();
2191
2192 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2193 // IdealLoopTree entries. Data nodes are NOT walked.
2194 build_loop_tree();
2195 // Check for bailout, and return
2196 if (C->failing()) {
2197 return;
2198 }
2199
2200 // No loops after all
2201 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2202
2203 // There should always be an outer loop containing the Root and Return nodes.
2204 // If not, we have a degenerate empty program. Bail out in this case.
2205 if (!has_node(C->root())) {
2206 if (!_verify_only) {
2207 C->clear_major_progress();
2208 C->record_method_not_compilable("empty program detected during loop optimization");
2209 }
2210 return;
2211 }
2212
2213 // Nothing to do, so get out
2214 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
2215 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2216 if (stop_early && !do_expensive_nodes) {
2217 _igvn.optimize(); // Cleanup NeverBranches
2218 return;
2219 }
2220
2221 // Set loop nesting depth
2222 _ltree_root->set_nest( 0 );
2223
2224 // Split shared headers and insert loop landing pads.
2225 // Do not bother doing this on the Root loop of course.
2226 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2227 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2228 if( _ltree_root->_child->beautify_loops( this ) ) {
2229 // Re-build loop tree!
2230 _ltree_root->_child = NULL;
2231 _nodes.clear();
2232 reallocate_preorders();
2233 build_loop_tree();
2234 // Check for bailout, and return
2235 if (C->failing()) {
2236 return;
2237 }
2238 // Reset loop nesting depth
2239 _ltree_root->set_nest( 0 );
2240
2241 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2242 }
2243 }
2244
2245 // Build Dominators for elision of NULL checks & loop finding.
2246 // Since nodes do not have a slot for immediate dominator, make
2247 // a persistent side array for that info indexed on node->_idx.
2248 _idom_size = C->unique();
2249 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2250 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2251 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2252 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2253
2254 Dominators();
2255
2256 if (!_verify_only) {
2257 // As a side effect, Dominators removed any unreachable CFG paths
2258 // into RegionNodes. It doesn't do this test against Root, so
2259 // we do it here.
2260 for( uint i = 1; i < C->root()->req(); i++ ) {
2261 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2262 _igvn.delete_input_of(C->root(), i);
2263 i--; // Rerun same iteration on compressed edges
2264 }
2265 }
2266
2267 // Given dominators, try to find inner loops with calls that must
2268 // always be executed (call dominates loop tail). These loops do
2269 // not need a separate safepoint.
2270 Node_List cisstack(a);
2271 _ltree_root->check_safepts(visited, cisstack);
2272 }
2273
2274 // Walk the DATA nodes and place into loops. Find earliest control
2275 // node. For CFG nodes, the _nodes array starts out and remains
2276 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2277 // _nodes array holds the earliest legal controlling CFG node.
2278
2279 // Allocate stack with enough space to avoid frequent realloc
2280 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2281 Node_Stack nstack( a, stack_size );
2282
2283 visited.Clear();
2284 Node_List worklist(a);
2285 // Don't need C->root() on worklist since
2286 // it will be processed among C->top() inputs
2287 worklist.push( C->top() );
2288 visited.set( C->top()->_idx ); // Set C->top() as visited now
2289 build_loop_early( visited, worklist, nstack );
2290
2291 // Given early legal placement, try finding counted loops. This placement
2292 // is good enough to discover most loop invariants.
2293 if( !_verify_me && !_verify_only )
2294 _ltree_root->counted_loop( this );
2295
2296 // Find latest loop placement. Find ideal loop placement.
2297 visited.Clear();
2298 init_dom_lca_tags();
2299 // Need C->root() on worklist when processing outs
2300 worklist.push( C->root() );
2301 NOT_PRODUCT( C->verify_graph_edges(); )
2302 worklist.push( C->top() );
2303 build_loop_late( visited, worklist, nstack );
2304
2305 if (_verify_only) {
2306 // restore major progress flag
2307 for (int i = 0; i < old_progress; i++)
2308 C->set_major_progress();
2309 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2310 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2311 return;
2312 }
2313
2314 // clear out the dead code after build_loop_late
2315 while (_deadlist.size()) {
2316 _igvn.remove_globally_dead_node(_deadlist.pop());
2317 }
2318
2319 if (stop_early) {
2320 assert(do_expensive_nodes, "why are we here?");
2321 if (process_expensive_nodes()) {
2322 // If we made some progress when processing expensive nodes then
2323 // the IGVN may modify the graph in a way that will allow us to
2324 // make some more progress: we need to try processing expensive
2325 // nodes again.
2326 C->set_major_progress();
2327 }
2328 _igvn.optimize();
2329 return;
2330 }
2331
2332 // Some parser-inserted loop predicates could never be used by loop
2333 // predication or they were moved away from loop during some optimizations.
2334 // For example, peeling. Eliminate them before next loop optimizations.
2335 if (UseLoopPredicate || LoopLimitCheck) {
2336 eliminate_useless_predicates();
2337 }
2338
2339 #ifndef PRODUCT
2340 C->verify_graph_edges();
2341 if (_verify_me) { // Nested verify pass?
2342 // Check to see if the verify mode is broken
2343 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2344 return;
2345 }
2346 if(VerifyLoopOptimizations) verify();
2347 if(TraceLoopOpts && C->has_loops()) {
2348 _ltree_root->dump();
2349 }
2350 #endif
2351
2352 if (skip_loop_opts) {
2353 // restore major progress flag
2354 for (int i = 0; i < old_progress; i++) {
2355 C->set_major_progress();
2356 }
2357
2358 // Cleanup any modified bits
2359 _igvn.optimize();
2360
2361 if (C->log() != NULL) {
2362 log_loop_tree(_ltree_root, _ltree_root, C->log());
2363 }
2364 return;
2365 }
2366
2367 if (ReassociateInvariants) {
2368 // Reassociate invariants and prep for split_thru_phi
2369 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2370 IdealLoopTree* lpt = iter.current();
2371 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2372
2373 lpt->reassociate_invariants(this);
2374
2375 // Because RCE opportunities can be masked by split_thru_phi,
2376 // look for RCE candidates and inhibit split_thru_phi
2377 // on just their loop-phi's for this pass of loop opts
2378 if (SplitIfBlocks && do_split_ifs) {
2379 if (lpt->policy_range_check(this)) {
2380 lpt->_rce_candidate = 1; // = true
2381 }
2382 }
2383 }
2384 }
2385
2386 // Check for aggressive application of split-if and other transforms
2387 // that require basic-block info (like cloning through Phi's)
2388 if( SplitIfBlocks && do_split_ifs ) {
2389 visited.Clear();
2390 split_if_with_blocks( visited, nstack );
2391 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2392 }
2393
2394 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2395 C->set_major_progress();
2396 }
2397
2398 // Perform loop predication before iteration splitting
2399 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2400 _ltree_root->_child->loop_predication(this);
2401 }
2402
2403 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2404 if (do_intrinsify_fill()) {
2405 C->set_major_progress();
2406 }
2407 }
2408
2409 // Perform iteration-splitting on inner loops. Split iterations to avoid
2410 // range checks or one-shot null checks.
2411
2412 // If split-if's didn't hack the graph too bad (no CFG changes)
2413 // then do loop opts.
2414 if (C->has_loops() && !C->major_progress()) {
2415 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2416 _ltree_root->_child->iteration_split( this, worklist );
2417 // No verify after peeling! GCM has hoisted code out of the loop.
2418 // After peeling, the hoisted code could sink inside the peeled area.
2419 // The peeling code does not try to recompute the best location for
2420 // all the code before the peeled area, so the verify pass will always
2421 // complain about it.
2422 }
2423 // Do verify graph edges in any case
2424 NOT_PRODUCT( C->verify_graph_edges(); );
2425
2426 if (!do_split_ifs) {
2427 // We saw major progress in Split-If to get here. We forced a
2428 // pass with unrolling and not split-if, however more split-if's
2429 // might make progress. If the unrolling didn't make progress
2430 // then the major-progress flag got cleared and we won't try
2431 // another round of Split-If. In particular the ever-common
2432 // instance-of/check-cast pattern requires at least 2 rounds of
2433 // Split-If to clear out.
2434 C->set_major_progress();
2435 }
2436
2437 // Repeat loop optimizations if new loops were seen
2438 if (created_loop_node()) {
2439 C->set_major_progress();
2440 }
2441
2442 // Keep loop predicates and perform optimizations with them
2443 // until no more loop optimizations could be done.
2444 // After that switch predicates off and do more loop optimizations.
2445 if (!C->major_progress() && (C->predicate_count() > 0)) {
2446 C->cleanup_loop_predicates(_igvn);
2447 #ifndef PRODUCT
2448 if (TraceLoopOpts) {
2449 tty->print_cr("PredicatesOff");
2450 }
2451 #endif
2452 C->set_major_progress();
2453 }
2454
2455 // Convert scalar to superword operations at the end of all loop opts.
2456 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2457 // SuperWord transform
2458 SuperWord sw(this);
2459 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2460 IdealLoopTree* lpt = iter.current();
2461 if (lpt->is_counted()) {
2462 sw.transform_loop(lpt);
2463 }
2464 }
2465 }
2466
2467 // Cleanup any modified bits
2468 _igvn.optimize();
2469
2470 // disable assert until issue with split_flow_path is resolved (6742111)
2471 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2472 // "shouldn't introduce irreducible loops");
2473
2474 if (C->log() != NULL) {
2475 log_loop_tree(_ltree_root, _ltree_root, C->log());
2476 }
2477 }
2478
2479 #ifndef PRODUCT
2480 //------------------------------print_statistics-------------------------------
2481 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2482 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2483 void PhaseIdealLoop::print_statistics() {
2484 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2485 }
2486
2487 //------------------------------verify-----------------------------------------
2488 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2489 static int fail; // debug only, so its multi-thread dont care
2490 void PhaseIdealLoop::verify() const {
2491 int old_progress = C->major_progress();
2492 ResourceMark rm;
2493 PhaseIdealLoop loop_verify( _igvn, this );
2494 VectorSet visited(Thread::current()->resource_area());
2495
2496 fail = 0;
2497 verify_compare( C->root(), &loop_verify, visited );
2498 assert( fail == 0, "verify loops failed" );
2499 // Verify loop structure is the same
2500 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2501 // Reset major-progress. It was cleared by creating a verify version of
2502 // PhaseIdealLoop.
2503 for( int i=0; i<old_progress; i++ )
2504 C->set_major_progress();
2505 }
2506
2507 //------------------------------verify_compare---------------------------------
2508 // Make sure me and the given PhaseIdealLoop agree on key data structures
2509 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2510 if( !n ) return;
2511 if( visited.test_set( n->_idx ) ) return;
2512 if( !_nodes[n->_idx] ) { // Unreachable
2513 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2514 return;
2515 }
2516
2517 uint i;
2518 for( i = 0; i < n->req(); i++ )
2519 verify_compare( n->in(i), loop_verify, visited );
2520
2521 // Check the '_nodes' block/loop structure
2522 i = n->_idx;
2523 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2524 if( _nodes[i] != loop_verify->_nodes[i] &&
2525 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2526 tty->print("Mismatched control setting for: ");
2527 n->dump();
2528 if( fail++ > 10 ) return;
2529 Node *c = get_ctrl_no_update(n);
2530 tty->print("We have it as: ");
2531 if( c->in(0) ) c->dump();
2532 else tty->print_cr("N%d",c->_idx);
2533 tty->print("Verify thinks: ");
2534 if( loop_verify->has_ctrl(n) )
2535 loop_verify->get_ctrl_no_update(n)->dump();
2536 else
2537 loop_verify->get_loop_idx(n)->dump();
2538 tty->cr();
2539 }
2540 } else { // We have a loop
2541 IdealLoopTree *us = get_loop_idx(n);
2542 if( loop_verify->has_ctrl(n) ) {
2543 tty->print("Mismatched loop setting for: ");
2544 n->dump();
2545 if( fail++ > 10 ) return;
2546 tty->print("We have it as: ");
2547 us->dump();
2548 tty->print("Verify thinks: ");
2549 loop_verify->get_ctrl_no_update(n)->dump();
2550 tty->cr();
2551 } else if (!C->major_progress()) {
2552 // Loop selection can be messed up if we did a major progress
2553 // operation, like split-if. Do not verify in that case.
2554 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2555 if( us->_head != them->_head || us->_tail != them->_tail ) {
2556 tty->print("Unequals loops for: ");
2557 n->dump();
2558 if( fail++ > 10 ) return;
2559 tty->print("We have it as: ");
2560 us->dump();
2561 tty->print("Verify thinks: ");
2562 them->dump();
2563 tty->cr();
2564 }
2565 }
2566 }
2567
2568 // Check for immediate dominators being equal
2569 if( i >= _idom_size ) {
2570 if( !n->is_CFG() ) return;
2571 tty->print("CFG Node with no idom: ");
2572 n->dump();
2573 return;
2574 }
2575 if( !n->is_CFG() ) return;
2576 if( n == C->root() ) return; // No IDOM here
2577
2578 assert(n->_idx == i, "sanity");
2579 Node *id = idom_no_update(n);
2580 if( id != loop_verify->idom_no_update(n) ) {
2581 tty->print("Unequals idoms for: ");
2582 n->dump();
2583 if( fail++ > 10 ) return;
2584 tty->print("We have it as: ");
2585 id->dump();
2586 tty->print("Verify thinks: ");
2587 loop_verify->idom_no_update(n)->dump();
2588 tty->cr();
2589 }
2590
2591 }
2592
2593 //------------------------------verify_tree------------------------------------
2594 // Verify that tree structures match. Because the CFG can change, siblings
2595 // within the loop tree can be reordered. We attempt to deal with that by
2596 // reordering the verify's loop tree if possible.
2597 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2598 assert( _parent == parent, "Badly formed loop tree" );
2599
2600 // Siblings not in same order? Attempt to re-order.
2601 if( _head != loop->_head ) {
2602 // Find _next pointer to update
2603 IdealLoopTree **pp = &loop->_parent->_child;
2604 while( *pp != loop )
2605 pp = &((*pp)->_next);
2606 // Find proper sibling to be next
2607 IdealLoopTree **nn = &loop->_next;
2608 while( (*nn) && (*nn)->_head != _head )
2609 nn = &((*nn)->_next);
2610
2611 // Check for no match.
2612 if( !(*nn) ) {
2613 // Annoyingly, irreducible loops can pick different headers
2614 // after a major_progress operation, so the rest of the loop
2615 // tree cannot be matched.
2616 if (_irreducible && Compile::current()->major_progress()) return;
2617 assert( 0, "failed to match loop tree" );
2618 }
2619
2620 // Move (*nn) to (*pp)
2621 IdealLoopTree *hit = *nn;
2622 *nn = hit->_next;
2623 hit->_next = loop;
2624 *pp = loop;
2625 loop = hit;
2626 // Now try again to verify
2627 }
2628
2629 assert( _head == loop->_head , "mismatched loop head" );
2630 Node *tail = _tail; // Inline a non-updating version of
2631 while( !tail->in(0) ) // the 'tail()' call.
2632 tail = tail->in(1);
2633 assert( tail == loop->_tail, "mismatched loop tail" );
2634
2635 // Counted loops that are guarded should be able to find their guards
2636 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2637 CountedLoopNode *cl = _head->as_CountedLoop();
2638 Node *init = cl->init_trip();
2639 Node *ctrl = cl->in(LoopNode::EntryControl);
2640 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2641 Node *iff = ctrl->in(0);
2642 assert( iff->Opcode() == Op_If, "" );
2643 Node *bol = iff->in(1);
2644 assert( bol->Opcode() == Op_Bool, "" );
2645 Node *cmp = bol->in(1);
2646 assert( cmp->Opcode() == Op_CmpI, "" );
2647 Node *add = cmp->in(1);
2648 Node *opaq;
2649 if( add->Opcode() == Op_Opaque1 ) {
2650 opaq = add;
2651 } else {
2652 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2653 assert( add == init, "" );
2654 opaq = cmp->in(2);
2655 }
2656 assert( opaq->Opcode() == Op_Opaque1, "" );
2657
2658 }
2659
2660 if (_child != NULL) _child->verify_tree(loop->_child, this);
2661 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2662 // Innermost loops need to verify loop bodies,
2663 // but only if no 'major_progress'
2664 int fail = 0;
2665 if (!Compile::current()->major_progress() && _child == NULL) {
2666 for( uint i = 0; i < _body.size(); i++ ) {
2667 Node *n = _body.at(i);
2668 if (n->outcnt() == 0) continue; // Ignore dead
2669 uint j;
2670 for( j = 0; j < loop->_body.size(); j++ )
2671 if( loop->_body.at(j) == n )
2672 break;
2673 if( j == loop->_body.size() ) { // Not found in loop body
2674 // Last ditch effort to avoid assertion: Its possible that we
2675 // have some users (so outcnt not zero) but are still dead.
2676 // Try to find from root.
2677 if (Compile::current()->root()->find(n->_idx)) {
2678 fail++;
2679 tty->print("We have that verify does not: ");
2680 n->dump();
2681 }
2682 }
2683 }
2684 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2685 Node *n = loop->_body.at(i2);
2686 if (n->outcnt() == 0) continue; // Ignore dead
2687 uint j;
2688 for( j = 0; j < _body.size(); j++ )
2689 if( _body.at(j) == n )
2690 break;
2691 if( j == _body.size() ) { // Not found in loop body
2692 // Last ditch effort to avoid assertion: Its possible that we
2693 // have some users (so outcnt not zero) but are still dead.
2694 // Try to find from root.
2695 if (Compile::current()->root()->find(n->_idx)) {
2696 fail++;
2697 tty->print("Verify has that we do not: ");
2698 n->dump();
2699 }
2700 }
2701 }
2702 assert( !fail, "loop body mismatch" );
2703 }
2704 }
2705
2706 #endif
2707
2708 //------------------------------set_idom---------------------------------------
2709 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2710 uint idx = d->_idx;
2711 if (idx >= _idom_size) {
2712 uint newsize = _idom_size<<1;
2713 while( idx >= newsize ) {
2714 newsize <<= 1;
2715 }
2716 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2717 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2718 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2719 _idom_size = newsize;
2720 }
2721 _idom[idx] = n;
2722 _dom_depth[idx] = dom_depth;
2723 }
2724
2725 //------------------------------recompute_dom_depth---------------------------------------
2726 // The dominator tree is constructed with only parent pointers.
2727 // This recomputes the depth in the tree by first tagging all
2728 // nodes as "no depth yet" marker. The next pass then runs up
2729 // the dom tree from each node marked "no depth yet", and computes
2730 // the depth on the way back down.
2731 void PhaseIdealLoop::recompute_dom_depth() {
2732 uint no_depth_marker = C->unique();
2733 uint i;
2734 // Initialize depth to "no depth yet"
2735 for (i = 0; i < _idom_size; i++) {
2736 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2737 _dom_depth[i] = no_depth_marker;
2738 }
2739 }
2740 if (_dom_stk == NULL) {
2741 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
2742 if (init_size < 10) init_size = 10;
2743 _dom_stk = new GrowableArray<uint>(init_size);
2744 }
2745 // Compute new depth for each node.
2746 for (i = 0; i < _idom_size; i++) {
2747 uint j = i;
2748 // Run up the dom tree to find a node with a depth
2749 while (_dom_depth[j] == no_depth_marker) {
2750 _dom_stk->push(j);
2751 j = _idom[j]->_idx;
2752 }
2753 // Compute the depth on the way back down this tree branch
2754 uint dd = _dom_depth[j] + 1;
2755 while (_dom_stk->length() > 0) {
2756 uint j = _dom_stk->pop();
2757 _dom_depth[j] = dd;
2758 dd++;
2759 }
2760 }
2761 }
2762
2763 //------------------------------sort-------------------------------------------
2764 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2765 // loop tree, not the root.
2766 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2767 if( !innermost ) return loop; // New innermost loop
2768
2769 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2770 assert( loop_preorder, "not yet post-walked loop" );
2771 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2772 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2773
2774 // Insert at start of list
2775 while( l ) { // Insertion sort based on pre-order
2776 if( l == loop ) return innermost; // Already on list!
2777 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2778 assert( l_preorder, "not yet post-walked l" );
2779 // Check header pre-order number to figure proper nesting
2780 if( loop_preorder > l_preorder )
2781 break; // End of insertion
2782 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2783 // Since I split shared headers, you'd think this could not happen.
2784 // BUT: I must first do the preorder numbering before I can discover I
2785 // have shared headers, so the split headers all get the same preorder
2786 // number as the RegionNode they split from.
2787 if( loop_preorder == l_preorder &&
2788 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2789 break; // Also check for shared headers (same pre#)
2790 pp = &l->_parent; // Chain up list
2791 l = *pp;
2792 }
2793 // Link into list
2794 // Point predecessor to me
2795 *pp = loop;
2796 // Point me to successor
2797 IdealLoopTree *p = loop->_parent;
2798 loop->_parent = l; // Point me to successor
2799 if( p ) sort( p, innermost ); // Insert my parents into list as well
2800 return innermost;
2801 }
2802
2803 //------------------------------build_loop_tree--------------------------------
2804 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2805 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2806 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2807 // tightest enclosing IdealLoopTree for post-walked.
2808 //
2809 // During my forward walk I do a short 1-layer lookahead to see if I can find
2810 // a loop backedge with that doesn't have any work on the backedge. This
2811 // helps me construct nested loops with shared headers better.
2812 //
2813 // Once I've done the forward recursion, I do the post-work. For each child
2814 // I check to see if there is a backedge. Backedges define a loop! I
2815 // insert an IdealLoopTree at the target of the backedge.
2816 //
2817 // During the post-work I also check to see if I have several children
2818 // belonging to different loops. If so, then this Node is a decision point
2819 // where control flow can choose to change loop nests. It is at this
2820 // decision point where I can figure out how loops are nested. At this
2821 // time I can properly order the different loop nests from my children.
2822 // Note that there may not be any backedges at the decision point!
2823 //
2824 // Since the decision point can be far removed from the backedges, I can't
2825 // order my loops at the time I discover them. Thus at the decision point
2826 // I need to inspect loop header pre-order numbers to properly nest my
2827 // loops. This means I need to sort my childrens' loops by pre-order.
2828 // The sort is of size number-of-control-children, which generally limits
2829 // it to size 2 (i.e., I just choose between my 2 target loops).
2830 void PhaseIdealLoop::build_loop_tree() {
2831 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
2832 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
2833 Node *n = C->root();
2834 bltstack.push(n);
2835 int pre_order = 1;
2836 int stack_size;
2837
2838 while ( ( stack_size = bltstack.length() ) != 0 ) {
2839 n = bltstack.top(); // Leave node on stack
2840 if ( !is_visited(n) ) {
2841 // ---- Pre-pass Work ----
2842 // Pre-walked but not post-walked nodes need a pre_order number.
2843
2844 set_preorder_visited( n, pre_order ); // set as visited
2845
2846 // ---- Scan over children ----
2847 // Scan first over control projections that lead to loop headers.
2848 // This helps us find inner-to-outer loops with shared headers better.
2849
2850 // Scan children's children for loop headers.
2851 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2852 Node* m = n->raw_out(i); // Child
2853 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2854 // Scan over children's children to find loop
2855 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2856 Node* l = m->fast_out(j);
2857 if( is_visited(l) && // Been visited?
2858 !is_postvisited(l) && // But not post-visited
2859 get_preorder(l) < pre_order ) { // And smaller pre-order
2860 // Found! Scan the DFS down this path before doing other paths
2861 bltstack.push(m);
2862 break;
2863 }
2864 }
2865 }
2866 }
2867 pre_order++;
2868 }
2869 else if ( !is_postvisited(n) ) {
2870 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2871 // such as com.sun.rsasign.am::a.
2872 // For non-recursive version, first, process current children.
2873 // On next iteration, check if additional children were added.
2874 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2875 Node* u = n->raw_out(k);
2876 if ( u->is_CFG() && !is_visited(u) ) {
2877 bltstack.push(u);
2878 }
2879 }
2880 if ( bltstack.length() == stack_size ) {
2881 // There were no additional children, post visit node now
2882 (void)bltstack.pop(); // Remove node from stack
2883 pre_order = build_loop_tree_impl( n, pre_order );
2884 // Check for bailout
2885 if (C->failing()) {
2886 return;
2887 }
2888 // Check to grow _preorders[] array for the case when
2889 // build_loop_tree_impl() adds new nodes.
2890 check_grow_preorders();
2891 }
2892 }
2893 else {
2894 (void)bltstack.pop(); // Remove post-visited node from stack
2895 }
2896 }
2897 }
2898
2899 //------------------------------build_loop_tree_impl---------------------------
2900 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2901 // ---- Post-pass Work ----
2902 // Pre-walked but not post-walked nodes need a pre_order number.
2903
2904 // Tightest enclosing loop for this Node
2905 IdealLoopTree *innermost = NULL;
2906
2907 // For all children, see if any edge is a backedge. If so, make a loop
2908 // for it. Then find the tightest enclosing loop for the self Node.
2909 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2910 Node* m = n->fast_out(i); // Child
2911 if( n == m ) continue; // Ignore control self-cycles
2912 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2913
2914 IdealLoopTree *l; // Child's loop
2915 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2916 // Found a backedge
2917 assert( get_preorder(m) < pre_order, "should be backedge" );
2918 // Check for the RootNode, which is already a LoopNode and is allowed
2919 // to have multiple "backedges".
2920 if( m == C->root()) { // Found the root?
2921 l = _ltree_root; // Root is the outermost LoopNode
2922 } else { // Else found a nested loop
2923 // Insert a LoopNode to mark this loop.
2924 l = new IdealLoopTree(this, m, n);
2925 } // End of Else found a nested loop
2926 if( !has_loop(m) ) // If 'm' does not already have a loop set
2927 set_loop(m, l); // Set loop header to loop now
2928
2929 } else { // Else not a nested loop
2930 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2931 l = get_loop(m); // Get previously determined loop
2932 // If successor is header of a loop (nest), move up-loop till it
2933 // is a member of some outer enclosing loop. Since there are no
2934 // shared headers (I've split them already) I only need to go up
2935 // at most 1 level.
2936 while( l && l->_head == m ) // Successor heads loop?
2937 l = l->_parent; // Move up 1 for me
2938 // If this loop is not properly parented, then this loop
2939 // has no exit path out, i.e. its an infinite loop.
2940 if( !l ) {
2941 // Make loop "reachable" from root so the CFG is reachable. Basically
2942 // insert a bogus loop exit that is never taken. 'm', the loop head,
2943 // points to 'n', one (of possibly many) fall-in paths. There may be
2944 // many backedges as well.
2945
2946 // Here I set the loop to be the root loop. I could have, after
2947 // inserting a bogus loop exit, restarted the recursion and found my
2948 // new loop exit. This would make the infinite loop a first-class
2949 // loop and it would then get properly optimized. What's the use of
2950 // optimizing an infinite loop?
2951 l = _ltree_root; // Oops, found infinite loop
2952
2953 if (!_verify_only) {
2954 // Insert the NeverBranch between 'm' and it's control user.
2955 NeverBranchNode *iff = new (C) NeverBranchNode( m );
2956 _igvn.register_new_node_with_optimizer(iff);
2957 set_loop(iff, l);
2958 Node *if_t = new (C) CProjNode( iff, 0 );
2959 _igvn.register_new_node_with_optimizer(if_t);
2960 set_loop(if_t, l);
2961
2962 Node* cfg = NULL; // Find the One True Control User of m
2963 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2964 Node* x = m->fast_out(j);
2965 if (x->is_CFG() && x != m && x != iff)
2966 { cfg = x; break; }
2967 }
2968 assert(cfg != NULL, "must find the control user of m");
2969 uint k = 0; // Probably cfg->in(0)
2970 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2971 cfg->set_req( k, if_t ); // Now point to NeverBranch
2972
2973 // Now create the never-taken loop exit
2974 Node *if_f = new (C) CProjNode( iff, 1 );
2975 _igvn.register_new_node_with_optimizer(if_f);
2976 set_loop(if_f, l);
2977 // Find frame ptr for Halt. Relies on the optimizer
2978 // V-N'ing. Easier and quicker than searching through
2979 // the program structure.
2980 Node *frame = new (C) ParmNode( C->start(), TypeFunc::FramePtr );
2981 _igvn.register_new_node_with_optimizer(frame);
2982 // Halt & Catch Fire
2983 Node *halt = new (C) HaltNode( if_f, frame );
2984 _igvn.register_new_node_with_optimizer(halt);
2985 set_loop(halt, l);
2986 C->root()->add_req(halt);
2987 }
2988 set_loop(C->root(), _ltree_root);
2989 }
2990 }
2991 // Weeny check for irreducible. This child was already visited (this
2992 // IS the post-work phase). Is this child's loop header post-visited
2993 // as well? If so, then I found another entry into the loop.
2994 if (!_verify_only) {
2995 while( is_postvisited(l->_head) ) {
2996 // found irreducible
2997 l->_irreducible = 1; // = true
2998 l = l->_parent;
2999 _has_irreducible_loops = true;
3000 // Check for bad CFG here to prevent crash, and bailout of compile
3001 if (l == NULL) {
3002 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3003 return pre_order;
3004 }
3005 }
3006 C->set_has_irreducible_loop(_has_irreducible_loops);
3007 }
3008
3009 // This Node might be a decision point for loops. It is only if
3010 // it's children belong to several different loops. The sort call
3011 // does a trivial amount of work if there is only 1 child or all
3012 // children belong to the same loop. If however, the children
3013 // belong to different loops, the sort call will properly set the
3014 // _parent pointers to show how the loops nest.
3015 //
3016 // In any case, it returns the tightest enclosing loop.
3017 innermost = sort( l, innermost );
3018 }
3019
3020 // Def-use info will have some dead stuff; dead stuff will have no
3021 // loop decided on.
3022
3023 // Am I a loop header? If so fix up my parent's child and next ptrs.
3024 if( innermost && innermost->_head == n ) {
3025 assert( get_loop(n) == innermost, "" );
3026 IdealLoopTree *p = innermost->_parent;
3027 IdealLoopTree *l = innermost;
3028 while( p && l->_head == n ) {
3029 l->_next = p->_child; // Put self on parents 'next child'
3030 p->_child = l; // Make self as first child of parent
3031 l = p; // Now walk up the parent chain
3032 p = l->_parent;
3033 }
3034 } else {
3035 // Note that it is possible for a LoopNode to reach here, if the
3036 // backedge has been made unreachable (hence the LoopNode no longer
3037 // denotes a Loop, and will eventually be removed).
3038
3039 // Record tightest enclosing loop for self. Mark as post-visited.
3040 set_loop(n, innermost);
3041 // Also record has_call flag early on
3042 if( innermost ) {
3043 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3044 // Do not count uncommon calls
3045 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3046 Node *iff = n->in(0)->in(0);
3047 // No any calls for vectorized loops.
3048 if( UseSuperWord || !iff->is_If() ||
3049 (n->in(0)->Opcode() == Op_IfFalse &&
3050 (1.0 - iff->as_If()->_prob) >= 0.01) ||
3051 (iff->as_If()->_prob >= 0.01) )
3052 innermost->_has_call = 1;
3053 }
3054 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3055 // Disable loop optimizations if the loop has a scalar replaceable
3056 // allocation. This disabling may cause a potential performance lost
3057 // if the allocation is not eliminated for some reason.
3058 innermost->_allow_optimizations = false;
3059 innermost->_has_call = 1; // = true
3060 } else if (n->Opcode() == Op_SafePoint) {
3061 // Record all safepoints in this loop.
3062 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3063 innermost->_safepts->push(n);
3064 }
3065 }
3066 }
3067
3068 // Flag as post-visited now
3069 set_postvisited(n);
3070 return pre_order;
3071 }
3072
3073
3074 //------------------------------build_loop_early-------------------------------
3075 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3076 // First pass computes the earliest controlling node possible. This is the
3077 // controlling input with the deepest dominating depth.
3078 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3079 while (worklist.size() != 0) {
3080 // Use local variables nstack_top_n & nstack_top_i to cache values
3081 // on nstack's top.
3082 Node *nstack_top_n = worklist.pop();
3083 uint nstack_top_i = 0;
3084 //while_nstack_nonempty:
3085 while (true) {
3086 // Get parent node and next input's index from stack's top.
3087 Node *n = nstack_top_n;
3088 uint i = nstack_top_i;
3089 uint cnt = n->req(); // Count of inputs
3090 if (i == 0) { // Pre-process the node.
3091 if( has_node(n) && // Have either loop or control already?
3092 !has_ctrl(n) ) { // Have loop picked out already?
3093 // During "merge_many_backedges" we fold up several nested loops
3094 // into a single loop. This makes the members of the original
3095 // loop bodies pointing to dead loops; they need to move up
3096 // to the new UNION'd larger loop. I set the _head field of these
3097 // dead loops to NULL and the _parent field points to the owning
3098 // loop. Shades of UNION-FIND algorithm.
3099 IdealLoopTree *ilt;
3100 while( !(ilt = get_loop(n))->_head ) {
3101 // Normally I would use a set_loop here. But in this one special
3102 // case, it is legal (and expected) to change what loop a Node
3103 // belongs to.
3104 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3105 }
3106 // Remove safepoints ONLY if I've already seen I don't need one.
3107 // (the old code here would yank a 2nd safepoint after seeing a
3108 // first one, even though the 1st did not dominate in the loop body
3109 // and thus could be avoided indefinitely)
3110 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3111 is_deleteable_safept(n)) {
3112 Node *in = n->in(TypeFunc::Control);
3113 lazy_replace(n,in); // Pull safepoint now
3114 if (ilt->_safepts != NULL) {
3115 ilt->_safepts->yank(n);
3116 }
3117 // Carry on with the recursion "as if" we are walking
3118 // only the control input
3119 if( !visited.test_set( in->_idx ) ) {
3120 worklist.push(in); // Visit this guy later, using worklist
3121 }
3122 // Get next node from nstack:
3123 // - skip n's inputs processing by setting i > cnt;
3124 // - we also will not call set_early_ctrl(n) since
3125 // has_node(n) == true (see the condition above).
3126 i = cnt + 1;
3127 }
3128 }
3129 } // if (i == 0)
3130
3131 // Visit all inputs
3132 bool done = true; // Assume all n's inputs will be processed
3133 while (i < cnt) {
3134 Node *in = n->in(i);
3135 ++i;
3136 if (in == NULL) continue;
3137 if (in->pinned() && !in->is_CFG())
3138 set_ctrl(in, in->in(0));
3139 int is_visited = visited.test_set( in->_idx );
3140 if (!has_node(in)) { // No controlling input yet?
3141 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3142 assert( !is_visited, "visit only once" );
3143 nstack.push(n, i); // Save parent node and next input's index.
3144 nstack_top_n = in; // Process current input now.
3145 nstack_top_i = 0;
3146 done = false; // Not all n's inputs processed.
3147 break; // continue while_nstack_nonempty;
3148 } else if (!is_visited) {
3149 // This guy has a location picked out for him, but has not yet
3150 // been visited. Happens to all CFG nodes, for instance.
3151 // Visit him using the worklist instead of recursion, to break
3152 // cycles. Since he has a location already we do not need to
3153 // find his location before proceeding with the current Node.
3154 worklist.push(in); // Visit this guy later, using worklist
3155 }
3156 }
3157 if (done) {
3158 // All of n's inputs have been processed, complete post-processing.
3159
3160 // Compute earliest point this Node can go.
3161 // CFG, Phi, pinned nodes already know their controlling input.
3162 if (!has_node(n)) {
3163 // Record earliest legal location
3164 set_early_ctrl( n );
3165 }
3166 if (nstack.is_empty()) {
3167 // Finished all nodes on stack.
3168 // Process next node on the worklist.
3169 break;
3170 }
3171 // Get saved parent node and next input's index.
3172 nstack_top_n = nstack.node();
3173 nstack_top_i = nstack.index();
3174 nstack.pop();
3175 }
3176 } // while (true)
3177 }
3178 }
3179
3180 //------------------------------dom_lca_internal--------------------------------
3181 // Pair-wise LCA
3182 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3183 if( !n1 ) return n2; // Handle NULL original LCA
3184 assert( n1->is_CFG(), "" );
3185 assert( n2->is_CFG(), "" );
3186 // find LCA of all uses
3187 uint d1 = dom_depth(n1);
3188 uint d2 = dom_depth(n2);
3189 while (n1 != n2) {
3190 if (d1 > d2) {
3191 n1 = idom(n1);
3192 d1 = dom_depth(n1);
3193 } else if (d1 < d2) {
3194 n2 = idom(n2);
3195 d2 = dom_depth(n2);
3196 } else {
3197 // Here d1 == d2. Due to edits of the dominator-tree, sections
3198 // of the tree might have the same depth. These sections have
3199 // to be searched more carefully.
3200
3201 // Scan up all the n1's with equal depth, looking for n2.
3202 Node *t1 = idom(n1);
3203 while (dom_depth(t1) == d1) {
3204 if (t1 == n2) return n2;
3205 t1 = idom(t1);
3206 }
3207 // Scan up all the n2's with equal depth, looking for n1.
3208 Node *t2 = idom(n2);
3209 while (dom_depth(t2) == d2) {
3210 if (t2 == n1) return n1;
3211 t2 = idom(t2);
3212 }
3213 // Move up to a new dominator-depth value as well as up the dom-tree.
3214 n1 = t1;
3215 n2 = t2;
3216 d1 = dom_depth(n1);
3217 d2 = dom_depth(n2);
3218 }
3219 }
3220 return n1;
3221 }
3222
3223 //------------------------------compute_idom-----------------------------------
3224 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3225 // IDOMs are correct.
3226 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3227 assert( region->is_Region(), "" );
3228 Node *LCA = NULL;
3229 for( uint i = 1; i < region->req(); i++ ) {
3230 if( region->in(i) != C->top() )
3231 LCA = dom_lca( LCA, region->in(i) );
3232 }
3233 return LCA;
3234 }
3235
3236 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3237 bool had_error = false;
3238 #ifdef ASSERT
3239 if (early != C->root()) {
3240 // Make sure that there's a dominance path from LCA to early
3241 Node* d = LCA;
3242 while (d != early) {
3243 if (d == C->root()) {
3244 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3245 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3246 had_error = true;
3247 break;
3248 }
3249 d = idom(d);
3250 }
3251 }
3252 #endif
3253 return had_error;
3254 }
3255
3256
3257 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3258 // Compute LCA over list of uses
3259 bool had_error = false;
3260 Node *LCA = NULL;
3261 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3262 Node* c = n->fast_out(i);
3263 if (_nodes[c->_idx] == NULL)
3264 continue; // Skip the occasional dead node
3265 if( c->is_Phi() ) { // For Phis, we must land above on the path
3266 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3267 if( c->in(j) == n ) { // Found matching input?
3268 Node *use = c->in(0)->in(j);
3269 if (_verify_only && use->is_top()) continue;
3270 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3271 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3272 }
3273 }
3274 } else {
3275 // For CFG data-users, use is in the block just prior
3276 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3277 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3278 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3279 }
3280 }
3281 assert(!had_error, "bad dominance");
3282 return LCA;
3283 }
3284
3285 // Check the shape of the graph at the loop entry. In some cases,
3286 // the shape of the graph does not match the shape outlined below.
3287 // That is caused by the Opaque1 node "protecting" the shape of
3288 // the graph being removed by, for example, the IGVN performed
3289 // in PhaseIdealLoop::build_and_optimize().
3290 //
3291 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3292 // loop unswitching, and IGVN, or a combination of them) can freely change
3293 // the graph's shape. As a result, the graph shape outlined below cannot
3294 // be guaranteed anymore.
3295 bool PhaseIdealLoop::is_canonical_main_loop_entry(CountedLoopNode* cl) {
3296 assert(cl->is_main_loop(), "check should be applied to main loops");
3297 Node* ctrl = cl->in(LoopNode::EntryControl);
3298 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3299 return false;
3300 }
3301 Node* iffm = ctrl->in(0);
3302 if (iffm == NULL || !iffm->is_If()) {
3303 return false;
3304 }
3305 Node* bolzm = iffm->in(1);
3306 if (bolzm == NULL || !bolzm->is_Bool()) {
3307 return false;
3308 }
3309 Node* cmpzm = bolzm->in(1);
3310 if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3311 return false;
3312 }
3313 Node* opqzm = cmpzm->in(2);
3314 if (opqzm == NULL || opqzm->Opcode() != Op_Opaque1) {
3315 return false;
3316 }
3317 return true;
3318 }
3319
3320 //------------------------------get_late_ctrl----------------------------------
3321 // Compute latest legal control.
3322 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3323 assert(early != NULL, "early control should not be NULL");
3324
3325 Node* LCA = compute_lca_of_uses(n, early);
3326 #ifdef ASSERT
3327 if (LCA == C->root() && LCA != early) {
3328 // def doesn't dominate uses so print some useful debugging output
3329 compute_lca_of_uses(n, early, true);
3330 }
3331 #endif
3332
3333 // if this is a load, check for anti-dependent stores
3334 // We use a conservative algorithm to identify potential interfering
3335 // instructions and for rescheduling the load. The users of the memory
3336 // input of this load are examined. Any use which is not a load and is
3337 // dominated by early is considered a potentially interfering store.
3338 // This can produce false positives.
3339 if (n->is_Load() && LCA != early) {
3340 Node_List worklist;
3341
3342 Node *mem = n->in(MemNode::Memory);
3343 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3344 Node* s = mem->fast_out(i);
3345 worklist.push(s);
3346 }
3347 while(worklist.size() != 0 && LCA != early) {
3348 Node* s = worklist.pop();
3349 if (s->is_Load()) {
3350 continue;
3351 } else if (s->is_MergeMem()) {
3352 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3353 Node* s1 = s->fast_out(i);
3354 worklist.push(s1);
3355 }
3356 } else {
3357 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3358 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3359 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3360 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3361 }
3362 }
3363 }
3364 }
3365
3366 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3367 return LCA;
3368 }
3369
3370 // true if CFG node d dominates CFG node n
3371 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3372 if (d == n)
3373 return true;
3374 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3375 uint dd = dom_depth(d);
3376 while (dom_depth(n) >= dd) {
3377 if (n == d)
3378 return true;
3379 n = idom(n);
3380 }
3381 return false;
3382 }
3383
3384 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3385 // Pair-wise LCA with tags.
3386 // Tag each index with the node 'tag' currently being processed
3387 // before advancing up the dominator chain using idom().
3388 // Later calls that find a match to 'tag' know that this path has already
3389 // been considered in the current LCA (which is input 'n1' by convention).
3390 // Since get_late_ctrl() is only called once for each node, the tag array
3391 // does not need to be cleared between calls to get_late_ctrl().
3392 // Algorithm trades a larger constant factor for better asymptotic behavior
3393 //
3394 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3395 uint d1 = dom_depth(n1);
3396 uint d2 = dom_depth(n2);
3397
3398 do {
3399 if (d1 > d2) {
3400 // current lca is deeper than n2
3401 _dom_lca_tags.map(n1->_idx, tag);
3402 n1 = idom(n1);
3403 d1 = dom_depth(n1);
3404 } else if (d1 < d2) {
3405 // n2 is deeper than current lca
3406 Node *memo = _dom_lca_tags[n2->_idx];
3407 if( memo == tag ) {
3408 return n1; // Return the current LCA
3409 }
3410 _dom_lca_tags.map(n2->_idx, tag);
3411 n2 = idom(n2);
3412 d2 = dom_depth(n2);
3413 } else {
3414 // Here d1 == d2. Due to edits of the dominator-tree, sections
3415 // of the tree might have the same depth. These sections have
3416 // to be searched more carefully.
3417
3418 // Scan up all the n1's with equal depth, looking for n2.
3419 _dom_lca_tags.map(n1->_idx, tag);
3420 Node *t1 = idom(n1);
3421 while (dom_depth(t1) == d1) {
3422 if (t1 == n2) return n2;
3423 _dom_lca_tags.map(t1->_idx, tag);
3424 t1 = idom(t1);
3425 }
3426 // Scan up all the n2's with equal depth, looking for n1.
3427 _dom_lca_tags.map(n2->_idx, tag);
3428 Node *t2 = idom(n2);
3429 while (dom_depth(t2) == d2) {
3430 if (t2 == n1) return n1;
3431 _dom_lca_tags.map(t2->_idx, tag);
3432 t2 = idom(t2);
3433 }
3434 // Move up to a new dominator-depth value as well as up the dom-tree.
3435 n1 = t1;
3436 n2 = t2;
3437 d1 = dom_depth(n1);
3438 d2 = dom_depth(n2);
3439 }
3440 } while (n1 != n2);
3441 return n1;
3442 }
3443
3444 //------------------------------init_dom_lca_tags------------------------------
3445 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3446 // Intended use does not involve any growth for the array, so it could
3447 // be of fixed size.
3448 void PhaseIdealLoop::init_dom_lca_tags() {
3449 uint limit = C->unique() + 1;
3450 _dom_lca_tags.map( limit, NULL );
3451 #ifdef ASSERT
3452 for( uint i = 0; i < limit; ++i ) {
3453 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3454 }
3455 #endif // ASSERT
3456 }
3457
3458 //------------------------------clear_dom_lca_tags------------------------------
3459 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3460 // Intended use does not involve any growth for the array, so it could
3461 // be of fixed size.
3462 void PhaseIdealLoop::clear_dom_lca_tags() {
3463 uint limit = C->unique() + 1;
3464 _dom_lca_tags.map( limit, NULL );
3465 _dom_lca_tags.clear();
3466 #ifdef ASSERT
3467 for( uint i = 0; i < limit; ++i ) {
3468 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3469 }
3470 #endif // ASSERT
3471 }
3472
3473 //------------------------------build_loop_late--------------------------------
3474 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3475 // Second pass finds latest legal placement, and ideal loop placement.
3476 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3477 while (worklist.size() != 0) {
3478 Node *n = worklist.pop();
3479 // Only visit once
3480 if (visited.test_set(n->_idx)) continue;
3481 uint cnt = n->outcnt();
3482 uint i = 0;
3483 while (true) {
3484 assert( _nodes[n->_idx], "no dead nodes" );
3485 // Visit all children
3486 if (i < cnt) {
3487 Node* use = n->raw_out(i);
3488 ++i;
3489 // Check for dead uses. Aggressively prune such junk. It might be
3490 // dead in the global sense, but still have local uses so I cannot
3491 // easily call 'remove_dead_node'.
3492 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3493 // Due to cycles, we might not hit the same fixed point in the verify
3494 // pass as we do in the regular pass. Instead, visit such phis as
3495 // simple uses of the loop head.
3496 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3497 if( !visited.test(use->_idx) )
3498 worklist.push(use);
3499 } else if( !visited.test_set(use->_idx) ) {
3500 nstack.push(n, i); // Save parent and next use's index.
3501 n = use; // Process all children of current use.
3502 cnt = use->outcnt();
3503 i = 0;
3504 }
3505 } else {
3506 // Do not visit around the backedge of loops via data edges.
3507 // push dead code onto a worklist
3508 _deadlist.push(use);
3509 }
3510 } else {
3511 // All of n's children have been processed, complete post-processing.
3512 build_loop_late_post(n);
3513 if (nstack.is_empty()) {
3514 // Finished all nodes on stack.
3515 // Process next node on the worklist.
3516 break;
3517 }
3518 // Get saved parent node and next use's index. Visit the rest of uses.
3519 n = nstack.node();
3520 cnt = n->outcnt();
3521 i = nstack.index();
3522 nstack.pop();
3523 }
3524 }
3525 }
3526 }
3527
3528 //------------------------------build_loop_late_post---------------------------
3529 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3530 // Second pass finds latest legal placement, and ideal loop placement.
3531 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3532
3533 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3534 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3535 }
3536
3537 #ifdef ASSERT
3538 if (_verify_only && !n->is_CFG()) {
3539 // Check def-use domination.
3540 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
3541 }
3542 #endif
3543
3544 // CFG and pinned nodes already handled
3545 if( n->in(0) ) {
3546 if( n->in(0)->is_top() ) return; // Dead?
3547
3548 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3549 // _must_ be pinned (they have to observe their control edge of course).
3550 // Unlike Stores (which modify an unallocable resource, the memory
3551 // state), Mods/Loads can float around. So free them up.
3552 bool pinned = true;
3553 switch( n->Opcode() ) {
3554 case Op_DivI:
3555 case Op_DivF:
3556 case Op_DivD:
3557 case Op_ModI:
3558 case Op_ModF:
3559 case Op_ModD:
3560 case Op_LoadB: // Same with Loads; they can sink
3561 case Op_LoadUB: // during loop optimizations.
3562 case Op_LoadUS:
3563 case Op_LoadD:
3564 case Op_LoadF:
3565 case Op_LoadI:
3566 case Op_LoadKlass:
3567 case Op_LoadNKlass:
3568 case Op_LoadL:
3569 case Op_LoadS:
3570 case Op_LoadP:
3571 case Op_LoadN:
3572 case Op_LoadRange:
3573 case Op_LoadD_unaligned:
3574 case Op_LoadL_unaligned:
3575 case Op_StrComp: // Does a bunch of load-like effects
3576 case Op_StrEquals:
3577 case Op_StrIndexOf:
3578 case Op_AryEq:
3579 pinned = false;
3580 }
3581 if( pinned ) {
3582 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3583 if( !chosen_loop->_child ) // Inner loop?
3584 chosen_loop->_body.push(n); // Collect inner loops
3585 return;
3586 }
3587 } else { // No slot zero
3588 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3589 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3590 return;
3591 }
3592 assert(!n->is_CFG() || n->outcnt() == 0, "");
3593 }
3594
3595 // Do I have a "safe range" I can select over?
3596 Node *early = get_ctrl(n);// Early location already computed
3597
3598 // Compute latest point this Node can go
3599 Node *LCA = get_late_ctrl( n, early );
3600 // LCA is NULL due to uses being dead
3601 if( LCA == NULL ) {
3602 #ifdef ASSERT
3603 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3604 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3605 }
3606 #endif
3607 _nodes.map(n->_idx, 0); // This node is useless
3608 _deadlist.push(n);
3609 return;
3610 }
3611 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3612
3613 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3614 Node *least = legal; // Best legal position so far
3615 while( early != legal ) { // While not at earliest legal
3616 #ifdef ASSERT
3617 if (legal->is_Start() && !early->is_Root()) {
3618 // Bad graph. Print idom path and fail.
3619 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
3620 assert(false, "Bad graph detected in build_loop_late");
3621 }
3622 #endif
3623 // Find least loop nesting depth
3624 legal = idom(legal); // Bump up the IDOM tree
3625 // Check for lower nesting depth
3626 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3627 least = legal;
3628 }
3629 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3630
3631 // Try not to place code on a loop entry projection
3632 // which can inhibit range check elimination.
3633 if (least != early) {
3634 Node* ctrl_out = least->unique_ctrl_out();
3635 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3636 least == ctrl_out->in(LoopNode::EntryControl)) {
3637 Node* least_dom = idom(least);
3638 if (get_loop(least_dom)->is_member(get_loop(least))) {
3639 least = least_dom;
3640 }
3641 }
3642 }
3643
3644 #ifdef ASSERT
3645 // If verifying, verify that 'verify_me' has a legal location
3646 // and choose it as our location.
3647 if( _verify_me ) {
3648 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3649 Node *legal = LCA;
3650 while( early != legal ) { // While not at earliest legal
3651 if( legal == v_ctrl ) break; // Check for prior good location
3652 legal = idom(legal) ;// Bump up the IDOM tree
3653 }
3654 // Check for prior good location
3655 if( legal == v_ctrl ) least = legal; // Keep prior if found
3656 }
3657 #endif
3658
3659 // Assign discovered "here or above" point
3660 least = find_non_split_ctrl(least);
3661 set_ctrl(n, least);
3662
3663 // Collect inner loop bodies
3664 IdealLoopTree *chosen_loop = get_loop(least);
3665 if( !chosen_loop->_child ) // Inner loop?
3666 chosen_loop->_body.push(n);// Collect inner loops
3667 }
3668
3669 #ifdef ASSERT
3670 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
3671 tty->print_cr("%s", msg);
3672 tty->print("n: "); n->dump();
3673 tty->print("early(n): "); early->dump();
3674 if (n->in(0) != NULL && !n->in(0)->is_top() &&
3675 n->in(0) != early && !n->in(0)->is_Root()) {
3676 tty->print("n->in(0): "); n->in(0)->dump();
3677 }
3678 for (uint i = 1; i < n->req(); i++) {
3679 Node* in1 = n->in(i);
3680 if (in1 != NULL && in1 != n && !in1->is_top()) {
3681 tty->print("n->in(%d): ", i); in1->dump();
3682 Node* in1_early = get_ctrl(in1);
3683 tty->print("early(n->in(%d)): ", i); in1_early->dump();
3684 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
3685 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3686 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3687 }
3688 for (uint j = 1; j < in1->req(); j++) {
3689 Node* in2 = in1->in(j);
3690 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3691 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3692 Node* in2_early = get_ctrl(in2);
3693 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3694 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
3695 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3696 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3697 }
3698 }
3699 }
3700 }
3701 }
3702 tty->cr();
3703 tty->print("LCA(n): "); LCA->dump();
3704 for (uint i = 0; i < n->outcnt(); i++) {
3705 Node* u1 = n->raw_out(i);
3706 if (u1 == n)
3707 continue;
3708 tty->print("n->out(%d): ", i); u1->dump();
3709 if (u1->is_CFG()) {
3710 for (uint j = 0; j < u1->outcnt(); j++) {
3711 Node* u2 = u1->raw_out(j);
3712 if (u2 != u1 && u2 != n && u2->is_CFG()) {
3713 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3714 }
3715 }
3716 } else {
3717 Node* u1_later = get_ctrl(u1);
3718 tty->print("later(n->out(%d)): ", i); u1_later->dump();
3719 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
3720 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3721 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3722 }
3723 for (uint j = 0; j < u1->outcnt(); j++) {
3724 Node* u2 = u1->raw_out(j);
3725 if (u2 == n || u2 == u1)
3726 continue;
3727 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3728 if (!u2->is_CFG()) {
3729 Node* u2_later = get_ctrl(u2);
3730 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3731 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
3732 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3733 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3734 }
3735 }
3736 }
3737 }
3738 }
3739 tty->cr();
3740 int ct = 0;
3741 Node *dbg_legal = LCA;
3742 while(!dbg_legal->is_Start() && ct < 100) {
3743 tty->print("idom[%d] ",ct); dbg_legal->dump();
3744 ct++;
3745 dbg_legal = idom(dbg_legal);
3746 }
3747 tty->cr();
3748 }
3749 #endif
3750
3751 #ifndef PRODUCT
3752 //------------------------------dump-------------------------------------------
3753 void PhaseIdealLoop::dump( ) const {
3754 ResourceMark rm;
3755 Arena* arena = Thread::current()->resource_area();
3756 Node_Stack stack(arena, C->live_nodes() >> 2);
3757 Node_List rpo_list;
3758 VectorSet visited(arena);
3759 visited.set(C->top()->_idx);
3760 rpo( C->root(), stack, visited, rpo_list );
3761 // Dump root loop indexed by last element in PO order
3762 dump( _ltree_root, rpo_list.size(), rpo_list );
3763 }
3764
3765 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3766 loop->dump_head();
3767
3768 // Now scan for CFG nodes in the same loop
3769 for( uint j=idx; j > 0; j-- ) {
3770 Node *n = rpo_list[j-1];
3771 if( !_nodes[n->_idx] ) // Skip dead nodes
3772 continue;
3773 if( get_loop(n) != loop ) { // Wrong loop nest
3774 if( get_loop(n)->_head == n && // Found nested loop?
3775 get_loop(n)->_parent == loop )
3776 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3777 continue;
3778 }
3779
3780 // Dump controlling node
3781 for( uint x = 0; x < loop->_nest; x++ )
3782 tty->print(" ");
3783 tty->print("C");
3784 if( n == C->root() ) {
3785 n->dump();
3786 } else {
3787 Node* cached_idom = idom_no_update(n);
3788 Node *computed_idom = n->in(0);
3789 if( n->is_Region() ) {
3790 computed_idom = compute_idom(n);
3791 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3792 // any MultiBranch ctrl node), so apply a similar transform to
3793 // the cached idom returned from idom_no_update.
3794 cached_idom = find_non_split_ctrl(cached_idom);
3795 }
3796 tty->print(" ID:%d",computed_idom->_idx);
3797 n->dump();
3798 if( cached_idom != computed_idom ) {
3799 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3800 computed_idom->_idx, cached_idom->_idx);
3801 }
3802 }
3803 // Dump nodes it controls
3804 for( uint k = 0; k < _nodes.Size(); k++ ) {
3805 // (k < C->unique() && get_ctrl(find(k)) == n)
3806 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3807 Node *m = C->root()->find(k);
3808 if( m && m->outcnt() > 0 ) {
3809 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3810 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3811 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3812 }
3813 for( uint j = 0; j < loop->_nest; j++ )
3814 tty->print(" ");
3815 tty->print(" ");
3816 m->dump();
3817 }
3818 }
3819 }
3820 }
3821 }
3822
3823 // Collect a R-P-O for the whole CFG.
3824 // Result list is in post-order (scan backwards for RPO)
3825 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3826 stk.push(start, 0);
3827 visited.set(start->_idx);
3828
3829 while (stk.is_nonempty()) {
3830 Node* m = stk.node();
3831 uint idx = stk.index();
3832 if (idx < m->outcnt()) {
3833 stk.set_index(idx + 1);
3834 Node* n = m->raw_out(idx);
3835 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3836 stk.push(n, 0);
3837 }
3838 } else {
3839 rpo_list.push(m);
3840 stk.pop();
3841 }
3842 }
3843 }
3844 #endif
3845
3846
3847 //=============================================================================
3848 //------------------------------LoopTreeIterator-----------------------------------
3849
3850 // Advance to next loop tree using a preorder, left-to-right traversal.
3851 void LoopTreeIterator::next() {
3852 assert(!done(), "must not be done.");
3853 if (_curnt->_child != NULL) {
3854 _curnt = _curnt->_child;
3855 } else if (_curnt->_next != NULL) {
3856 _curnt = _curnt->_next;
3857 } else {
3858 while (_curnt != _root && _curnt->_next == NULL) {
3859 _curnt = _curnt->_parent;
3860 }
3861 if (_curnt == _root) {
3862 _curnt = NULL;
3863 assert(done(), "must be done.");
3864 } else {
3865 assert(_curnt->_next != NULL, "must be more to do");
3866 _curnt = _curnt->_next;
3867 }
3868 }
3869 }