1 /*
2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "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) ) {
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 return early;
136 }
137
138 //------------------------------set_early_ctrl---------------------------------
139 // Set earliest legal control
140 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
141 Node *early = get_early_ctrl(n);
142
143 // Record earliest legal location
144 set_ctrl(n, early);
145 }
146
147 //------------------------------set_subtree_ctrl-------------------------------
148 // set missing _ctrl entries on new nodes
149 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
150 // Already set? Get out.
151 if( _nodes[n->_idx] ) return;
152 // Recursively set _nodes array to indicate where the Node goes
153 uint i;
154 for( i = 0; i < n->req(); ++i ) {
155 Node *m = n->in(i);
156 if( m && m != C->root() )
157 set_subtree_ctrl( m );
158 }
159
160 // Fixup self
161 set_early_ctrl( n );
162 }
163
164 //------------------------------is_counted_loop--------------------------------
165 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
166 PhaseGVN *gvn = &_igvn;
167
168 // Counted loop head must be a good RegionNode with only 3 not NULL
169 // control input edges: Self, Entry, LoopBack.
170 if (x->in(LoopNode::Self) == NULL || x->req() != 3)
171 return false;
172
173 Node *init_control = x->in(LoopNode::EntryControl);
174 Node *back_control = x->in(LoopNode::LoopBackControl);
175 if (init_control == NULL || back_control == NULL) // Partially dead
176 return false;
177 // Must also check for TOP when looking for a dead loop
178 if (init_control->is_top() || back_control->is_top())
179 return false;
180
181 // Allow funny placement of Safepoint
182 if (back_control->Opcode() == Op_SafePoint)
183 back_control = back_control->in(TypeFunc::Control);
184
185 // Controlling test for loop
186 Node *iftrue = back_control;
187 uint iftrue_op = iftrue->Opcode();
188 if (iftrue_op != Op_IfTrue &&
189 iftrue_op != Op_IfFalse)
190 // I have a weird back-control. Probably the loop-exit test is in
191 // the middle of the loop and I am looking at some trailing control-flow
192 // merge point. To fix this I would have to partially peel the loop.
193 return false; // Obscure back-control
194
195 // Get boolean guarding loop-back test
196 Node *iff = iftrue->in(0);
197 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
198 return false;
199 BoolNode *test = iff->in(1)->as_Bool();
200 BoolTest::mask bt = test->_test._test;
201 float cl_prob = iff->as_If()->_prob;
202 if (iftrue_op == Op_IfFalse) {
203 bt = BoolTest(bt).negate();
204 cl_prob = 1.0 - cl_prob;
205 }
206 // Get backedge compare
207 Node *cmp = test->in(1);
208 int cmp_op = cmp->Opcode();
209 if (cmp_op != Op_CmpI)
210 return false; // Avoid pointer & float compares
211
212 // Find the trip-counter increment & limit. Limit must be loop invariant.
213 Node *incr = cmp->in(1);
214 Node *limit = cmp->in(2);
215
216 // ---------
217 // need 'loop()' test to tell if limit is loop invariant
218 // ---------
219
220 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
221 Node *tmp = incr; // Then reverse order into the CmpI
222 incr = limit;
223 limit = tmp;
224 bt = BoolTest(bt).commute(); // And commute the exit test
225 }
226 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
227 return false;
228 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
229 return false;
230
231 Node* phi_incr = NULL;
232 // Trip-counter increment must be commutative & associative.
233 if (incr->is_Phi()) {
234 if (incr->as_Phi()->region() != x || incr->req() != 3)
235 return false; // Not simple trip counter expression
236 phi_incr = incr;
237 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
238 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
239 return false;
240 }
241
242 Node* trunc1 = NULL;
243 Node* trunc2 = NULL;
244 const TypeInt* iv_trunc_t = NULL;
245 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
246 return false; // Funny increment opcode
247 }
248 assert(incr->Opcode() == Op_AddI, "wrong increment code");
249
250 // Get merge point
251 Node *xphi = incr->in(1);
252 Node *stride = incr->in(2);
253 if (!stride->is_Con()) { // Oops, swap these
254 if (!xphi->is_Con()) // Is the other guy a constant?
255 return false; // Nope, unknown stride, bail out
256 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
257 xphi = stride;
258 stride = tmp;
259 }
260 // Stride must be constant
261 int stride_con = stride->get_int();
262 if (stride_con == 0)
263 return false; // missed some peephole opt
264
265 if (!xphi->is_Phi())
266 return false; // Too much math on the trip counter
267 if (phi_incr != NULL && phi_incr != xphi)
268 return false;
269 PhiNode *phi = xphi->as_Phi();
270
271 // Phi must be of loop header; backedge must wrap to increment
272 if (phi->region() != x)
273 return false;
274 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
275 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
276 return false;
277 }
278 Node *init_trip = phi->in(LoopNode::EntryControl);
279
280 // If iv trunc type is smaller than int, check for possible wrap.
281 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
282 assert(trunc1 != NULL, "must have found some truncation");
283
284 // Get a better type for the phi (filtered thru if's)
285 const TypeInt* phi_ft = filtered_type(phi);
286
287 // Can iv take on a value that will wrap?
288 //
289 // Ensure iv's limit is not within "stride" of the wrap value.
290 //
291 // Example for "short" type
292 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
293 // If the stride is +10, then the last value of the induction
294 // variable before the increment (phi_ft->_hi) must be
295 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
296 // ensure no truncation occurs after the increment.
297
298 if (stride_con > 0) {
299 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
300 iv_trunc_t->_lo > phi_ft->_lo) {
301 return false; // truncation may occur
302 }
303 } else if (stride_con < 0) {
304 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
305 iv_trunc_t->_hi < phi_ft->_hi) {
306 return false; // truncation may occur
307 }
308 }
309 // No possibility of wrap so truncation can be discarded
310 // Promote iv type to Int
311 } else {
312 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
313 }
314
315 // If the condition is inverted and we will be rolling
316 // through MININT to MAXINT, then bail out.
317 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
318 // Odd stride
319 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
320 // Count down loop rolls through MAXINT
321 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
322 // Count up loop rolls through MININT
323 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
324 return false; // Bail out
325 }
326
327 const TypeInt* init_t = gvn->type(init_trip)->is_int();
328 const TypeInt* limit_t = gvn->type(limit)->is_int();
329
330 if (stride_con > 0) {
331 jlong init_p = (jlong)init_t->_lo + stride_con;
332 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
333 return false; // cyclic loop or this loop trips only once
334 } else {
335 jlong init_p = (jlong)init_t->_hi + stride_con;
336 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
337 return false; // cyclic loop or this loop trips only once
338 }
339
340 // =================================================
341 // ---- SUCCESS! Found A Trip-Counted Loop! -----
342 //
343 assert(x->Opcode() == Op_Loop, "regular loops only");
344 C->print_method("Before CountedLoop", 3);
345
346 Node *hook = new (C) Node(6);
347
348 if (LoopLimitCheck) {
349
350 // ===================================================
351 // Generate loop limit check to avoid integer overflow
352 // in cases like next (cyclic loops):
353 //
354 // for (i=0; i <= max_jint; i++) {}
355 // for (i=0; i < max_jint; i+=2) {}
356 //
357 //
358 // Limit check predicate depends on the loop test:
359 //
360 // for(;i != limit; i++) --> limit <= (max_jint)
361 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
362 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
363 //
364
365 // Check if limit is excluded to do more precise int overflow check.
366 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
367 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
368
369 // If compare points directly to the phi we need to adjust
370 // the compare so that it points to the incr. Limit have
371 // to be adjusted to keep trip count the same and the
372 // adjusted limit should be checked for int overflow.
373 if (phi_incr != NULL) {
374 stride_m += stride_con;
375 }
376
377 if (limit->is_Con()) {
378 int limit_con = limit->get_int();
379 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
380 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
381 // Bailout: it could be integer overflow.
382 return false;
383 }
384 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
385 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
386 // Limit's type may satisfy the condition, for example,
387 // when it is an array length.
388 } else {
389 // Generate loop's limit check.
390 // Loop limit check predicate should be near the loop.
391 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
392 if (!limit_check_proj) {
393 // The limit check predicate is not generated if this method trapped here before.
394 #ifdef ASSERT
395 if (TraceLoopLimitCheck) {
396 tty->print("missing loop limit check:");
397 loop->dump_head();
398 x->dump(1);
399 }
400 #endif
401 return false;
402 }
403
404 IfNode* check_iff = limit_check_proj->in(0)->as_If();
405 Node* cmp_limit;
406 Node* bol;
407
408 if (stride_con > 0) {
409 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
410 bol = new (C) BoolNode(cmp_limit, BoolTest::le);
411 } else {
412 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
413 bol = new (C) BoolNode(cmp_limit, BoolTest::ge);
414 }
415 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
416 bol = _igvn.register_new_node_with_optimizer(bol);
417 set_subtree_ctrl(bol);
418
419 // Replace condition in original predicate but preserve Opaque node
420 // so that previous predicates could be found.
421 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
422 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
423 Node* opq = check_iff->in(1)->in(1);
424 _igvn.hash_delete(opq);
425 opq->set_req(1, bol);
426 // Update ctrl.
427 set_ctrl(opq, check_iff->in(0));
428 set_ctrl(check_iff->in(1), check_iff->in(0));
429
430 #ifndef PRODUCT
431 // report that the loop predication has been actually performed
432 // for this loop
433 if (TraceLoopLimitCheck) {
434 tty->print_cr("Counted Loop Limit Check generated:");
435 debug_only( bol->dump(2); )
436 }
437 #endif
438 }
439
440 if (phi_incr != NULL) {
441 // If compare points directly to the phi we need to adjust
442 // the compare so that it points to the incr. Limit have
443 // to be adjusted to keep trip count the same and we
444 // should avoid int overflow.
445 //
446 // i = init; do {} while(i++ < limit);
447 // is converted to
448 // i = init; do {} while(++i < limit+1);
449 //
450 limit = gvn->transform(new (C) AddINode(limit, stride));
451 }
452
453 // Now we need to canonicalize loop condition.
454 if (bt == BoolTest::ne) {
455 assert(stride_con == 1 || stride_con == -1, "simple increment only");
456 // 'ne' can be replaced with 'lt' only when init < limit.
457 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
458 bt = BoolTest::lt;
459 // 'ne' can be replaced with 'gt' only when init > limit.
460 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
461 bt = BoolTest::gt;
462 }
463
464 if (incl_limit) {
465 // The limit check guaranties that 'limit <= (max_jint - stride)' so
466 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
467 //
468 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
469 limit = gvn->transform(new (C) AddINode(limit, one));
470 if (bt == BoolTest::le)
471 bt = BoolTest::lt;
472 else if (bt == BoolTest::ge)
473 bt = BoolTest::gt;
474 else
475 ShouldNotReachHere();
476 }
477 set_subtree_ctrl( limit );
478
479 } else { // LoopLimitCheck
480
481 // If compare points to incr, we are ok. Otherwise the compare
482 // can directly point to the phi; in this case adjust the compare so that
483 // it points to the incr by adjusting the limit.
484 if (cmp->in(1) == phi || cmp->in(2) == phi)
485 limit = gvn->transform(new (C) AddINode(limit,stride));
486
487 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
488 // Final value for iterator should be: trip_count * stride + init_trip.
489 Node *one_p = gvn->intcon( 1);
490 Node *one_m = gvn->intcon(-1);
491
492 Node *trip_count = NULL;
493 switch( bt ) {
494 case BoolTest::eq:
495 ShouldNotReachHere();
496 case BoolTest::ne: // Ahh, the case we desire
497 if (stride_con == 1)
498 trip_count = gvn->transform(new (C) SubINode(limit,init_trip));
499 else if (stride_con == -1)
500 trip_count = gvn->transform(new (C) SubINode(init_trip,limit));
501 else
502 ShouldNotReachHere();
503 set_subtree_ctrl(trip_count);
504 //_loop.map(trip_count->_idx,loop(limit));
505 break;
506 case BoolTest::le: // Maybe convert to '<' case
507 limit = gvn->transform(new (C) AddINode(limit,one_p));
508 set_subtree_ctrl( limit );
509 hook->init_req(4, limit);
510
511 bt = BoolTest::lt;
512 // Make the new limit be in the same loop nest as the old limit
513 //_loop.map(limit->_idx,limit_loop);
514 // Fall into next case
515 case BoolTest::lt: { // Maybe convert to '!=' case
516 if (stride_con < 0) // Count down loop rolls through MAXINT
517 ShouldNotReachHere();
518 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
519 set_subtree_ctrl( range );
520 hook->init_req(0, range);
521
522 Node *bias = gvn->transform(new (C) AddINode(range,stride));
523 set_subtree_ctrl( bias );
524 hook->init_req(1, bias);
525
526 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_m));
527 set_subtree_ctrl( bias1 );
528 hook->init_req(2, bias1);
529
530 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
531 set_subtree_ctrl( trip_count );
532 hook->init_req(3, trip_count);
533 break;
534 }
535
536 case BoolTest::ge: // Maybe convert to '>' case
537 limit = gvn->transform(new (C) AddINode(limit,one_m));
538 set_subtree_ctrl( limit );
539 hook->init_req(4 ,limit);
540
541 bt = BoolTest::gt;
542 // Make the new limit be in the same loop nest as the old limit
543 //_loop.map(limit->_idx,limit_loop);
544 // Fall into next case
545 case BoolTest::gt: { // Maybe convert to '!=' case
546 if (stride_con > 0) // count up loop rolls through MININT
547 ShouldNotReachHere();
548 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
549 set_subtree_ctrl( range );
550 hook->init_req(0, range);
551
552 Node *bias = gvn->transform(new (C) AddINode(range,stride));
553 set_subtree_ctrl( bias );
554 hook->init_req(1, bias);
555
556 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_p));
557 set_subtree_ctrl( bias1 );
558 hook->init_req(2, bias1);
559
560 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
561 set_subtree_ctrl( trip_count );
562 hook->init_req(3, trip_count);
563 break;
564 }
565 } // switch( bt )
566
567 Node *span = gvn->transform(new (C) MulINode(trip_count,stride));
568 set_subtree_ctrl( span );
569 hook->init_req(5, span);
570
571 limit = gvn->transform(new (C) AddINode(span,init_trip));
572 set_subtree_ctrl( limit );
573
574 } // LoopLimitCheck
575
576 // Check for SafePoint on backedge and remove
577 Node *sfpt = x->in(LoopNode::LoopBackControl);
578 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
579 lazy_replace( sfpt, iftrue );
580 if (loop->_safepts != NULL) {
581 loop->_safepts->yank(sfpt);
582 }
583 loop->_tail = iftrue;
584 }
585
586 // Build a canonical trip test.
587 // Clone code, as old values may be in use.
588 incr = incr->clone();
589 incr->set_req(1,phi);
590 incr->set_req(2,stride);
591 incr = _igvn.register_new_node_with_optimizer(incr);
592 set_early_ctrl( incr );
593 _igvn.hash_delete(phi);
594 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
595
596 // If phi type is more restrictive than Int, raise to
597 // Int to prevent (almost) infinite recursion in igvn
598 // which can only handle integer types for constants or minint..maxint.
599 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
600 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
601 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
602 nphi = _igvn.register_new_node_with_optimizer(nphi);
603 set_ctrl(nphi, get_ctrl(phi));
604 _igvn.replace_node(phi, nphi);
605 phi = nphi->as_Phi();
606 }
607 cmp = cmp->clone();
608 cmp->set_req(1,incr);
609 cmp->set_req(2,limit);
610 cmp = _igvn.register_new_node_with_optimizer(cmp);
611 set_ctrl(cmp, iff->in(0));
612
613 test = test->clone()->as_Bool();
614 (*(BoolTest*)&test->_test)._test = bt;
615 test->set_req(1,cmp);
616 _igvn.register_new_node_with_optimizer(test);
617 set_ctrl(test, iff->in(0));
618
619 // Replace the old IfNode with a new LoopEndNode
620 Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
621 IfNode *le = lex->as_If();
622 uint dd = dom_depth(iff);
623 set_idom(le, le->in(0), dd); // Update dominance for loop exit
624 set_loop(le, loop);
625
626 // Get the loop-exit control
627 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
628
629 // Need to swap loop-exit and loop-back control?
630 if (iftrue_op == Op_IfFalse) {
631 Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le));
632 Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le));
633
634 loop->_tail = back_control = ift2;
635 set_loop(ift2, loop);
636 set_loop(iff2, get_loop(iffalse));
637
638 // Lazy update of 'get_ctrl' mechanism.
639 lazy_replace_proj( iffalse, iff2 );
640 lazy_replace_proj( iftrue, ift2 );
641
642 // Swap names
643 iffalse = iff2;
644 iftrue = ift2;
645 } else {
646 _igvn.hash_delete(iffalse);
647 _igvn.hash_delete(iftrue);
648 iffalse->set_req_X( 0, le, &_igvn );
649 iftrue ->set_req_X( 0, le, &_igvn );
650 }
651
652 set_idom(iftrue, le, dd+1);
653 set_idom(iffalse, le, dd+1);
654 assert(iff->outcnt() == 0, "should be dead now");
655 lazy_replace( iff, le ); // fix 'get_ctrl'
656
657 // Now setup a new CountedLoopNode to replace the existing LoopNode
658 CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control);
659 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
660 // The following assert is approximately true, and defines the intention
661 // of can_be_counted_loop. It fails, however, because phase->type
662 // is not yet initialized for this loop and its parts.
663 //assert(l->can_be_counted_loop(this), "sanity");
664 _igvn.register_new_node_with_optimizer(l);
665 set_loop(l, loop);
666 loop->_head = l;
667 // Fix all data nodes placed at the old loop head.
668 // Uses the lazy-update mechanism of 'get_ctrl'.
669 lazy_replace( x, l );
670 set_idom(l, init_control, dom_depth(x));
671
672 // Check for immediately preceding SafePoint and remove
673 Node *sfpt2 = le->in(0);
674 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
675 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
676 if (loop->_safepts != NULL) {
677 loop->_safepts->yank(sfpt2);
678 }
679 }
680
681 // Free up intermediate goo
682 _igvn.remove_dead_node(hook);
683
684 #ifdef ASSERT
685 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
686 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
687 #endif
688 #ifndef PRODUCT
689 if (TraceLoopOpts) {
690 tty->print("Counted ");
691 loop->dump_head();
692 }
693 #endif
694
695 C->print_method("After CountedLoop", 3);
696
697 return true;
698 }
699
700 //----------------------exact_limit-------------------------------------------
701 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
702 assert(loop->_head->is_CountedLoop(), "");
703 CountedLoopNode *cl = loop->_head->as_CountedLoop();
704 assert(cl->is_valid_counted_loop(), "");
705
706 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
707 cl->limit()->Opcode() == Op_LoopLimit) {
708 // Old code has exact limit (it could be incorrect in case of int overflow).
709 // Loop limit is exact with stride == 1. And loop may already have exact limit.
710 return cl->limit();
711 }
712 Node *limit = NULL;
713 #ifdef ASSERT
714 BoolTest::mask bt = cl->loopexit()->test_trip();
715 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
716 #endif
717 if (cl->has_exact_trip_count()) {
718 // Simple case: loop has constant boundaries.
719 // Use jlongs to avoid integer overflow.
720 int stride_con = cl->stride_con();
721 jlong init_con = cl->init_trip()->get_int();
722 jlong limit_con = cl->limit()->get_int();
723 julong trip_cnt = cl->trip_count();
724 jlong final_con = init_con + trip_cnt*stride_con;
725 int final_int = (int)final_con;
726 // The final value should be in integer range since the loop
727 // is counted and the limit was checked for overflow.
728 assert(final_con == (jlong)final_int, "final value should be integer");
729 limit = _igvn.intcon(final_int);
730 } else {
731 // Create new LoopLimit node to get exact limit (final iv value).
732 limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
733 register_new_node(limit, cl->in(LoopNode::EntryControl));
734 }
735 assert(limit != NULL, "sanity");
736 return limit;
737 }
738
739 //------------------------------Ideal------------------------------------------
740 // Return a node which is more "ideal" than the current node.
741 // Attempt to convert into a counted-loop.
742 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
743 if (!can_be_counted_loop(phase)) {
744 phase->C->set_major_progress();
745 }
746 return RegionNode::Ideal(phase, can_reshape);
747 }
748
749
750 //=============================================================================
751 //------------------------------Ideal------------------------------------------
752 // Return a node which is more "ideal" than the current node.
753 // Attempt to convert into a counted-loop.
754 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
755 return RegionNode::Ideal(phase, can_reshape);
756 }
757
758 //------------------------------dump_spec--------------------------------------
759 // Dump special per-node info
760 #ifndef PRODUCT
761 void CountedLoopNode::dump_spec(outputStream *st) const {
762 LoopNode::dump_spec(st);
763 if (stride_is_con()) {
764 st->print("stride: %d ",stride_con());
765 }
766 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
767 if (is_main_loop()) st->print("main of N%d", _idx);
768 if (is_post_loop()) st->print("post of N%d", _main_idx);
769 }
770 #endif
771
772 //=============================================================================
773 int CountedLoopEndNode::stride_con() const {
774 return stride()->bottom_type()->is_int()->get_con();
775 }
776
777 //=============================================================================
778 //------------------------------Value-----------------------------------------
779 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
780 const Type* init_t = phase->type(in(Init));
781 const Type* limit_t = phase->type(in(Limit));
782 const Type* stride_t = phase->type(in(Stride));
783 // Either input is TOP ==> the result is TOP
784 if (init_t == Type::TOP) return Type::TOP;
785 if (limit_t == Type::TOP) return Type::TOP;
786 if (stride_t == Type::TOP) return Type::TOP;
787
788 int stride_con = stride_t->is_int()->get_con();
789 if (stride_con == 1)
790 return NULL; // Identity
791
792 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
793 // Use jlongs to avoid integer overflow.
794 jlong init_con = init_t->is_int()->get_con();
795 jlong limit_con = limit_t->is_int()->get_con();
796 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
797 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
798 jlong final_con = init_con + stride_con*trip_count;
799 int final_int = (int)final_con;
800 // The final value should be in integer range since the loop
801 // is counted and the limit was checked for overflow.
802 assert(final_con == (jlong)final_int, "final value should be integer");
803 return TypeInt::make(final_int);
804 }
805
806 return bottom_type(); // TypeInt::INT
807 }
808
809 //------------------------------Ideal------------------------------------------
810 // Return a node which is more "ideal" than the current node.
811 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
812 if (phase->type(in(Init)) == Type::TOP ||
813 phase->type(in(Limit)) == Type::TOP ||
814 phase->type(in(Stride)) == Type::TOP)
815 return NULL; // Dead
816
817 int stride_con = phase->type(in(Stride))->is_int()->get_con();
818 if (stride_con == 1)
819 return NULL; // Identity
820
821 if (in(Init)->is_Con() && in(Limit)->is_Con())
822 return NULL; // Value
823
824 // Delay following optimizations until all loop optimizations
825 // done to keep Ideal graph simple.
826 if (!can_reshape || phase->C->major_progress())
827 return NULL;
828
829 const TypeInt* init_t = phase->type(in(Init) )->is_int();
830 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
831 int stride_p;
832 jlong lim, ini;
833 julong max;
834 if (stride_con > 0) {
835 stride_p = stride_con;
836 lim = limit_t->_hi;
837 ini = init_t->_lo;
838 max = (julong)max_jint;
839 } else {
840 stride_p = -stride_con;
841 lim = init_t->_hi;
842 ini = limit_t->_lo;
843 max = (julong)min_jint;
844 }
845 julong range = lim - ini + stride_p;
846 if (range <= max) {
847 // Convert to integer expression if it is not overflow.
848 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
849 Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init)));
850 Node *bias = phase->transform(new (phase->C) AddINode(range, stride_m));
851 Node *trip = phase->transform(new (phase->C) DivINode(0, bias, in(Stride)));
852 Node *span = phase->transform(new (phase->C) MulINode(trip, in(Stride)));
853 return new (phase->C) AddINode(span, in(Init)); // exact limit
854 }
855
856 if (is_power_of_2(stride_p) || // divisor is 2^n
857 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
858 // Convert to long expression to avoid integer overflow
859 // and let igvn optimizer convert this division.
860 //
861 Node* init = phase->transform( new (phase->C) ConvI2LNode(in(Init)));
862 Node* limit = phase->transform( new (phase->C) ConvI2LNode(in(Limit)));
863 Node* stride = phase->longcon(stride_con);
864 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
865
866 Node *range = phase->transform(new (phase->C) SubLNode(limit, init));
867 Node *bias = phase->transform(new (phase->C) AddLNode(range, stride_m));
868 Node *span;
869 if (stride_con > 0 && is_power_of_2(stride_p)) {
870 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
871 // and avoid generating rounding for division. Zero trip guard should
872 // guarantee that init < limit but sometimes the guard is missing and
873 // we can get situation when init > limit. Note, for the empty loop
874 // optimization zero trip guard is generated explicitly which leaves
875 // only RCE predicate where exact limit is used and the predicate
876 // will simply fail forcing recompilation.
877 Node* neg_stride = phase->longcon(-stride_con);
878 span = phase->transform(new (phase->C) AndLNode(bias, neg_stride));
879 } else {
880 Node *trip = phase->transform(new (phase->C) DivLNode(0, bias, stride));
881 span = phase->transform(new (phase->C) MulLNode(trip, stride));
882 }
883 // Convert back to int
884 Node *span_int = phase->transform(new (phase->C) ConvL2INode(span));
885 return new (phase->C) AddINode(span_int, in(Init)); // exact limit
886 }
887
888 return NULL; // No progress
889 }
890
891 //------------------------------Identity---------------------------------------
892 // If stride == 1 return limit node.
893 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
894 int stride_con = phase->type(in(Stride))->is_int()->get_con();
895 if (stride_con == 1 || stride_con == -1)
896 return in(Limit);
897 return this;
898 }
899
900 //=============================================================================
901 //----------------------match_incr_with_optional_truncation--------------------
902 // Match increment with optional truncation:
903 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
904 // Return NULL for failure. Success returns the increment node.
905 Node* CountedLoopNode::match_incr_with_optional_truncation(
906 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
907 // Quick cutouts:
908 if (expr == NULL || expr->req() != 3) return NULL;
909
910 Node *t1 = NULL;
911 Node *t2 = NULL;
912 const TypeInt* trunc_t = TypeInt::INT;
913 Node* n1 = expr;
914 int n1op = n1->Opcode();
915
916 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
917 if (n1op == Op_AndI &&
918 n1->in(2)->is_Con() &&
919 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
920 // %%% This check should match any mask of 2**K-1.
921 t1 = n1;
922 n1 = t1->in(1);
923 n1op = n1->Opcode();
924 trunc_t = TypeInt::CHAR;
925 } else if (n1op == Op_RShiftI &&
926 n1->in(1) != NULL &&
927 n1->in(1)->Opcode() == Op_LShiftI &&
928 n1->in(2) == n1->in(1)->in(2) &&
929 n1->in(2)->is_Con()) {
930 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
931 // %%% This check should match any shift in [1..31].
932 if (shift == 16 || shift == 8) {
933 t1 = n1;
934 t2 = t1->in(1);
935 n1 = t2->in(1);
936 n1op = n1->Opcode();
937 if (shift == 16) {
938 trunc_t = TypeInt::SHORT;
939 } else if (shift == 8) {
940 trunc_t = TypeInt::BYTE;
941 }
942 }
943 }
944
945 // If (maybe after stripping) it is an AddI, we won:
946 if (n1op == Op_AddI) {
947 *trunc1 = t1;
948 *trunc2 = t2;
949 *trunc_type = trunc_t;
950 return n1;
951 }
952
953 // failed
954 return NULL;
955 }
956
957
958 //------------------------------filtered_type--------------------------------
959 // Return a type based on condition control flow
960 // A successful return will be a type that is restricted due
961 // to a series of dominating if-tests, such as:
962 // if (i < 10) {
963 // if (i > 0) {
964 // here: "i" type is [1..10)
965 // }
966 // }
967 // or a control flow merge
968 // if (i < 10) {
969 // do {
970 // phi( , ) -- at top of loop type is [min_int..10)
971 // i = ?
972 // } while ( i < 10)
973 //
974 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
975 assert(n && n->bottom_type()->is_int(), "must be int");
976 const TypeInt* filtered_t = NULL;
977 if (!n->is_Phi()) {
978 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
979 filtered_t = filtered_type_from_dominators(n, n_ctrl);
980
981 } else {
982 Node* phi = n->as_Phi();
983 Node* region = phi->in(0);
984 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
985 if (region && region != C->top()) {
986 for (uint i = 1; i < phi->req(); i++) {
987 Node* val = phi->in(i);
988 Node* use_c = region->in(i);
989 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
990 if (val_t != NULL) {
991 if (filtered_t == NULL) {
992 filtered_t = val_t;
993 } else {
994 filtered_t = filtered_t->meet(val_t)->is_int();
995 }
996 }
997 }
998 }
999 }
1000 const TypeInt* n_t = _igvn.type(n)->is_int();
1001 if (filtered_t != NULL) {
1002 n_t = n_t->join(filtered_t)->is_int();
1003 }
1004 return n_t;
1005 }
1006
1007
1008 //------------------------------filtered_type_from_dominators--------------------------------
1009 // Return a possibly more restrictive type for val based on condition control flow of dominators
1010 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1011 if (val->is_Con()) {
1012 return val->bottom_type()->is_int();
1013 }
1014 uint if_limit = 10; // Max number of dominating if's visited
1015 const TypeInt* rtn_t = NULL;
1016
1017 if (use_ctrl && use_ctrl != C->top()) {
1018 Node* val_ctrl = get_ctrl(val);
1019 uint val_dom_depth = dom_depth(val_ctrl);
1020 Node* pred = use_ctrl;
1021 uint if_cnt = 0;
1022 while (if_cnt < if_limit) {
1023 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1024 if_cnt++;
1025 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1026 if (if_t != NULL) {
1027 if (rtn_t == NULL) {
1028 rtn_t = if_t;
1029 } else {
1030 rtn_t = rtn_t->join(if_t)->is_int();
1031 }
1032 }
1033 }
1034 pred = idom(pred);
1035 if (pred == NULL || pred == C->top()) {
1036 break;
1037 }
1038 // Stop if going beyond definition block of val
1039 if (dom_depth(pred) < val_dom_depth) {
1040 break;
1041 }
1042 }
1043 }
1044 return rtn_t;
1045 }
1046
1047
1048 //------------------------------dump_spec--------------------------------------
1049 // Dump special per-node info
1050 #ifndef PRODUCT
1051 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1052 if( in(TestValue)->is_Bool() ) {
1053 BoolTest bt( test_trip()); // Added this for g++.
1054
1055 st->print("[");
1056 bt.dump_on(st);
1057 st->print("]");
1058 }
1059 st->print(" ");
1060 IfNode::dump_spec(st);
1061 }
1062 #endif
1063
1064 //=============================================================================
1065 //------------------------------is_member--------------------------------------
1066 // Is 'l' a member of 'this'?
1067 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1068 while( l->_nest > _nest ) l = l->_parent;
1069 return l == this;
1070 }
1071
1072 //------------------------------set_nest---------------------------------------
1073 // Set loop tree nesting depth. Accumulate _has_call bits.
1074 int IdealLoopTree::set_nest( uint depth ) {
1075 _nest = depth;
1076 int bits = _has_call;
1077 if( _child ) bits |= _child->set_nest(depth+1);
1078 if( bits ) _has_call = 1;
1079 if( _next ) bits |= _next ->set_nest(depth );
1080 return bits;
1081 }
1082
1083 //------------------------------split_fall_in----------------------------------
1084 // Split out multiple fall-in edges from the loop header. Move them to a
1085 // private RegionNode before the loop. This becomes the loop landing pad.
1086 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1087 PhaseIterGVN &igvn = phase->_igvn;
1088 uint i;
1089
1090 // Make a new RegionNode to be the landing pad.
1091 Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 );
1092 phase->set_loop(landing_pad,_parent);
1093 // Gather all the fall-in control paths into the landing pad
1094 uint icnt = fall_in_cnt;
1095 uint oreq = _head->req();
1096 for( i = oreq-1; i>0; i-- )
1097 if( !phase->is_member( this, _head->in(i) ) )
1098 landing_pad->set_req(icnt--,_head->in(i));
1099
1100 // Peel off PhiNode edges as well
1101 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1102 Node *oj = _head->fast_out(j);
1103 if( oj->is_Phi() ) {
1104 PhiNode* old_phi = oj->as_Phi();
1105 assert( old_phi->region() == _head, "" );
1106 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1107 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1108 uint icnt = fall_in_cnt;
1109 for( i = oreq-1; i>0; i-- ) {
1110 if( !phase->is_member( this, _head->in(i) ) ) {
1111 p->init_req(icnt--, old_phi->in(i));
1112 // Go ahead and clean out old edges from old phi
1113 old_phi->del_req(i);
1114 }
1115 }
1116 // Search for CSE's here, because ZKM.jar does a lot of
1117 // loop hackery and we need to be a little incremental
1118 // with the CSE to avoid O(N^2) node blow-up.
1119 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1120 if( p2 ) { // Found CSE
1121 p->destruct(); // Recover useless new node
1122 p = p2; // Use old node
1123 } else {
1124 igvn.register_new_node_with_optimizer(p, old_phi);
1125 }
1126 // Make old Phi refer to new Phi.
1127 old_phi->add_req(p);
1128 // Check for the special case of making the old phi useless and
1129 // disappear it. In JavaGrande I have a case where this useless
1130 // Phi is the loop limit and prevents recognizing a CountedLoop
1131 // which in turn prevents removing an empty loop.
1132 Node *id_old_phi = old_phi->Identity( &igvn );
1133 if( id_old_phi != old_phi ) { // Found a simple identity?
1134 // Note that I cannot call 'replace_node' here, because
1135 // that will yank the edge from old_phi to the Region and
1136 // I'm mid-iteration over the Region's uses.
1137 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1138 Node* use = old_phi->last_out(i);
1139 igvn.rehash_node_delayed(use);
1140 uint uses_found = 0;
1141 for (uint j = 0; j < use->len(); j++) {
1142 if (use->in(j) == old_phi) {
1143 if (j < use->req()) use->set_req (j, id_old_phi);
1144 else use->set_prec(j, id_old_phi);
1145 uses_found++;
1146 }
1147 }
1148 i -= uses_found; // we deleted 1 or more copies of this edge
1149 }
1150 }
1151 igvn._worklist.push(old_phi);
1152 }
1153 }
1154 // Finally clean out the fall-in edges from the RegionNode
1155 for( i = oreq-1; i>0; i-- ) {
1156 if( !phase->is_member( this, _head->in(i) ) ) {
1157 _head->del_req(i);
1158 }
1159 }
1160 // Transform landing pad
1161 igvn.register_new_node_with_optimizer(landing_pad, _head);
1162 // Insert landing pad into the header
1163 _head->add_req(landing_pad);
1164 }
1165
1166 //------------------------------split_outer_loop-------------------------------
1167 // Split out the outermost loop from this shared header.
1168 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1169 PhaseIterGVN &igvn = phase->_igvn;
1170
1171 // Find index of outermost loop; it should also be my tail.
1172 uint outer_idx = 1;
1173 while( _head->in(outer_idx) != _tail ) outer_idx++;
1174
1175 // Make a LoopNode for the outermost loop.
1176 Node *ctl = _head->in(LoopNode::EntryControl);
1177 Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) );
1178 outer = igvn.register_new_node_with_optimizer(outer, _head);
1179 phase->set_created_loop_node();
1180
1181 // Outermost loop falls into '_head' loop
1182 _head->set_req(LoopNode::EntryControl, outer);
1183 _head->del_req(outer_idx);
1184 // Split all the Phis up between '_head' loop and 'outer' loop.
1185 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1186 Node *out = _head->fast_out(j);
1187 if( out->is_Phi() ) {
1188 PhiNode *old_phi = out->as_Phi();
1189 assert( old_phi->region() == _head, "" );
1190 Node *phi = PhiNode::make_blank(outer, old_phi);
1191 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1192 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1193 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1194 // Make old Phi point to new Phi on the fall-in path
1195 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1196 old_phi->del_req(outer_idx);
1197 }
1198 }
1199
1200 // Use the new loop head instead of the old shared one
1201 _head = outer;
1202 phase->set_loop(_head, this);
1203 }
1204
1205 //------------------------------fix_parent-------------------------------------
1206 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1207 loop->_parent = parent;
1208 if( loop->_child ) fix_parent( loop->_child, loop );
1209 if( loop->_next ) fix_parent( loop->_next , parent );
1210 }
1211
1212 //------------------------------estimate_path_freq-----------------------------
1213 static float estimate_path_freq( Node *n ) {
1214 // Try to extract some path frequency info
1215 IfNode *iff;
1216 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1217 uint nop = n->Opcode();
1218 if( nop == Op_SafePoint ) { // Skip any safepoint
1219 n = n->in(0);
1220 continue;
1221 }
1222 if( nop == Op_CatchProj ) { // Get count from a prior call
1223 // Assume call does not always throw exceptions: means the call-site
1224 // count is also the frequency of the fall-through path.
1225 assert( n->is_CatchProj(), "" );
1226 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1227 return 0.0f; // Assume call exception path is rare
1228 Node *call = n->in(0)->in(0)->in(0);
1229 assert( call->is_Call(), "expect a call here" );
1230 const JVMState *jvms = ((CallNode*)call)->jvms();
1231 ciMethodData* methodData = jvms->method()->method_data();
1232 if (!methodData->is_mature()) return 0.0f; // No call-site data
1233 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1234 if ((data == NULL) || !data->is_CounterData()) {
1235 // no call profile available, try call's control input
1236 n = n->in(0);
1237 continue;
1238 }
1239 return data->as_CounterData()->count()/FreqCountInvocations;
1240 }
1241 // See if there's a gating IF test
1242 Node *n_c = n->in(0);
1243 if( !n_c->is_If() ) break; // No estimate available
1244 iff = n_c->as_If();
1245 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1246 // Compute how much count comes on this path
1247 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1248 // Have no count info. Skip dull uncommon-trap like branches.
1249 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1250 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1251 break;
1252 // Skip through never-taken branch; look for a real loop exit.
1253 n = iff->in(0);
1254 }
1255 return 0.0f; // No estimate available
1256 }
1257
1258 //------------------------------merge_many_backedges---------------------------
1259 // Merge all the backedges from the shared header into a private Region.
1260 // Feed that region as the one backedge to this loop.
1261 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1262 uint i;
1263
1264 // Scan for the top 2 hottest backedges
1265 float hotcnt = 0.0f;
1266 float warmcnt = 0.0f;
1267 uint hot_idx = 0;
1268 // Loop starts at 2 because slot 1 is the fall-in path
1269 for( i = 2; i < _head->req(); i++ ) {
1270 float cnt = estimate_path_freq(_head->in(i));
1271 if( cnt > hotcnt ) { // Grab hottest path
1272 warmcnt = hotcnt;
1273 hotcnt = cnt;
1274 hot_idx = i;
1275 } else if( cnt > warmcnt ) { // And 2nd hottest path
1276 warmcnt = cnt;
1277 }
1278 }
1279
1280 // See if the hottest backedge is worthy of being an inner loop
1281 // by being much hotter than the next hottest backedge.
1282 if( hotcnt <= 0.0001 ||
1283 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1284
1285 // Peel out the backedges into a private merge point; peel
1286 // them all except optionally hot_idx.
1287 PhaseIterGVN &igvn = phase->_igvn;
1288
1289 Node *hot_tail = NULL;
1290 // Make a Region for the merge point
1291 Node *r = new (phase->C) RegionNode(1);
1292 for( i = 2; i < _head->req(); i++ ) {
1293 if( i != hot_idx )
1294 r->add_req( _head->in(i) );
1295 else hot_tail = _head->in(i);
1296 }
1297 igvn.register_new_node_with_optimizer(r, _head);
1298 // Plug region into end of loop _head, followed by hot_tail
1299 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1300 _head->set_req(2, r);
1301 if( hot_idx ) _head->add_req(hot_tail);
1302
1303 // Split all the Phis up between '_head' loop and the Region 'r'
1304 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1305 Node *out = _head->fast_out(j);
1306 if( out->is_Phi() ) {
1307 PhiNode* n = out->as_Phi();
1308 igvn.hash_delete(n); // Delete from hash before hacking edges
1309 Node *hot_phi = NULL;
1310 Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type());
1311 // Check all inputs for the ones to peel out
1312 uint j = 1;
1313 for( uint i = 2; i < n->req(); i++ ) {
1314 if( i != hot_idx )
1315 phi->set_req( j++, n->in(i) );
1316 else hot_phi = n->in(i);
1317 }
1318 // Register the phi but do not transform until whole place transforms
1319 igvn.register_new_node_with_optimizer(phi, n);
1320 // Add the merge phi to the old Phi
1321 while( n->req() > 3 ) n->del_req( n->req()-1 );
1322 n->set_req(2, phi);
1323 if( hot_idx ) n->add_req(hot_phi);
1324 }
1325 }
1326
1327
1328 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1329 // of self loop tree. Turn self into a loop headed by _head and with
1330 // tail being the new merge point.
1331 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1332 phase->set_loop(_tail,ilt); // Adjust tail
1333 _tail = r; // Self's tail is new merge point
1334 phase->set_loop(r,this);
1335 ilt->_child = _child; // New guy has my children
1336 _child = ilt; // Self has new guy as only child
1337 ilt->_parent = this; // new guy has self for parent
1338 ilt->_nest = _nest; // Same nesting depth (for now)
1339
1340 // Starting with 'ilt', look for child loop trees using the same shared
1341 // header. Flatten these out; they will no longer be loops in the end.
1342 IdealLoopTree **pilt = &_child;
1343 while( ilt ) {
1344 if( ilt->_head == _head ) {
1345 uint i;
1346 for( i = 2; i < _head->req(); i++ )
1347 if( _head->in(i) == ilt->_tail )
1348 break; // Still a loop
1349 if( i == _head->req() ) { // No longer a loop
1350 // Flatten ilt. Hang ilt's "_next" list from the end of
1351 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1352 IdealLoopTree **cp = &ilt->_child;
1353 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1354 *cp = ilt->_next; // Hang next list at end of child list
1355 *pilt = ilt->_child; // Move child up to replace ilt
1356 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1357 ilt = ilt->_child; // Repeat using new ilt
1358 continue; // do not advance over ilt->_child
1359 }
1360 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1361 phase->set_loop(_head,ilt);
1362 }
1363 pilt = &ilt->_child; // Advance to next
1364 ilt = *pilt;
1365 }
1366
1367 if( _child ) fix_parent( _child, this );
1368 }
1369
1370 //------------------------------beautify_loops---------------------------------
1371 // Split shared headers and insert loop landing pads.
1372 // Insert a LoopNode to replace the RegionNode.
1373 // Return TRUE if loop tree is structurally changed.
1374 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1375 bool result = false;
1376 // Cache parts in locals for easy
1377 PhaseIterGVN &igvn = phase->_igvn;
1378
1379 igvn.hash_delete(_head); // Yank from hash before hacking edges
1380
1381 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1382 int fall_in_cnt = 0;
1383 for( uint i = 1; i < _head->req(); i++ )
1384 if( !phase->is_member( this, _head->in(i) ) )
1385 fall_in_cnt++;
1386 assert( fall_in_cnt, "at least 1 fall-in path" );
1387 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1388 split_fall_in( phase, fall_in_cnt );
1389
1390 // Swap inputs to the _head and all Phis to move the fall-in edge to
1391 // the left.
1392 fall_in_cnt = 1;
1393 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1394 fall_in_cnt++;
1395 if( fall_in_cnt > 1 ) {
1396 // Since I am just swapping inputs I do not need to update def-use info
1397 Node *tmp = _head->in(1);
1398 _head->set_req( 1, _head->in(fall_in_cnt) );
1399 _head->set_req( fall_in_cnt, tmp );
1400 // Swap also all Phis
1401 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1402 Node* phi = _head->fast_out(i);
1403 if( phi->is_Phi() ) {
1404 igvn.hash_delete(phi); // Yank from hash before hacking edges
1405 tmp = phi->in(1);
1406 phi->set_req( 1, phi->in(fall_in_cnt) );
1407 phi->set_req( fall_in_cnt, tmp );
1408 }
1409 }
1410 }
1411 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1412 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1413
1414 // If I am a shared header (multiple backedges), peel off the many
1415 // backedges into a private merge point and use the merge point as
1416 // the one true backedge.
1417 if( _head->req() > 3 ) {
1418 // Merge the many backedges into a single backedge but leave
1419 // the hottest backedge as separate edge for the following peel.
1420 merge_many_backedges( phase );
1421 result = true;
1422 }
1423
1424 // If I have one hot backedge, peel off myself loop.
1425 // I better be the outermost loop.
1426 if( _head->req() > 3 ) {
1427 split_outer_loop( phase );
1428 result = true;
1429
1430 } else if( !_head->is_Loop() && !_irreducible ) {
1431 // Make a new LoopNode to replace the old loop head
1432 Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) );
1433 l = igvn.register_new_node_with_optimizer(l, _head);
1434 phase->set_created_loop_node();
1435 // Go ahead and replace _head
1436 phase->_igvn.replace_node( _head, l );
1437 _head = l;
1438 phase->set_loop(_head, this);
1439 }
1440
1441 // Now recursively beautify nested loops
1442 if( _child ) result |= _child->beautify_loops( phase );
1443 if( _next ) result |= _next ->beautify_loops( phase );
1444 return result;
1445 }
1446
1447 //------------------------------allpaths_check_safepts----------------------------
1448 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1449 // encountered. Helper for check_safepts.
1450 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1451 assert(stack.size() == 0, "empty stack");
1452 stack.push(_tail);
1453 visited.Clear();
1454 visited.set(_tail->_idx);
1455 while (stack.size() > 0) {
1456 Node* n = stack.pop();
1457 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1458 // Terminate this path
1459 } else if (n->Opcode() == Op_SafePoint) {
1460 if (_phase->get_loop(n) != this) {
1461 if (_required_safept == NULL) _required_safept = new Node_List();
1462 _required_safept->push(n); // save the one closest to the tail
1463 }
1464 // Terminate this path
1465 } else {
1466 uint start = n->is_Region() ? 1 : 0;
1467 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1468 for (uint i = start; i < end; i++) {
1469 Node* in = n->in(i);
1470 assert(in->is_CFG(), "must be");
1471 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1472 stack.push(in);
1473 }
1474 }
1475 }
1476 }
1477 }
1478
1479 //------------------------------check_safepts----------------------------
1480 // Given dominators, try to find loops with calls that must always be
1481 // executed (call dominates loop tail). These loops do not need non-call
1482 // safepoints (ncsfpt).
1483 //
1484 // A complication is that a safepoint in a inner loop may be needed
1485 // by an outer loop. In the following, the inner loop sees it has a
1486 // call (block 3) on every path from the head (block 2) to the
1487 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1488 // in block 2, _but_ this leaves the outer loop without a safepoint.
1489 //
1490 // entry 0
1491 // |
1492 // v
1493 // outer 1,2 +->1
1494 // | |
1495 // | v
1496 // | 2<---+ ncsfpt in 2
1497 // |_/|\ |
1498 // | v |
1499 // inner 2,3 / 3 | call in 3
1500 // / | |
1501 // v +--+
1502 // exit 4
1503 //
1504 //
1505 // This method creates a list (_required_safept) of ncsfpt nodes that must
1506 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1507 // is first looked for in the lists for the outer loops of the current loop.
1508 //
1509 // The insights into the problem:
1510 // A) counted loops are okay
1511 // B) innermost loops are okay (only an inner loop can delete
1512 // a ncsfpt needed by an outer loop)
1513 // C) a loop is immune from an inner loop deleting a safepoint
1514 // if the loop has a call on the idom-path
1515 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1516 // idom-path that is not in a nested loop
1517 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1518 // loop needs to be prevented from deletion by an inner loop
1519 //
1520 // There are two analyses:
1521 // 1) The first, and cheaper one, scans the loop body from
1522 // tail to head following the idom (immediate dominator)
1523 // chain, looking for the cases (C,D,E) above.
1524 // Since inner loops are scanned before outer loops, there is summary
1525 // information about inner loops. Inner loops can be skipped over
1526 // when the tail of an inner loop is encountered.
1527 //
1528 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1529 // the idom path (which is rare), scans all predecessor control paths
1530 // from the tail to the head, terminating a path when a call or sfpt
1531 // is encountered, to find the ncsfpt's that are closest to the tail.
1532 //
1533 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1534 // Bottom up traversal
1535 IdealLoopTree* ch = _child;
1536 if (_child) _child->check_safepts(visited, stack);
1537 if (_next) _next ->check_safepts(visited, stack);
1538
1539 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1540 bool has_call = false; // call on dom-path
1541 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1542 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1543 // Scan the dom-path nodes from tail to head
1544 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1545 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1546 has_call = true;
1547 _has_sfpt = 1; // Then no need for a safept!
1548 break;
1549 } else if (n->Opcode() == Op_SafePoint) {
1550 if (_phase->get_loop(n) == this) {
1551 has_local_ncsfpt = true;
1552 break;
1553 }
1554 if (nonlocal_ncsfpt == NULL) {
1555 nonlocal_ncsfpt = n; // save the one closest to the tail
1556 }
1557 } else {
1558 IdealLoopTree* nlpt = _phase->get_loop(n);
1559 if (this != nlpt) {
1560 // If at an inner loop tail, see if the inner loop has already
1561 // recorded seeing a call on the dom-path (and stop.) If not,
1562 // jump to the head of the inner loop.
1563 assert(is_member(nlpt), "nested loop");
1564 Node* tail = nlpt->_tail;
1565 if (tail->in(0)->is_If()) tail = tail->in(0);
1566 if (n == tail) {
1567 // If inner loop has call on dom-path, so does outer loop
1568 if (nlpt->_has_sfpt) {
1569 has_call = true;
1570 _has_sfpt = 1;
1571 break;
1572 }
1573 // Skip to head of inner loop
1574 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1575 n = nlpt->_head;
1576 }
1577 }
1578 }
1579 }
1580 // Record safept's that this loop needs preserved when an
1581 // inner loop attempts to delete it's safepoints.
1582 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1583 if (nonlocal_ncsfpt != NULL) {
1584 if (_required_safept == NULL) _required_safept = new Node_List();
1585 _required_safept->push(nonlocal_ncsfpt);
1586 } else {
1587 // Failed to find a suitable safept on the dom-path. Now use
1588 // an all paths walk from tail to head, looking for safepoints to preserve.
1589 allpaths_check_safepts(visited, stack);
1590 }
1591 }
1592 }
1593 }
1594
1595 //---------------------------is_deleteable_safept----------------------------
1596 // Is safept not required by an outer loop?
1597 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1598 assert(sfpt->Opcode() == Op_SafePoint, "");
1599 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1600 while (lp != NULL) {
1601 Node_List* sfpts = lp->_required_safept;
1602 if (sfpts != NULL) {
1603 for (uint i = 0; i < sfpts->size(); i++) {
1604 if (sfpt == sfpts->at(i))
1605 return false;
1606 }
1607 }
1608 lp = lp->_parent;
1609 }
1610 return true;
1611 }
1612
1613 //---------------------------replace_parallel_iv-------------------------------
1614 // Replace parallel induction variable (parallel to trip counter)
1615 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1616 assert(loop->_head->is_CountedLoop(), "");
1617 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1618 if (!cl->is_valid_counted_loop())
1619 return; // skip malformed counted loop
1620 Node *incr = cl->incr();
1621 if (incr == NULL)
1622 return; // Dead loop?
1623 Node *init = cl->init_trip();
1624 Node *phi = cl->phi();
1625 int stride_con = cl->stride_con();
1626
1627 // Visit all children, looking for Phis
1628 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1629 Node *out = cl->out(i);
1630 // Look for other phis (secondary IVs). Skip dead ones
1631 if (!out->is_Phi() || out == phi || !has_node(out))
1632 continue;
1633 PhiNode* phi2 = out->as_Phi();
1634 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1635 // Look for induction variables of the form: X += constant
1636 if (phi2->region() != loop->_head ||
1637 incr2->req() != 3 ||
1638 incr2->in(1) != phi2 ||
1639 incr2 == incr ||
1640 incr2->Opcode() != Op_AddI ||
1641 !incr2->in(2)->is_Con())
1642 continue;
1643
1644 // Check for parallel induction variable (parallel to trip counter)
1645 // via an affine function. In particular, count-down loops with
1646 // count-up array indices are common. We only RCE references off
1647 // the trip-counter, so we need to convert all these to trip-counter
1648 // expressions.
1649 Node *init2 = phi2->in( LoopNode::EntryControl );
1650 int stride_con2 = incr2->in(2)->get_int();
1651
1652 // The general case here gets a little tricky. We want to find the
1653 // GCD of all possible parallel IV's and make a new IV using this
1654 // GCD for the loop. Then all possible IVs are simple multiples of
1655 // the GCD. In practice, this will cover very few extra loops.
1656 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1657 // where +/-1 is the common case, but other integer multiples are
1658 // also easy to handle.
1659 int ratio_con = stride_con2/stride_con;
1660
1661 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1662 #ifndef PRODUCT
1663 if (TraceLoopOpts) {
1664 tty->print("Parallel IV: %d ", phi2->_idx);
1665 loop->dump_head();
1666 }
1667 #endif
1668 // Convert to using the trip counter. The parallel induction
1669 // variable differs from the trip counter by a loop-invariant
1670 // amount, the difference between their respective initial values.
1671 // It is scaled by the 'ratio_con'.
1672 Node* ratio = _igvn.intcon(ratio_con);
1673 set_ctrl(ratio, C->root());
1674 Node* ratio_init = new (C) MulINode(init, ratio);
1675 _igvn.register_new_node_with_optimizer(ratio_init, init);
1676 set_early_ctrl(ratio_init);
1677 Node* diff = new (C) SubINode(init2, ratio_init);
1678 _igvn.register_new_node_with_optimizer(diff, init2);
1679 set_early_ctrl(diff);
1680 Node* ratio_idx = new (C) MulINode(phi, ratio);
1681 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1682 set_ctrl(ratio_idx, cl);
1683 Node* add = new (C) AddINode(ratio_idx, diff);
1684 _igvn.register_new_node_with_optimizer(add);
1685 set_ctrl(add, cl);
1686 _igvn.replace_node( phi2, add );
1687 // Sometimes an induction variable is unused
1688 if (add->outcnt() == 0) {
1689 _igvn.remove_dead_node(add);
1690 }
1691 --i; // deleted this phi; rescan starting with next position
1692 continue;
1693 }
1694 }
1695 }
1696
1697 //------------------------------counted_loop-----------------------------------
1698 // Convert to counted loops where possible
1699 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1700
1701 // For grins, set the inner-loop flag here
1702 if (!_child) {
1703 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1704 }
1705
1706 if (_head->is_CountedLoop() ||
1707 phase->is_counted_loop(_head, this)) {
1708 _has_sfpt = 1; // Indicate we do not need a safepoint here
1709
1710 // Look for safepoints to remove.
1711 Node_List* sfpts = _safepts;
1712 if (sfpts != NULL) {
1713 for (uint i = 0; i < sfpts->size(); i++) {
1714 Node* n = sfpts->at(i);
1715 assert(phase->get_loop(n) == this, "");
1716 if (phase->is_deleteable_safept(n)) {
1717 phase->lazy_replace(n, n->in(TypeFunc::Control));
1718 }
1719 }
1720 }
1721
1722 // Look for induction variables
1723 phase->replace_parallel_iv(this);
1724
1725 } else if (_parent != NULL && !_irreducible) {
1726 // Not a counted loop.
1727 // Look for a safepoint on the idom-path.
1728 Node* sfpt = tail();
1729 for (; sfpt != _head; sfpt = phase->idom(sfpt)) {
1730 if (sfpt->Opcode() == Op_SafePoint && phase->get_loop(sfpt) == this)
1731 break; // Found one
1732 }
1733 // Delete other safepoints in this loop.
1734 Node_List* sfpts = _safepts;
1735 if (sfpts != NULL && sfpt != _head && sfpt->Opcode() == Op_SafePoint) {
1736 for (uint i = 0; i < sfpts->size(); i++) {
1737 Node* n = sfpts->at(i);
1738 assert(phase->get_loop(n) == this, "");
1739 if (n != sfpt && phase->is_deleteable_safept(n)) {
1740 phase->lazy_replace(n, n->in(TypeFunc::Control));
1741 }
1742 }
1743 }
1744 }
1745
1746 // Recursively
1747 if (_child) _child->counted_loop( phase );
1748 if (_next) _next ->counted_loop( phase );
1749 }
1750
1751 #ifndef PRODUCT
1752 //------------------------------dump_head--------------------------------------
1753 // Dump 1 liner for loop header info
1754 void IdealLoopTree::dump_head( ) const {
1755 for (uint i=0; i<_nest; i++)
1756 tty->print(" ");
1757 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1758 if (_irreducible) tty->print(" IRREDUCIBLE");
1759 Node* entry = _head->in(LoopNode::EntryControl);
1760 if (LoopLimitCheck) {
1761 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1762 if (predicate != NULL ) {
1763 tty->print(" limit_check");
1764 entry = entry->in(0)->in(0);
1765 }
1766 }
1767 if (UseLoopPredicate) {
1768 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1769 if (entry != NULL) {
1770 tty->print(" predicated");
1771 }
1772 }
1773 if (_head->is_CountedLoop()) {
1774 CountedLoopNode *cl = _head->as_CountedLoop();
1775 tty->print(" counted");
1776
1777 Node* init_n = cl->init_trip();
1778 if (init_n != NULL && init_n->is_Con())
1779 tty->print(" [%d,", cl->init_trip()->get_int());
1780 else
1781 tty->print(" [int,");
1782 Node* limit_n = cl->limit();
1783 if (limit_n != NULL && limit_n->is_Con())
1784 tty->print("%d),", cl->limit()->get_int());
1785 else
1786 tty->print("int),");
1787 int stride_con = cl->stride_con();
1788 if (stride_con > 0) tty->print("+");
1789 tty->print("%d", stride_con);
1790
1791 tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
1792
1793 if (cl->is_pre_loop ()) tty->print(" pre" );
1794 if (cl->is_main_loop()) tty->print(" main");
1795 if (cl->is_post_loop()) tty->print(" post");
1796 }
1797 tty->cr();
1798 }
1799
1800 //------------------------------dump-------------------------------------------
1801 // Dump loops by loop tree
1802 void IdealLoopTree::dump( ) const {
1803 dump_head();
1804 if (_child) _child->dump();
1805 if (_next) _next ->dump();
1806 }
1807
1808 #endif
1809
1810 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1811 if (loop == root) {
1812 if (loop->_child != NULL) {
1813 log->begin_head("loop_tree");
1814 log->end_head();
1815 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1816 log->tail("loop_tree");
1817 assert(loop->_next == NULL, "what?");
1818 }
1819 } else {
1820 Node* head = loop->_head;
1821 log->begin_head("loop");
1822 log->print(" idx='%d' ", head->_idx);
1823 if (loop->_irreducible) log->print("irreducible='1' ");
1824 if (head->is_Loop()) {
1825 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1826 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1827 }
1828 if (head->is_CountedLoop()) {
1829 CountedLoopNode* cl = head->as_CountedLoop();
1830 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1831 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1832 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1833 }
1834 log->end_head();
1835 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1836 log->tail("loop");
1837 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1838 }
1839 }
1840
1841 //---------------------collect_potentially_useful_predicates-----------------------
1842 // Helper function to collect potentially useful predicates to prevent them from
1843 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1844 void PhaseIdealLoop::collect_potentially_useful_predicates(
1845 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1846 if (loop->_child) { // child
1847 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1848 }
1849
1850 // self (only loops that we can apply loop predication may use their predicates)
1851 if (loop->_head->is_Loop() &&
1852 !loop->_irreducible &&
1853 !loop->tail()->is_top()) {
1854 LoopNode* lpn = loop->_head->as_Loop();
1855 Node* entry = lpn->in(LoopNode::EntryControl);
1856 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1857 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1858 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1859 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1860 entry = entry->in(0)->in(0);
1861 }
1862 predicate_proj = find_predicate(entry); // Predicate
1863 if (predicate_proj != NULL ) {
1864 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1865 }
1866 }
1867
1868 if (loop->_next) { // sibling
1869 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1870 }
1871 }
1872
1873 //------------------------eliminate_useless_predicates-----------------------------
1874 // Eliminate all inserted predicates if they could not be used by loop predication.
1875 // Note: it will also eliminates loop limits check predicate since it also uses
1876 // Opaque1 node (see Parse::add_predicate()).
1877 void PhaseIdealLoop::eliminate_useless_predicates() {
1878 if (C->predicate_count() == 0)
1879 return; // no predicate left
1880
1881 Unique_Node_List useful_predicates; // to store useful predicates
1882 if (C->has_loops()) {
1883 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1884 }
1885
1886 for (int i = C->predicate_count(); i > 0; i--) {
1887 Node * n = C->predicate_opaque1_node(i-1);
1888 assert(n->Opcode() == Op_Opaque1, "must be");
1889 if (!useful_predicates.member(n)) { // not in the useful list
1890 _igvn.replace_node(n, n->in(1));
1891 }
1892 }
1893 }
1894
1895 //=============================================================================
1896 //----------------------------build_and_optimize-------------------------------
1897 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1898 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1899 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
1900 ResourceMark rm;
1901
1902 int old_progress = C->major_progress();
1903 uint orig_worklist_size = _igvn._worklist.size();
1904
1905 // Reset major-progress flag for the driver's heuristics
1906 C->clear_major_progress();
1907
1908 #ifndef PRODUCT
1909 // Capture for later assert
1910 uint unique = C->unique();
1911 _loop_invokes++;
1912 _loop_work += unique;
1913 #endif
1914
1915 // True if the method has at least 1 irreducible loop
1916 _has_irreducible_loops = false;
1917
1918 _created_loop_node = false;
1919
1920 Arena *a = Thread::current()->resource_area();
1921 VectorSet visited(a);
1922 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1923 _nodes.map(C->unique(), NULL);
1924 memset(_nodes.adr(), 0, wordSize * C->unique());
1925
1926 // Pre-build the top-level outermost loop tree entry
1927 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1928 // Do not need a safepoint at the top level
1929 _ltree_root->_has_sfpt = 1;
1930
1931 // Initialize Dominators.
1932 // Checked in clone_loop_predicate() during beautify_loops().
1933 _idom_size = 0;
1934 _idom = NULL;
1935 _dom_depth = NULL;
1936 _dom_stk = NULL;
1937
1938 // Empty pre-order array
1939 allocate_preorders();
1940
1941 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1942 // IdealLoopTree entries. Data nodes are NOT walked.
1943 build_loop_tree();
1944 // Check for bailout, and return
1945 if (C->failing()) {
1946 return;
1947 }
1948
1949 // No loops after all
1950 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
1951
1952 // There should always be an outer loop containing the Root and Return nodes.
1953 // If not, we have a degenerate empty program. Bail out in this case.
1954 if (!has_node(C->root())) {
1955 if (!_verify_only) {
1956 C->clear_major_progress();
1957 C->record_method_not_compilable("empty program detected during loop optimization");
1958 }
1959 return;
1960 }
1961
1962 // Nothing to do, so get out
1963 if( !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only ) {
1964 _igvn.optimize(); // Cleanup NeverBranches
1965 return;
1966 }
1967
1968 // Set loop nesting depth
1969 _ltree_root->set_nest( 0 );
1970
1971 // Split shared headers and insert loop landing pads.
1972 // Do not bother doing this on the Root loop of course.
1973 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
1974 C->print_method("Before beautify loops", 3);
1975 if( _ltree_root->_child->beautify_loops( this ) ) {
1976 // Re-build loop tree!
1977 _ltree_root->_child = NULL;
1978 _nodes.clear();
1979 reallocate_preorders();
1980 build_loop_tree();
1981 // Check for bailout, and return
1982 if (C->failing()) {
1983 return;
1984 }
1985 // Reset loop nesting depth
1986 _ltree_root->set_nest( 0 );
1987
1988 C->print_method("After beautify loops", 3);
1989 }
1990 }
1991
1992 // Build Dominators for elision of NULL checks & loop finding.
1993 // Since nodes do not have a slot for immediate dominator, make
1994 // a persistent side array for that info indexed on node->_idx.
1995 _idom_size = C->unique();
1996 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1997 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1998 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1999 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2000
2001 Dominators();
2002
2003 if (!_verify_only) {
2004 // As a side effect, Dominators removed any unreachable CFG paths
2005 // into RegionNodes. It doesn't do this test against Root, so
2006 // we do it here.
2007 for( uint i = 1; i < C->root()->req(); i++ ) {
2008 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2009 _igvn.delete_input_of(C->root(), i);
2010 i--; // Rerun same iteration on compressed edges
2011 }
2012 }
2013
2014 // Given dominators, try to find inner loops with calls that must
2015 // always be executed (call dominates loop tail). These loops do
2016 // not need a separate safepoint.
2017 Node_List cisstack(a);
2018 _ltree_root->check_safepts(visited, cisstack);
2019 }
2020
2021 // Walk the DATA nodes and place into loops. Find earliest control
2022 // node. For CFG nodes, the _nodes array starts out and remains
2023 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2024 // _nodes array holds the earliest legal controlling CFG node.
2025
2026 // Allocate stack with enough space to avoid frequent realloc
2027 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
2028 Node_Stack nstack( a, stack_size );
2029
2030 visited.Clear();
2031 Node_List worklist(a);
2032 // Don't need C->root() on worklist since
2033 // it will be processed among C->top() inputs
2034 worklist.push( C->top() );
2035 visited.set( C->top()->_idx ); // Set C->top() as visited now
2036 build_loop_early( visited, worklist, nstack );
2037
2038 // Given early legal placement, try finding counted loops. This placement
2039 // is good enough to discover most loop invariants.
2040 if( !_verify_me && !_verify_only )
2041 _ltree_root->counted_loop( this );
2042
2043 // Find latest loop placement. Find ideal loop placement.
2044 visited.Clear();
2045 init_dom_lca_tags();
2046 // Need C->root() on worklist when processing outs
2047 worklist.push( C->root() );
2048 NOT_PRODUCT( C->verify_graph_edges(); )
2049 worklist.push( C->top() );
2050 build_loop_late( visited, worklist, nstack );
2051
2052 if (_verify_only) {
2053 // restore major progress flag
2054 for (int i = 0; i < old_progress; i++)
2055 C->set_major_progress();
2056 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2057 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2058 return;
2059 }
2060
2061 // Some parser-inserted loop predicates could never be used by loop
2062 // predication or they were moved away from loop during some optimizations.
2063 // For example, peeling. Eliminate them before next loop optimizations.
2064 if (UseLoopPredicate || LoopLimitCheck) {
2065 eliminate_useless_predicates();
2066 }
2067
2068 // clear out the dead code
2069 while(_deadlist.size()) {
2070 _igvn.remove_globally_dead_node(_deadlist.pop());
2071 }
2072
2073 #ifndef PRODUCT
2074 C->verify_graph_edges();
2075 if (_verify_me) { // Nested verify pass?
2076 // Check to see if the verify mode is broken
2077 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2078 return;
2079 }
2080 if(VerifyLoopOptimizations) verify();
2081 if(TraceLoopOpts && C->has_loops()) {
2082 _ltree_root->dump();
2083 }
2084 #endif
2085
2086 if (skip_loop_opts) {
2087 // Cleanup any modified bits
2088 _igvn.optimize();
2089
2090 if (C->log() != NULL) {
2091 log_loop_tree(_ltree_root, _ltree_root, C->log());
2092 }
2093 return;
2094 }
2095
2096 if (ReassociateInvariants) {
2097 // Reassociate invariants and prep for split_thru_phi
2098 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2099 IdealLoopTree* lpt = iter.current();
2100 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2101
2102 lpt->reassociate_invariants(this);
2103
2104 // Because RCE opportunities can be masked by split_thru_phi,
2105 // look for RCE candidates and inhibit split_thru_phi
2106 // on just their loop-phi's for this pass of loop opts
2107 if (SplitIfBlocks && do_split_ifs) {
2108 if (lpt->policy_range_check(this)) {
2109 lpt->_rce_candidate = 1; // = true
2110 }
2111 }
2112 }
2113 }
2114
2115 // Check for aggressive application of split-if and other transforms
2116 // that require basic-block info (like cloning through Phi's)
2117 if( SplitIfBlocks && do_split_ifs ) {
2118 visited.Clear();
2119 split_if_with_blocks( visited, nstack );
2120 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2121 }
2122
2123 // Perform loop predication before iteration splitting
2124 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2125 _ltree_root->_child->loop_predication(this);
2126 }
2127
2128 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2129 if (do_intrinsify_fill()) {
2130 C->set_major_progress();
2131 }
2132 }
2133
2134 // Perform iteration-splitting on inner loops. Split iterations to avoid
2135 // range checks or one-shot null checks.
2136
2137 // If split-if's didn't hack the graph too bad (no CFG changes)
2138 // then do loop opts.
2139 if (C->has_loops() && !C->major_progress()) {
2140 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2141 _ltree_root->_child->iteration_split( this, worklist );
2142 // No verify after peeling! GCM has hoisted code out of the loop.
2143 // After peeling, the hoisted code could sink inside the peeled area.
2144 // The peeling code does not try to recompute the best location for
2145 // all the code before the peeled area, so the verify pass will always
2146 // complain about it.
2147 }
2148 // Do verify graph edges in any case
2149 NOT_PRODUCT( C->verify_graph_edges(); );
2150
2151 if (!do_split_ifs) {
2152 // We saw major progress in Split-If to get here. We forced a
2153 // pass with unrolling and not split-if, however more split-if's
2154 // might make progress. If the unrolling didn't make progress
2155 // then the major-progress flag got cleared and we won't try
2156 // another round of Split-If. In particular the ever-common
2157 // instance-of/check-cast pattern requires at least 2 rounds of
2158 // Split-If to clear out.
2159 C->set_major_progress();
2160 }
2161
2162 // Repeat loop optimizations if new loops were seen
2163 if (created_loop_node()) {
2164 C->set_major_progress();
2165 }
2166
2167 // Keep loop predicates and perform optimizations with them
2168 // until no more loop optimizations could be done.
2169 // After that switch predicates off and do more loop optimizations.
2170 if (!C->major_progress() && (C->predicate_count() > 0)) {
2171 C->cleanup_loop_predicates(_igvn);
2172 #ifndef PRODUCT
2173 if (TraceLoopOpts) {
2174 tty->print_cr("PredicatesOff");
2175 }
2176 #endif
2177 C->set_major_progress();
2178 }
2179
2180 // Convert scalar to superword operations at the end of all loop opts.
2181 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2182 // SuperWord transform
2183 SuperWord sw(this);
2184 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2185 IdealLoopTree* lpt = iter.current();
2186 if (lpt->is_counted()) {
2187 sw.transform_loop(lpt);
2188 }
2189 }
2190 }
2191
2192 // Cleanup any modified bits
2193 _igvn.optimize();
2194
2195 // disable assert until issue with split_flow_path is resolved (6742111)
2196 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2197 // "shouldn't introduce irreducible loops");
2198
2199 if (C->log() != NULL) {
2200 log_loop_tree(_ltree_root, _ltree_root, C->log());
2201 }
2202 }
2203
2204 #ifndef PRODUCT
2205 //------------------------------print_statistics-------------------------------
2206 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2207 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2208 void PhaseIdealLoop::print_statistics() {
2209 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2210 }
2211
2212 //------------------------------verify-----------------------------------------
2213 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2214 static int fail; // debug only, so its multi-thread dont care
2215 void PhaseIdealLoop::verify() const {
2216 int old_progress = C->major_progress();
2217 ResourceMark rm;
2218 PhaseIdealLoop loop_verify( _igvn, this );
2219 VectorSet visited(Thread::current()->resource_area());
2220
2221 fail = 0;
2222 verify_compare( C->root(), &loop_verify, visited );
2223 assert( fail == 0, "verify loops failed" );
2224 // Verify loop structure is the same
2225 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2226 // Reset major-progress. It was cleared by creating a verify version of
2227 // PhaseIdealLoop.
2228 for( int i=0; i<old_progress; i++ )
2229 C->set_major_progress();
2230 }
2231
2232 //------------------------------verify_compare---------------------------------
2233 // Make sure me and the given PhaseIdealLoop agree on key data structures
2234 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2235 if( !n ) return;
2236 if( visited.test_set( n->_idx ) ) return;
2237 if( !_nodes[n->_idx] ) { // Unreachable
2238 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2239 return;
2240 }
2241
2242 uint i;
2243 for( i = 0; i < n->req(); i++ )
2244 verify_compare( n->in(i), loop_verify, visited );
2245
2246 // Check the '_nodes' block/loop structure
2247 i = n->_idx;
2248 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2249 if( _nodes[i] != loop_verify->_nodes[i] &&
2250 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2251 tty->print("Mismatched control setting for: ");
2252 n->dump();
2253 if( fail++ > 10 ) return;
2254 Node *c = get_ctrl_no_update(n);
2255 tty->print("We have it as: ");
2256 if( c->in(0) ) c->dump();
2257 else tty->print_cr("N%d",c->_idx);
2258 tty->print("Verify thinks: ");
2259 if( loop_verify->has_ctrl(n) )
2260 loop_verify->get_ctrl_no_update(n)->dump();
2261 else
2262 loop_verify->get_loop_idx(n)->dump();
2263 tty->cr();
2264 }
2265 } else { // We have a loop
2266 IdealLoopTree *us = get_loop_idx(n);
2267 if( loop_verify->has_ctrl(n) ) {
2268 tty->print("Mismatched loop setting for: ");
2269 n->dump();
2270 if( fail++ > 10 ) return;
2271 tty->print("We have it as: ");
2272 us->dump();
2273 tty->print("Verify thinks: ");
2274 loop_verify->get_ctrl_no_update(n)->dump();
2275 tty->cr();
2276 } else if (!C->major_progress()) {
2277 // Loop selection can be messed up if we did a major progress
2278 // operation, like split-if. Do not verify in that case.
2279 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2280 if( us->_head != them->_head || us->_tail != them->_tail ) {
2281 tty->print("Unequals loops for: ");
2282 n->dump();
2283 if( fail++ > 10 ) return;
2284 tty->print("We have it as: ");
2285 us->dump();
2286 tty->print("Verify thinks: ");
2287 them->dump();
2288 tty->cr();
2289 }
2290 }
2291 }
2292
2293 // Check for immediate dominators being equal
2294 if( i >= _idom_size ) {
2295 if( !n->is_CFG() ) return;
2296 tty->print("CFG Node with no idom: ");
2297 n->dump();
2298 return;
2299 }
2300 if( !n->is_CFG() ) return;
2301 if( n == C->root() ) return; // No IDOM here
2302
2303 assert(n->_idx == i, "sanity");
2304 Node *id = idom_no_update(n);
2305 if( id != loop_verify->idom_no_update(n) ) {
2306 tty->print("Unequals idoms for: ");
2307 n->dump();
2308 if( fail++ > 10 ) return;
2309 tty->print("We have it as: ");
2310 id->dump();
2311 tty->print("Verify thinks: ");
2312 loop_verify->idom_no_update(n)->dump();
2313 tty->cr();
2314 }
2315
2316 }
2317
2318 //------------------------------verify_tree------------------------------------
2319 // Verify that tree structures match. Because the CFG can change, siblings
2320 // within the loop tree can be reordered. We attempt to deal with that by
2321 // reordering the verify's loop tree if possible.
2322 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2323 assert( _parent == parent, "Badly formed loop tree" );
2324
2325 // Siblings not in same order? Attempt to re-order.
2326 if( _head != loop->_head ) {
2327 // Find _next pointer to update
2328 IdealLoopTree **pp = &loop->_parent->_child;
2329 while( *pp != loop )
2330 pp = &((*pp)->_next);
2331 // Find proper sibling to be next
2332 IdealLoopTree **nn = &loop->_next;
2333 while( (*nn) && (*nn)->_head != _head )
2334 nn = &((*nn)->_next);
2335
2336 // Check for no match.
2337 if( !(*nn) ) {
2338 // Annoyingly, irreducible loops can pick different headers
2339 // after a major_progress operation, so the rest of the loop
2340 // tree cannot be matched.
2341 if (_irreducible && Compile::current()->major_progress()) return;
2342 assert( 0, "failed to match loop tree" );
2343 }
2344
2345 // Move (*nn) to (*pp)
2346 IdealLoopTree *hit = *nn;
2347 *nn = hit->_next;
2348 hit->_next = loop;
2349 *pp = loop;
2350 loop = hit;
2351 // Now try again to verify
2352 }
2353
2354 assert( _head == loop->_head , "mismatched loop head" );
2355 Node *tail = _tail; // Inline a non-updating version of
2356 while( !tail->in(0) ) // the 'tail()' call.
2357 tail = tail->in(1);
2358 assert( tail == loop->_tail, "mismatched loop tail" );
2359
2360 // Counted loops that are guarded should be able to find their guards
2361 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2362 CountedLoopNode *cl = _head->as_CountedLoop();
2363 Node *init = cl->init_trip();
2364 Node *ctrl = cl->in(LoopNode::EntryControl);
2365 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2366 Node *iff = ctrl->in(0);
2367 assert( iff->Opcode() == Op_If, "" );
2368 Node *bol = iff->in(1);
2369 assert( bol->Opcode() == Op_Bool, "" );
2370 Node *cmp = bol->in(1);
2371 assert( cmp->Opcode() == Op_CmpI, "" );
2372 Node *add = cmp->in(1);
2373 Node *opaq;
2374 if( add->Opcode() == Op_Opaque1 ) {
2375 opaq = add;
2376 } else {
2377 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2378 assert( add == init, "" );
2379 opaq = cmp->in(2);
2380 }
2381 assert( opaq->Opcode() == Op_Opaque1, "" );
2382
2383 }
2384
2385 if (_child != NULL) _child->verify_tree(loop->_child, this);
2386 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2387 // Innermost loops need to verify loop bodies,
2388 // but only if no 'major_progress'
2389 int fail = 0;
2390 if (!Compile::current()->major_progress() && _child == NULL) {
2391 for( uint i = 0; i < _body.size(); i++ ) {
2392 Node *n = _body.at(i);
2393 if (n->outcnt() == 0) continue; // Ignore dead
2394 uint j;
2395 for( j = 0; j < loop->_body.size(); j++ )
2396 if( loop->_body.at(j) == n )
2397 break;
2398 if( j == loop->_body.size() ) { // Not found in loop body
2399 // Last ditch effort to avoid assertion: Its possible that we
2400 // have some users (so outcnt not zero) but are still dead.
2401 // Try to find from root.
2402 if (Compile::current()->root()->find(n->_idx)) {
2403 fail++;
2404 tty->print("We have that verify does not: ");
2405 n->dump();
2406 }
2407 }
2408 }
2409 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2410 Node *n = loop->_body.at(i2);
2411 if (n->outcnt() == 0) continue; // Ignore dead
2412 uint j;
2413 for( j = 0; j < _body.size(); j++ )
2414 if( _body.at(j) == n )
2415 break;
2416 if( j == _body.size() ) { // Not found in loop body
2417 // Last ditch effort to avoid assertion: Its possible that we
2418 // have some users (so outcnt not zero) but are still dead.
2419 // Try to find from root.
2420 if (Compile::current()->root()->find(n->_idx)) {
2421 fail++;
2422 tty->print("Verify has that we do not: ");
2423 n->dump();
2424 }
2425 }
2426 }
2427 assert( !fail, "loop body mismatch" );
2428 }
2429 }
2430
2431 #endif
2432
2433 //------------------------------set_idom---------------------------------------
2434 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2435 uint idx = d->_idx;
2436 if (idx >= _idom_size) {
2437 uint newsize = _idom_size<<1;
2438 while( idx >= newsize ) {
2439 newsize <<= 1;
2440 }
2441 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2442 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2443 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2444 _idom_size = newsize;
2445 }
2446 _idom[idx] = n;
2447 _dom_depth[idx] = dom_depth;
2448 }
2449
2450 //------------------------------recompute_dom_depth---------------------------------------
2451 // The dominator tree is constructed with only parent pointers.
2452 // This recomputes the depth in the tree by first tagging all
2453 // nodes as "no depth yet" marker. The next pass then runs up
2454 // the dom tree from each node marked "no depth yet", and computes
2455 // the depth on the way back down.
2456 void PhaseIdealLoop::recompute_dom_depth() {
2457 uint no_depth_marker = C->unique();
2458 uint i;
2459 // Initialize depth to "no depth yet"
2460 for (i = 0; i < _idom_size; i++) {
2461 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2462 _dom_depth[i] = no_depth_marker;
2463 }
2464 }
2465 if (_dom_stk == NULL) {
2466 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
2467 if (init_size < 10) init_size = 10;
2468 _dom_stk = new GrowableArray<uint>(init_size);
2469 }
2470 // Compute new depth for each node.
2471 for (i = 0; i < _idom_size; i++) {
2472 uint j = i;
2473 // Run up the dom tree to find a node with a depth
2474 while (_dom_depth[j] == no_depth_marker) {
2475 _dom_stk->push(j);
2476 j = _idom[j]->_idx;
2477 }
2478 // Compute the depth on the way back down this tree branch
2479 uint dd = _dom_depth[j] + 1;
2480 while (_dom_stk->length() > 0) {
2481 uint j = _dom_stk->pop();
2482 _dom_depth[j] = dd;
2483 dd++;
2484 }
2485 }
2486 }
2487
2488 //------------------------------sort-------------------------------------------
2489 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2490 // loop tree, not the root.
2491 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2492 if( !innermost ) return loop; // New innermost loop
2493
2494 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2495 assert( loop_preorder, "not yet post-walked loop" );
2496 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2497 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2498
2499 // Insert at start of list
2500 while( l ) { // Insertion sort based on pre-order
2501 if( l == loop ) return innermost; // Already on list!
2502 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2503 assert( l_preorder, "not yet post-walked l" );
2504 // Check header pre-order number to figure proper nesting
2505 if( loop_preorder > l_preorder )
2506 break; // End of insertion
2507 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2508 // Since I split shared headers, you'd think this could not happen.
2509 // BUT: I must first do the preorder numbering before I can discover I
2510 // have shared headers, so the split headers all get the same preorder
2511 // number as the RegionNode they split from.
2512 if( loop_preorder == l_preorder &&
2513 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2514 break; // Also check for shared headers (same pre#)
2515 pp = &l->_parent; // Chain up list
2516 l = *pp;
2517 }
2518 // Link into list
2519 // Point predecessor to me
2520 *pp = loop;
2521 // Point me to successor
2522 IdealLoopTree *p = loop->_parent;
2523 loop->_parent = l; // Point me to successor
2524 if( p ) sort( p, innermost ); // Insert my parents into list as well
2525 return innermost;
2526 }
2527
2528 //------------------------------build_loop_tree--------------------------------
2529 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2530 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2531 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2532 // tightest enclosing IdealLoopTree for post-walked.
2533 //
2534 // During my forward walk I do a short 1-layer lookahead to see if I can find
2535 // a loop backedge with that doesn't have any work on the backedge. This
2536 // helps me construct nested loops with shared headers better.
2537 //
2538 // Once I've done the forward recursion, I do the post-work. For each child
2539 // I check to see if there is a backedge. Backedges define a loop! I
2540 // insert an IdealLoopTree at the target of the backedge.
2541 //
2542 // During the post-work I also check to see if I have several children
2543 // belonging to different loops. If so, then this Node is a decision point
2544 // where control flow can choose to change loop nests. It is at this
2545 // decision point where I can figure out how loops are nested. At this
2546 // time I can properly order the different loop nests from my children.
2547 // Note that there may not be any backedges at the decision point!
2548 //
2549 // Since the decision point can be far removed from the backedges, I can't
2550 // order my loops at the time I discover them. Thus at the decision point
2551 // I need to inspect loop header pre-order numbers to properly nest my
2552 // loops. This means I need to sort my childrens' loops by pre-order.
2553 // The sort is of size number-of-control-children, which generally limits
2554 // it to size 2 (i.e., I just choose between my 2 target loops).
2555 void PhaseIdealLoop::build_loop_tree() {
2556 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2557 GrowableArray <Node *> bltstack(C->unique() >> 1);
2558 Node *n = C->root();
2559 bltstack.push(n);
2560 int pre_order = 1;
2561 int stack_size;
2562
2563 while ( ( stack_size = bltstack.length() ) != 0 ) {
2564 n = bltstack.top(); // Leave node on stack
2565 if ( !is_visited(n) ) {
2566 // ---- Pre-pass Work ----
2567 // Pre-walked but not post-walked nodes need a pre_order number.
2568
2569 set_preorder_visited( n, pre_order ); // set as visited
2570
2571 // ---- Scan over children ----
2572 // Scan first over control projections that lead to loop headers.
2573 // This helps us find inner-to-outer loops with shared headers better.
2574
2575 // Scan children's children for loop headers.
2576 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2577 Node* m = n->raw_out(i); // Child
2578 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2579 // Scan over children's children to find loop
2580 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2581 Node* l = m->fast_out(j);
2582 if( is_visited(l) && // Been visited?
2583 !is_postvisited(l) && // But not post-visited
2584 get_preorder(l) < pre_order ) { // And smaller pre-order
2585 // Found! Scan the DFS down this path before doing other paths
2586 bltstack.push(m);
2587 break;
2588 }
2589 }
2590 }
2591 }
2592 pre_order++;
2593 }
2594 else if ( !is_postvisited(n) ) {
2595 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2596 // such as com.sun.rsasign.am::a.
2597 // For non-recursive version, first, process current children.
2598 // On next iteration, check if additional children were added.
2599 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2600 Node* u = n->raw_out(k);
2601 if ( u->is_CFG() && !is_visited(u) ) {
2602 bltstack.push(u);
2603 }
2604 }
2605 if ( bltstack.length() == stack_size ) {
2606 // There were no additional children, post visit node now
2607 (void)bltstack.pop(); // Remove node from stack
2608 pre_order = build_loop_tree_impl( n, pre_order );
2609 // Check for bailout
2610 if (C->failing()) {
2611 return;
2612 }
2613 // Check to grow _preorders[] array for the case when
2614 // build_loop_tree_impl() adds new nodes.
2615 check_grow_preorders();
2616 }
2617 }
2618 else {
2619 (void)bltstack.pop(); // Remove post-visited node from stack
2620 }
2621 }
2622 }
2623
2624 //------------------------------build_loop_tree_impl---------------------------
2625 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2626 // ---- Post-pass Work ----
2627 // Pre-walked but not post-walked nodes need a pre_order number.
2628
2629 // Tightest enclosing loop for this Node
2630 IdealLoopTree *innermost = NULL;
2631
2632 // For all children, see if any edge is a backedge. If so, make a loop
2633 // for it. Then find the tightest enclosing loop for the self Node.
2634 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2635 Node* m = n->fast_out(i); // Child
2636 if( n == m ) continue; // Ignore control self-cycles
2637 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2638
2639 IdealLoopTree *l; // Child's loop
2640 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2641 // Found a backedge
2642 assert( get_preorder(m) < pre_order, "should be backedge" );
2643 // Check for the RootNode, which is already a LoopNode and is allowed
2644 // to have multiple "backedges".
2645 if( m == C->root()) { // Found the root?
2646 l = _ltree_root; // Root is the outermost LoopNode
2647 } else { // Else found a nested loop
2648 // Insert a LoopNode to mark this loop.
2649 l = new IdealLoopTree(this, m, n);
2650 } // End of Else found a nested loop
2651 if( !has_loop(m) ) // If 'm' does not already have a loop set
2652 set_loop(m, l); // Set loop header to loop now
2653
2654 } else { // Else not a nested loop
2655 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2656 l = get_loop(m); // Get previously determined loop
2657 // If successor is header of a loop (nest), move up-loop till it
2658 // is a member of some outer enclosing loop. Since there are no
2659 // shared headers (I've split them already) I only need to go up
2660 // at most 1 level.
2661 while( l && l->_head == m ) // Successor heads loop?
2662 l = l->_parent; // Move up 1 for me
2663 // If this loop is not properly parented, then this loop
2664 // has no exit path out, i.e. its an infinite loop.
2665 if( !l ) {
2666 // Make loop "reachable" from root so the CFG is reachable. Basically
2667 // insert a bogus loop exit that is never taken. 'm', the loop head,
2668 // points to 'n', one (of possibly many) fall-in paths. There may be
2669 // many backedges as well.
2670
2671 // Here I set the loop to be the root loop. I could have, after
2672 // inserting a bogus loop exit, restarted the recursion and found my
2673 // new loop exit. This would make the infinite loop a first-class
2674 // loop and it would then get properly optimized. What's the use of
2675 // optimizing an infinite loop?
2676 l = _ltree_root; // Oops, found infinite loop
2677
2678 if (!_verify_only) {
2679 // Insert the NeverBranch between 'm' and it's control user.
2680 NeverBranchNode *iff = new (C) NeverBranchNode( m );
2681 _igvn.register_new_node_with_optimizer(iff);
2682 set_loop(iff, l);
2683 Node *if_t = new (C) CProjNode( iff, 0 );
2684 _igvn.register_new_node_with_optimizer(if_t);
2685 set_loop(if_t, l);
2686
2687 Node* cfg = NULL; // Find the One True Control User of m
2688 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2689 Node* x = m->fast_out(j);
2690 if (x->is_CFG() && x != m && x != iff)
2691 { cfg = x; break; }
2692 }
2693 assert(cfg != NULL, "must find the control user of m");
2694 uint k = 0; // Probably cfg->in(0)
2695 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2696 cfg->set_req( k, if_t ); // Now point to NeverBranch
2697
2698 // Now create the never-taken loop exit
2699 Node *if_f = new (C) CProjNode( iff, 1 );
2700 _igvn.register_new_node_with_optimizer(if_f);
2701 set_loop(if_f, l);
2702 // Find frame ptr for Halt. Relies on the optimizer
2703 // V-N'ing. Easier and quicker than searching through
2704 // the program structure.
2705 Node *frame = new (C) ParmNode( C->start(), TypeFunc::FramePtr );
2706 _igvn.register_new_node_with_optimizer(frame);
2707 // Halt & Catch Fire
2708 Node *halt = new (C) HaltNode( if_f, frame );
2709 _igvn.register_new_node_with_optimizer(halt);
2710 set_loop(halt, l);
2711 C->root()->add_req(halt);
2712 }
2713 set_loop(C->root(), _ltree_root);
2714 }
2715 }
2716 // Weeny check for irreducible. This child was already visited (this
2717 // IS the post-work phase). Is this child's loop header post-visited
2718 // as well? If so, then I found another entry into the loop.
2719 if (!_verify_only) {
2720 while( is_postvisited(l->_head) ) {
2721 // found irreducible
2722 l->_irreducible = 1; // = true
2723 l = l->_parent;
2724 _has_irreducible_loops = true;
2725 // Check for bad CFG here to prevent crash, and bailout of compile
2726 if (l == NULL) {
2727 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2728 return pre_order;
2729 }
2730 }
2731 }
2732
2733 // This Node might be a decision point for loops. It is only if
2734 // it's children belong to several different loops. The sort call
2735 // does a trivial amount of work if there is only 1 child or all
2736 // children belong to the same loop. If however, the children
2737 // belong to different loops, the sort call will properly set the
2738 // _parent pointers to show how the loops nest.
2739 //
2740 // In any case, it returns the tightest enclosing loop.
2741 innermost = sort( l, innermost );
2742 }
2743
2744 // Def-use info will have some dead stuff; dead stuff will have no
2745 // loop decided on.
2746
2747 // Am I a loop header? If so fix up my parent's child and next ptrs.
2748 if( innermost && innermost->_head == n ) {
2749 assert( get_loop(n) == innermost, "" );
2750 IdealLoopTree *p = innermost->_parent;
2751 IdealLoopTree *l = innermost;
2752 while( p && l->_head == n ) {
2753 l->_next = p->_child; // Put self on parents 'next child'
2754 p->_child = l; // Make self as first child of parent
2755 l = p; // Now walk up the parent chain
2756 p = l->_parent;
2757 }
2758 } else {
2759 // Note that it is possible for a LoopNode to reach here, if the
2760 // backedge has been made unreachable (hence the LoopNode no longer
2761 // denotes a Loop, and will eventually be removed).
2762
2763 // Record tightest enclosing loop for self. Mark as post-visited.
2764 set_loop(n, innermost);
2765 // Also record has_call flag early on
2766 if( innermost ) {
2767 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2768 // Do not count uncommon calls
2769 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2770 Node *iff = n->in(0)->in(0);
2771 // No any calls for vectorized loops.
2772 if( UseSuperWord || !iff->is_If() ||
2773 (n->in(0)->Opcode() == Op_IfFalse &&
2774 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2775 (iff->as_If()->_prob >= 0.01) )
2776 innermost->_has_call = 1;
2777 }
2778 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2779 // Disable loop optimizations if the loop has a scalar replaceable
2780 // allocation. This disabling may cause a potential performance lost
2781 // if the allocation is not eliminated for some reason.
2782 innermost->_allow_optimizations = false;
2783 innermost->_has_call = 1; // = true
2784 } else if (n->Opcode() == Op_SafePoint) {
2785 // Record all safepoints in this loop.
2786 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
2787 innermost->_safepts->push(n);
2788 }
2789 }
2790 }
2791
2792 // Flag as post-visited now
2793 set_postvisited(n);
2794 return pre_order;
2795 }
2796
2797
2798 //------------------------------build_loop_early-------------------------------
2799 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2800 // First pass computes the earliest controlling node possible. This is the
2801 // controlling input with the deepest dominating depth.
2802 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2803 while (worklist.size() != 0) {
2804 // Use local variables nstack_top_n & nstack_top_i to cache values
2805 // on nstack's top.
2806 Node *nstack_top_n = worklist.pop();
2807 uint nstack_top_i = 0;
2808 //while_nstack_nonempty:
2809 while (true) {
2810 // Get parent node and next input's index from stack's top.
2811 Node *n = nstack_top_n;
2812 uint i = nstack_top_i;
2813 uint cnt = n->req(); // Count of inputs
2814 if (i == 0) { // Pre-process the node.
2815 if( has_node(n) && // Have either loop or control already?
2816 !has_ctrl(n) ) { // Have loop picked out already?
2817 // During "merge_many_backedges" we fold up several nested loops
2818 // into a single loop. This makes the members of the original
2819 // loop bodies pointing to dead loops; they need to move up
2820 // to the new UNION'd larger loop. I set the _head field of these
2821 // dead loops to NULL and the _parent field points to the owning
2822 // loop. Shades of UNION-FIND algorithm.
2823 IdealLoopTree *ilt;
2824 while( !(ilt = get_loop(n))->_head ) {
2825 // Normally I would use a set_loop here. But in this one special
2826 // case, it is legal (and expected) to change what loop a Node
2827 // belongs to.
2828 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2829 }
2830 // Remove safepoints ONLY if I've already seen I don't need one.
2831 // (the old code here would yank a 2nd safepoint after seeing a
2832 // first one, even though the 1st did not dominate in the loop body
2833 // and thus could be avoided indefinitely)
2834 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2835 is_deleteable_safept(n)) {
2836 Node *in = n->in(TypeFunc::Control);
2837 lazy_replace(n,in); // Pull safepoint now
2838 if (ilt->_safepts != NULL) {
2839 ilt->_safepts->yank(n);
2840 }
2841 // Carry on with the recursion "as if" we are walking
2842 // only the control input
2843 if( !visited.test_set( in->_idx ) ) {
2844 worklist.push(in); // Visit this guy later, using worklist
2845 }
2846 // Get next node from nstack:
2847 // - skip n's inputs processing by setting i > cnt;
2848 // - we also will not call set_early_ctrl(n) since
2849 // has_node(n) == true (see the condition above).
2850 i = cnt + 1;
2851 }
2852 }
2853 } // if (i == 0)
2854
2855 // Visit all inputs
2856 bool done = true; // Assume all n's inputs will be processed
2857 while (i < cnt) {
2858 Node *in = n->in(i);
2859 ++i;
2860 if (in == NULL) continue;
2861 if (in->pinned() && !in->is_CFG())
2862 set_ctrl(in, in->in(0));
2863 int is_visited = visited.test_set( in->_idx );
2864 if (!has_node(in)) { // No controlling input yet?
2865 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2866 assert( !is_visited, "visit only once" );
2867 nstack.push(n, i); // Save parent node and next input's index.
2868 nstack_top_n = in; // Process current input now.
2869 nstack_top_i = 0;
2870 done = false; // Not all n's inputs processed.
2871 break; // continue while_nstack_nonempty;
2872 } else if (!is_visited) {
2873 // This guy has a location picked out for him, but has not yet
2874 // been visited. Happens to all CFG nodes, for instance.
2875 // Visit him using the worklist instead of recursion, to break
2876 // cycles. Since he has a location already we do not need to
2877 // find his location before proceeding with the current Node.
2878 worklist.push(in); // Visit this guy later, using worklist
2879 }
2880 }
2881 if (done) {
2882 // All of n's inputs have been processed, complete post-processing.
2883
2884 // Compute earliest point this Node can go.
2885 // CFG, Phi, pinned nodes already know their controlling input.
2886 if (!has_node(n)) {
2887 // Record earliest legal location
2888 set_early_ctrl( n );
2889 }
2890 if (nstack.is_empty()) {
2891 // Finished all nodes on stack.
2892 // Process next node on the worklist.
2893 break;
2894 }
2895 // Get saved parent node and next input's index.
2896 nstack_top_n = nstack.node();
2897 nstack_top_i = nstack.index();
2898 nstack.pop();
2899 }
2900 } // while (true)
2901 }
2902 }
2903
2904 //------------------------------dom_lca_internal--------------------------------
2905 // Pair-wise LCA
2906 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2907 if( !n1 ) return n2; // Handle NULL original LCA
2908 assert( n1->is_CFG(), "" );
2909 assert( n2->is_CFG(), "" );
2910 // find LCA of all uses
2911 uint d1 = dom_depth(n1);
2912 uint d2 = dom_depth(n2);
2913 while (n1 != n2) {
2914 if (d1 > d2) {
2915 n1 = idom(n1);
2916 d1 = dom_depth(n1);
2917 } else if (d1 < d2) {
2918 n2 = idom(n2);
2919 d2 = dom_depth(n2);
2920 } else {
2921 // Here d1 == d2. Due to edits of the dominator-tree, sections
2922 // of the tree might have the same depth. These sections have
2923 // to be searched more carefully.
2924
2925 // Scan up all the n1's with equal depth, looking for n2.
2926 Node *t1 = idom(n1);
2927 while (dom_depth(t1) == d1) {
2928 if (t1 == n2) return n2;
2929 t1 = idom(t1);
2930 }
2931 // Scan up all the n2's with equal depth, looking for n1.
2932 Node *t2 = idom(n2);
2933 while (dom_depth(t2) == d2) {
2934 if (t2 == n1) return n1;
2935 t2 = idom(t2);
2936 }
2937 // Move up to a new dominator-depth value as well as up the dom-tree.
2938 n1 = t1;
2939 n2 = t2;
2940 d1 = dom_depth(n1);
2941 d2 = dom_depth(n2);
2942 }
2943 }
2944 return n1;
2945 }
2946
2947 //------------------------------compute_idom-----------------------------------
2948 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2949 // IDOMs are correct.
2950 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2951 assert( region->is_Region(), "" );
2952 Node *LCA = NULL;
2953 for( uint i = 1; i < region->req(); i++ ) {
2954 if( region->in(i) != C->top() )
2955 LCA = dom_lca( LCA, region->in(i) );
2956 }
2957 return LCA;
2958 }
2959
2960 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
2961 bool had_error = false;
2962 #ifdef ASSERT
2963 if (early != C->root()) {
2964 // Make sure that there's a dominance path from use to LCA
2965 Node* d = use;
2966 while (d != LCA) {
2967 d = idom(d);
2968 if (d == C->root()) {
2969 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
2970 n->dump();
2971 use->dump();
2972 had_error = true;
2973 break;
2974 }
2975 }
2976 }
2977 #endif
2978 return had_error;
2979 }
2980
2981
2982 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
2983 // Compute LCA over list of uses
2984 bool had_error = false;
2985 Node *LCA = NULL;
2986 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2987 Node* c = n->fast_out(i);
2988 if (_nodes[c->_idx] == NULL)
2989 continue; // Skip the occasional dead node
2990 if( c->is_Phi() ) { // For Phis, we must land above on the path
2991 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2992 if( c->in(j) == n ) { // Found matching input?
2993 Node *use = c->in(0)->in(j);
2994 if (_verify_only && use->is_top()) continue;
2995 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2996 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2997 }
2998 }
2999 } else {
3000 // For CFG data-users, use is in the block just prior
3001 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3002 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3003 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3004 }
3005 }
3006 assert(!had_error, "bad dominance");
3007 return LCA;
3008 }
3009
3010 //------------------------------get_late_ctrl----------------------------------
3011 // Compute latest legal control.
3012 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3013 assert(early != NULL, "early control should not be NULL");
3014
3015 Node* LCA = compute_lca_of_uses(n, early);
3016 #ifdef ASSERT
3017 if (LCA == C->root() && LCA != early) {
3018 // def doesn't dominate uses so print some useful debugging output
3019 compute_lca_of_uses(n, early, true);
3020 }
3021 #endif
3022
3023 // if this is a load, check for anti-dependent stores
3024 // We use a conservative algorithm to identify potential interfering
3025 // instructions and for rescheduling the load. The users of the memory
3026 // input of this load are examined. Any use which is not a load and is
3027 // dominated by early is considered a potentially interfering store.
3028 // This can produce false positives.
3029 if (n->is_Load() && LCA != early) {
3030 Node_List worklist;
3031
3032 Node *mem = n->in(MemNode::Memory);
3033 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3034 Node* s = mem->fast_out(i);
3035 worklist.push(s);
3036 }
3037 while(worklist.size() != 0 && LCA != early) {
3038 Node* s = worklist.pop();
3039 if (s->is_Load()) {
3040 continue;
3041 } else if (s->is_MergeMem()) {
3042 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3043 Node* s1 = s->fast_out(i);
3044 worklist.push(s1);
3045 }
3046 } else {
3047 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3048 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3049 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3050 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3051 }
3052 }
3053 }
3054 }
3055
3056 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3057 return LCA;
3058 }
3059
3060 // true if CFG node d dominates CFG node n
3061 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3062 if (d == n)
3063 return true;
3064 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3065 uint dd = dom_depth(d);
3066 while (dom_depth(n) >= dd) {
3067 if (n == d)
3068 return true;
3069 n = idom(n);
3070 }
3071 return false;
3072 }
3073
3074 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3075 // Pair-wise LCA with tags.
3076 // Tag each index with the node 'tag' currently being processed
3077 // before advancing up the dominator chain using idom().
3078 // Later calls that find a match to 'tag' know that this path has already
3079 // been considered in the current LCA (which is input 'n1' by convention).
3080 // Since get_late_ctrl() is only called once for each node, the tag array
3081 // does not need to be cleared between calls to get_late_ctrl().
3082 // Algorithm trades a larger constant factor for better asymptotic behavior
3083 //
3084 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3085 uint d1 = dom_depth(n1);
3086 uint d2 = dom_depth(n2);
3087
3088 do {
3089 if (d1 > d2) {
3090 // current lca is deeper than n2
3091 _dom_lca_tags.map(n1->_idx, tag);
3092 n1 = idom(n1);
3093 d1 = dom_depth(n1);
3094 } else if (d1 < d2) {
3095 // n2 is deeper than current lca
3096 Node *memo = _dom_lca_tags[n2->_idx];
3097 if( memo == tag ) {
3098 return n1; // Return the current LCA
3099 }
3100 _dom_lca_tags.map(n2->_idx, tag);
3101 n2 = idom(n2);
3102 d2 = dom_depth(n2);
3103 } else {
3104 // Here d1 == d2. Due to edits of the dominator-tree, sections
3105 // of the tree might have the same depth. These sections have
3106 // to be searched more carefully.
3107
3108 // Scan up all the n1's with equal depth, looking for n2.
3109 _dom_lca_tags.map(n1->_idx, tag);
3110 Node *t1 = idom(n1);
3111 while (dom_depth(t1) == d1) {
3112 if (t1 == n2) return n2;
3113 _dom_lca_tags.map(t1->_idx, tag);
3114 t1 = idom(t1);
3115 }
3116 // Scan up all the n2's with equal depth, looking for n1.
3117 _dom_lca_tags.map(n2->_idx, tag);
3118 Node *t2 = idom(n2);
3119 while (dom_depth(t2) == d2) {
3120 if (t2 == n1) return n1;
3121 _dom_lca_tags.map(t2->_idx, tag);
3122 t2 = idom(t2);
3123 }
3124 // Move up to a new dominator-depth value as well as up the dom-tree.
3125 n1 = t1;
3126 n2 = t2;
3127 d1 = dom_depth(n1);
3128 d2 = dom_depth(n2);
3129 }
3130 } while (n1 != n2);
3131 return n1;
3132 }
3133
3134 //------------------------------init_dom_lca_tags------------------------------
3135 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3136 // Intended use does not involve any growth for the array, so it could
3137 // be of fixed size.
3138 void PhaseIdealLoop::init_dom_lca_tags() {
3139 uint limit = C->unique() + 1;
3140 _dom_lca_tags.map( limit, NULL );
3141 #ifdef ASSERT
3142 for( uint i = 0; i < limit; ++i ) {
3143 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3144 }
3145 #endif // ASSERT
3146 }
3147
3148 //------------------------------clear_dom_lca_tags------------------------------
3149 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3150 // Intended use does not involve any growth for the array, so it could
3151 // be of fixed size.
3152 void PhaseIdealLoop::clear_dom_lca_tags() {
3153 uint limit = C->unique() + 1;
3154 _dom_lca_tags.map( limit, NULL );
3155 _dom_lca_tags.clear();
3156 #ifdef ASSERT
3157 for( uint i = 0; i < limit; ++i ) {
3158 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3159 }
3160 #endif // ASSERT
3161 }
3162
3163 //------------------------------build_loop_late--------------------------------
3164 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3165 // Second pass finds latest legal placement, and ideal loop placement.
3166 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3167 while (worklist.size() != 0) {
3168 Node *n = worklist.pop();
3169 // Only visit once
3170 if (visited.test_set(n->_idx)) continue;
3171 uint cnt = n->outcnt();
3172 uint i = 0;
3173 while (true) {
3174 assert( _nodes[n->_idx], "no dead nodes" );
3175 // Visit all children
3176 if (i < cnt) {
3177 Node* use = n->raw_out(i);
3178 ++i;
3179 // Check for dead uses. Aggressively prune such junk. It might be
3180 // dead in the global sense, but still have local uses so I cannot
3181 // easily call 'remove_dead_node'.
3182 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3183 // Due to cycles, we might not hit the same fixed point in the verify
3184 // pass as we do in the regular pass. Instead, visit such phis as
3185 // simple uses of the loop head.
3186 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3187 if( !visited.test(use->_idx) )
3188 worklist.push(use);
3189 } else if( !visited.test_set(use->_idx) ) {
3190 nstack.push(n, i); // Save parent and next use's index.
3191 n = use; // Process all children of current use.
3192 cnt = use->outcnt();
3193 i = 0;
3194 }
3195 } else {
3196 // Do not visit around the backedge of loops via data edges.
3197 // push dead code onto a worklist
3198 _deadlist.push(use);
3199 }
3200 } else {
3201 // All of n's children have been processed, complete post-processing.
3202 build_loop_late_post(n);
3203 if (nstack.is_empty()) {
3204 // Finished all nodes on stack.
3205 // Process next node on the worklist.
3206 break;
3207 }
3208 // Get saved parent node and next use's index. Visit the rest of uses.
3209 n = nstack.node();
3210 cnt = n->outcnt();
3211 i = nstack.index();
3212 nstack.pop();
3213 }
3214 }
3215 }
3216 }
3217
3218 //------------------------------build_loop_late_post---------------------------
3219 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3220 // Second pass finds latest legal placement, and ideal loop placement.
3221 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3222
3223 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3224 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3225 }
3226
3227 // CFG and pinned nodes already handled
3228 if( n->in(0) ) {
3229 if( n->in(0)->is_top() ) return; // Dead?
3230
3231 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3232 // _must_ be pinned (they have to observe their control edge of course).
3233 // Unlike Stores (which modify an unallocable resource, the memory
3234 // state), Mods/Loads can float around. So free them up.
3235 bool pinned = true;
3236 switch( n->Opcode() ) {
3237 case Op_DivI:
3238 case Op_DivF:
3239 case Op_DivD:
3240 case Op_ModI:
3241 case Op_ModF:
3242 case Op_ModD:
3243 case Op_LoadB: // Same with Loads; they can sink
3244 case Op_LoadUB: // during loop optimizations.
3245 case Op_LoadUS:
3246 case Op_LoadD:
3247 case Op_LoadF:
3248 case Op_LoadI:
3249 case Op_LoadKlass:
3250 case Op_LoadNKlass:
3251 case Op_LoadL:
3252 case Op_LoadS:
3253 case Op_LoadP:
3254 case Op_LoadN:
3255 case Op_LoadRange:
3256 case Op_LoadD_unaligned:
3257 case Op_LoadL_unaligned:
3258 case Op_StrComp: // Does a bunch of load-like effects
3259 case Op_StrEquals:
3260 case Op_StrIndexOf:
3261 case Op_AryEq:
3262 pinned = false;
3263 }
3264 if( pinned ) {
3265 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3266 if( !chosen_loop->_child ) // Inner loop?
3267 chosen_loop->_body.push(n); // Collect inner loops
3268 return;
3269 }
3270 } else { // No slot zero
3271 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3272 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3273 return;
3274 }
3275 assert(!n->is_CFG() || n->outcnt() == 0, "");
3276 }
3277
3278 // Do I have a "safe range" I can select over?
3279 Node *early = get_ctrl(n);// Early location already computed
3280
3281 // Compute latest point this Node can go
3282 Node *LCA = get_late_ctrl( n, early );
3283 // LCA is NULL due to uses being dead
3284 if( LCA == NULL ) {
3285 #ifdef ASSERT
3286 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3287 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3288 }
3289 #endif
3290 _nodes.map(n->_idx, 0); // This node is useless
3291 _deadlist.push(n);
3292 return;
3293 }
3294 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3295
3296 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3297 Node *least = legal; // Best legal position so far
3298 while( early != legal ) { // While not at earliest legal
3299 #ifdef ASSERT
3300 if (legal->is_Start() && !early->is_Root()) {
3301 // Bad graph. Print idom path and fail.
3302 dump_bad_graph(n, early, LCA);
3303 assert(false, "Bad graph detected in build_loop_late");
3304 }
3305 #endif
3306 // Find least loop nesting depth
3307 legal = idom(legal); // Bump up the IDOM tree
3308 // Check for lower nesting depth
3309 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3310 least = legal;
3311 }
3312 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3313
3314 // Try not to place code on a loop entry projection
3315 // which can inhibit range check elimination.
3316 if (least != early) {
3317 Node* ctrl_out = least->unique_ctrl_out();
3318 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3319 least == ctrl_out->in(LoopNode::EntryControl)) {
3320 Node* least_dom = idom(least);
3321 if (get_loop(least_dom)->is_member(get_loop(least))) {
3322 least = least_dom;
3323 }
3324 }
3325 }
3326
3327 #ifdef ASSERT
3328 // If verifying, verify that 'verify_me' has a legal location
3329 // and choose it as our location.
3330 if( _verify_me ) {
3331 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3332 Node *legal = LCA;
3333 while( early != legal ) { // While not at earliest legal
3334 if( legal == v_ctrl ) break; // Check for prior good location
3335 legal = idom(legal) ;// Bump up the IDOM tree
3336 }
3337 // Check for prior good location
3338 if( legal == v_ctrl ) least = legal; // Keep prior if found
3339 }
3340 #endif
3341
3342 // Assign discovered "here or above" point
3343 least = find_non_split_ctrl(least);
3344 set_ctrl(n, least);
3345
3346 // Collect inner loop bodies
3347 IdealLoopTree *chosen_loop = get_loop(least);
3348 if( !chosen_loop->_child ) // Inner loop?
3349 chosen_loop->_body.push(n);// Collect inner loops
3350 }
3351
3352 #ifdef ASSERT
3353 void PhaseIdealLoop::dump_bad_graph(Node* n, Node* early, Node* LCA) {
3354 tty->print_cr( "Bad graph detected in build_loop_late");
3355 tty->print("n: "); n->dump();
3356 tty->print("early(n): "); early->dump();
3357 if (n->in(0) != NULL && !n->in(0)->is_top() &&
3358 n->in(0) != early && !n->in(0)->is_Root()) {
3359 tty->print("n->in(0): "); n->in(0)->dump();
3360 }
3361 for (uint i = 1; i < n->req(); i++) {
3362 Node* in1 = n->in(i);
3363 if (in1 != NULL && in1 != n && !in1->is_top()) {
3364 tty->print("n->in(%d): ", i); in1->dump();
3365 Node* in1_early = get_ctrl(in1);
3366 tty->print("early(n->in(%d)): ", i); in1_early->dump();
3367 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
3368 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3369 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3370 }
3371 for (uint j = 1; j < in1->req(); j++) {
3372 Node* in2 = in1->in(j);
3373 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3374 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3375 Node* in2_early = get_ctrl(in2);
3376 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3377 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
3378 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3379 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3380 }
3381 }
3382 }
3383 }
3384 }
3385 tty->cr();
3386 tty->print("LCA(n): "); LCA->dump();
3387 for (uint i = 0; i < n->outcnt(); i++) {
3388 Node* u1 = n->raw_out(i);
3389 if (u1 == n)
3390 continue;
3391 tty->print("n->out(%d): ", i); u1->dump();
3392 if (u1->is_CFG()) {
3393 for (uint j = 0; j < u1->outcnt(); j++) {
3394 Node* u2 = u1->raw_out(j);
3395 if (u2 != u1 && u2 != n && u2->is_CFG()) {
3396 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3397 }
3398 }
3399 } else {
3400 Node* u1_later = get_ctrl(u1);
3401 tty->print("later(n->out(%d)): ", i); u1_later->dump();
3402 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
3403 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3404 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3405 }
3406 for (uint j = 0; j < u1->outcnt(); j++) {
3407 Node* u2 = u1->raw_out(j);
3408 if (u2 == n || u2 == u1)
3409 continue;
3410 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3411 if (!u2->is_CFG()) {
3412 Node* u2_later = get_ctrl(u2);
3413 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3414 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
3415 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3416 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3417 }
3418 }
3419 }
3420 }
3421 }
3422 tty->cr();
3423 int ct = 0;
3424 Node *dbg_legal = LCA;
3425 while(!dbg_legal->is_Start() && ct < 100) {
3426 tty->print("idom[%d] ",ct); dbg_legal->dump();
3427 ct++;
3428 dbg_legal = idom(dbg_legal);
3429 }
3430 tty->cr();
3431 }
3432 #endif
3433
3434 #ifndef PRODUCT
3435 //------------------------------dump-------------------------------------------
3436 void PhaseIdealLoop::dump( ) const {
3437 ResourceMark rm;
3438 Arena* arena = Thread::current()->resource_area();
3439 Node_Stack stack(arena, C->unique() >> 2);
3440 Node_List rpo_list;
3441 VectorSet visited(arena);
3442 visited.set(C->top()->_idx);
3443 rpo( C->root(), stack, visited, rpo_list );
3444 // Dump root loop indexed by last element in PO order
3445 dump( _ltree_root, rpo_list.size(), rpo_list );
3446 }
3447
3448 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3449 loop->dump_head();
3450
3451 // Now scan for CFG nodes in the same loop
3452 for( uint j=idx; j > 0; j-- ) {
3453 Node *n = rpo_list[j-1];
3454 if( !_nodes[n->_idx] ) // Skip dead nodes
3455 continue;
3456 if( get_loop(n) != loop ) { // Wrong loop nest
3457 if( get_loop(n)->_head == n && // Found nested loop?
3458 get_loop(n)->_parent == loop )
3459 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3460 continue;
3461 }
3462
3463 // Dump controlling node
3464 for( uint x = 0; x < loop->_nest; x++ )
3465 tty->print(" ");
3466 tty->print("C");
3467 if( n == C->root() ) {
3468 n->dump();
3469 } else {
3470 Node* cached_idom = idom_no_update(n);
3471 Node *computed_idom = n->in(0);
3472 if( n->is_Region() ) {
3473 computed_idom = compute_idom(n);
3474 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3475 // any MultiBranch ctrl node), so apply a similar transform to
3476 // the cached idom returned from idom_no_update.
3477 cached_idom = find_non_split_ctrl(cached_idom);
3478 }
3479 tty->print(" ID:%d",computed_idom->_idx);
3480 n->dump();
3481 if( cached_idom != computed_idom ) {
3482 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3483 computed_idom->_idx, cached_idom->_idx);
3484 }
3485 }
3486 // Dump nodes it controls
3487 for( uint k = 0; k < _nodes.Size(); k++ ) {
3488 // (k < C->unique() && get_ctrl(find(k)) == n)
3489 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3490 Node *m = C->root()->find(k);
3491 if( m && m->outcnt() > 0 ) {
3492 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3493 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3494 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3495 }
3496 for( uint j = 0; j < loop->_nest; j++ )
3497 tty->print(" ");
3498 tty->print(" ");
3499 m->dump();
3500 }
3501 }
3502 }
3503 }
3504 }
3505
3506 // Collect a R-P-O for the whole CFG.
3507 // Result list is in post-order (scan backwards for RPO)
3508 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3509 stk.push(start, 0);
3510 visited.set(start->_idx);
3511
3512 while (stk.is_nonempty()) {
3513 Node* m = stk.node();
3514 uint idx = stk.index();
3515 if (idx < m->outcnt()) {
3516 stk.set_index(idx + 1);
3517 Node* n = m->raw_out(idx);
3518 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3519 stk.push(n, 0);
3520 }
3521 } else {
3522 rpo_list.push(m);
3523 stk.pop();
3524 }
3525 }
3526 }
3527 #endif
3528
3529
3530 //=============================================================================
3531 //------------------------------LoopTreeIterator-----------------------------------
3532
3533 // Advance to next loop tree using a preorder, left-to-right traversal.
3534 void LoopTreeIterator::next() {
3535 assert(!done(), "must not be done.");
3536 if (_curnt->_child != NULL) {
3537 _curnt = _curnt->_child;
3538 } else if (_curnt->_next != NULL) {
3539 _curnt = _curnt->_next;
3540 } else {
3541 while (_curnt != _root && _curnt->_next == NULL) {
3542 _curnt = _curnt->_parent;
3543 }
3544 if (_curnt == _root) {
3545 _curnt = NULL;
3546 assert(done(), "must be done.");
3547 } else {
3548 assert(_curnt->_next != NULL, "must be more to do");
3549 _curnt = _curnt->_next;
3550 }
3551 }
3552 }