Fri, 16 Oct 2009 02:05:46 -0700
6888898: CMS: ReduceInitialCardMarks unsafe in the presence of cms precleaning
6889757: G1: enable card mark elision for initializing writes from compiled code (ReduceInitialCardMarks)
Summary: Defer the (compiler-elided) card-mark upon a slow-path allocation until after the store and before the next subsequent safepoint; G1 now answers yes to can_elide_tlab_write_barriers().
Reviewed-by: jcoomes, kvn, never
1 /*
2 * Copyright 1998-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_loopnode.cpp.incl"
28 //=============================================================================
29 //------------------------------is_loop_iv-------------------------------------
30 // Determine if a node is Counted loop induction variable.
31 // The method is declared in node.hpp.
32 const Node* Node::is_loop_iv() const {
33 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
34 this->as_Phi()->region()->is_CountedLoop() &&
35 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
36 return this;
37 } else {
38 return NULL;
39 }
40 }
42 //=============================================================================
43 //------------------------------dump_spec--------------------------------------
44 // Dump special per-node info
45 #ifndef PRODUCT
46 void LoopNode::dump_spec(outputStream *st) const {
47 if( is_inner_loop () ) st->print( "inner " );
48 if( is_partial_peel_loop () ) st->print( "partial_peel " );
49 if( partial_peel_has_failed () ) st->print( "partial_peel_failed " );
50 }
51 #endif
53 //------------------------------get_early_ctrl---------------------------------
54 // Compute earliest legal control
55 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
56 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
57 uint i;
58 Node *early;
59 if( n->in(0) ) {
60 early = n->in(0);
61 if( !early->is_CFG() ) // Might be a non-CFG multi-def
62 early = get_ctrl(early); // So treat input as a straight data input
63 i = 1;
64 } else {
65 early = get_ctrl(n->in(1));
66 i = 2;
67 }
68 uint e_d = dom_depth(early);
69 assert( early, "" );
70 for( ; i < n->req(); i++ ) {
71 Node *cin = get_ctrl(n->in(i));
72 assert( cin, "" );
73 // Keep deepest dominator depth
74 uint c_d = dom_depth(cin);
75 if( c_d > e_d ) { // Deeper guy?
76 early = cin; // Keep deepest found so far
77 e_d = c_d;
78 } else if( c_d == e_d && // Same depth?
79 early != cin ) { // If not equal, must use slower algorithm
80 // If same depth but not equal, one _must_ dominate the other
81 // and we want the deeper (i.e., dominated) guy.
82 Node *n1 = early;
83 Node *n2 = cin;
84 while( 1 ) {
85 n1 = idom(n1); // Walk up until break cycle
86 n2 = idom(n2);
87 if( n1 == cin || // Walked early up to cin
88 dom_depth(n2) < c_d )
89 break; // early is deeper; keep him
90 if( n2 == early || // Walked cin up to early
91 dom_depth(n1) < c_d ) {
92 early = cin; // cin is deeper; keep him
93 break;
94 }
95 }
96 e_d = dom_depth(early); // Reset depth register cache
97 }
98 }
100 // Return earliest legal location
101 assert(early == find_non_split_ctrl(early), "unexpected early control");
103 return early;
104 }
106 //------------------------------set_early_ctrl---------------------------------
107 // Set earliest legal control
108 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
109 Node *early = get_early_ctrl(n);
111 // Record earliest legal location
112 set_ctrl(n, early);
113 }
115 //------------------------------set_subtree_ctrl-------------------------------
116 // set missing _ctrl entries on new nodes
117 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
118 // Already set? Get out.
119 if( _nodes[n->_idx] ) return;
120 // Recursively set _nodes array to indicate where the Node goes
121 uint i;
122 for( i = 0; i < n->req(); ++i ) {
123 Node *m = n->in(i);
124 if( m && m != C->root() )
125 set_subtree_ctrl( m );
126 }
128 // Fixup self
129 set_early_ctrl( n );
130 }
132 //------------------------------is_counted_loop--------------------------------
133 Node *PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
134 PhaseGVN *gvn = &_igvn;
136 // Counted loop head must be a good RegionNode with only 3 not NULL
137 // control input edges: Self, Entry, LoopBack.
138 if ( x->in(LoopNode::Self) == NULL || x->req() != 3 )
139 return NULL;
141 Node *init_control = x->in(LoopNode::EntryControl);
142 Node *back_control = x->in(LoopNode::LoopBackControl);
143 if( init_control == NULL || back_control == NULL ) // Partially dead
144 return NULL;
145 // Must also check for TOP when looking for a dead loop
146 if( init_control->is_top() || back_control->is_top() )
147 return NULL;
149 // Allow funny placement of Safepoint
150 if( back_control->Opcode() == Op_SafePoint )
151 back_control = back_control->in(TypeFunc::Control);
153 // Controlling test for loop
154 Node *iftrue = back_control;
155 uint iftrue_op = iftrue->Opcode();
156 if( iftrue_op != Op_IfTrue &&
157 iftrue_op != Op_IfFalse )
158 // I have a weird back-control. Probably the loop-exit test is in
159 // the middle of the loop and I am looking at some trailing control-flow
160 // merge point. To fix this I would have to partially peel the loop.
161 return NULL; // Obscure back-control
163 // Get boolean guarding loop-back test
164 Node *iff = iftrue->in(0);
165 if( get_loop(iff) != loop || !iff->in(1)->is_Bool() ) return NULL;
166 BoolNode *test = iff->in(1)->as_Bool();
167 BoolTest::mask bt = test->_test._test;
168 float cl_prob = iff->as_If()->_prob;
169 if( iftrue_op == Op_IfFalse ) {
170 bt = BoolTest(bt).negate();
171 cl_prob = 1.0 - cl_prob;
172 }
173 // Get backedge compare
174 Node *cmp = test->in(1);
175 int cmp_op = cmp->Opcode();
176 if( cmp_op != Op_CmpI )
177 return NULL; // Avoid pointer & float compares
179 // Find the trip-counter increment & limit. Limit must be loop invariant.
180 Node *incr = cmp->in(1);
181 Node *limit = cmp->in(2);
183 // ---------
184 // need 'loop()' test to tell if limit is loop invariant
185 // ---------
187 if( !is_member( loop, get_ctrl(incr) ) ) { // Swapped trip counter and limit?
188 Node *tmp = incr; // Then reverse order into the CmpI
189 incr = limit;
190 limit = tmp;
191 bt = BoolTest(bt).commute(); // And commute the exit test
192 }
193 if( is_member( loop, get_ctrl(limit) ) ) // Limit must loop-invariant
194 return NULL;
196 // Trip-counter increment must be commutative & associative.
197 uint incr_op = incr->Opcode();
198 if( incr_op == Op_Phi && incr->req() == 3 ) {
199 incr = incr->in(2); // Assume incr is on backedge of Phi
200 incr_op = incr->Opcode();
201 }
202 Node* trunc1 = NULL;
203 Node* trunc2 = NULL;
204 const TypeInt* iv_trunc_t = NULL;
205 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
206 return NULL; // Funny increment opcode
207 }
209 // Get merge point
210 Node *xphi = incr->in(1);
211 Node *stride = incr->in(2);
212 if( !stride->is_Con() ) { // Oops, swap these
213 if( !xphi->is_Con() ) // Is the other guy a constant?
214 return NULL; // Nope, unknown stride, bail out
215 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
216 xphi = stride;
217 stride = tmp;
218 }
219 //if( loop(xphi) != l) return NULL;// Merge point is in inner loop??
220 if( !xphi->is_Phi() ) return NULL; // Too much math on the trip counter
221 PhiNode *phi = xphi->as_Phi();
223 // Stride must be constant
224 const Type *stride_t = stride->bottom_type();
225 int stride_con = stride_t->is_int()->get_con();
226 assert( stride_con, "missed some peephole opt" );
228 // Phi must be of loop header; backedge must wrap to increment
229 if( phi->region() != x ) return NULL;
230 if( trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
231 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1 ) {
232 return NULL;
233 }
234 Node *init_trip = phi->in(LoopNode::EntryControl);
235 //if (!init_trip->is_Con()) return NULL; // avoid rolling over MAXINT/MININT
237 // If iv trunc type is smaller than int, check for possible wrap.
238 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
239 assert(trunc1 != NULL, "must have found some truncation");
241 // Get a better type for the phi (filtered thru if's)
242 const TypeInt* phi_ft = filtered_type(phi);
244 // Can iv take on a value that will wrap?
245 //
246 // Ensure iv's limit is not within "stride" of the wrap value.
247 //
248 // Example for "short" type
249 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
250 // If the stride is +10, then the last value of the induction
251 // variable before the increment (phi_ft->_hi) must be
252 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
253 // ensure no truncation occurs after the increment.
255 if (stride_con > 0) {
256 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
257 iv_trunc_t->_lo > phi_ft->_lo) {
258 return NULL; // truncation may occur
259 }
260 } else if (stride_con < 0) {
261 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
262 iv_trunc_t->_hi < phi_ft->_hi) {
263 return NULL; // truncation may occur
264 }
265 }
266 // No possibility of wrap so truncation can be discarded
267 // Promote iv type to Int
268 } else {
269 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
270 }
272 // =================================================
273 // ---- SUCCESS! Found A Trip-Counted Loop! -----
274 //
275 // Canonicalize the condition on the test. If we can exactly determine
276 // the trip-counter exit value, then set limit to that value and use
277 // a '!=' test. Otherwise use condition '<' for count-up loops and
278 // '>' for count-down loops. If the condition is inverted and we will
279 // be rolling through MININT to MAXINT, then bail out.
281 C->print_method("Before CountedLoop", 3);
283 // Check for SafePoint on backedge and remove
284 Node *sfpt = x->in(LoopNode::LoopBackControl);
285 if( sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
286 lazy_replace( sfpt, iftrue );
287 loop->_tail = iftrue;
288 }
291 // If compare points to incr, we are ok. Otherwise the compare
292 // can directly point to the phi; in this case adjust the compare so that
293 // it points to the incr by adjusting the limit.
294 if( cmp->in(1) == phi || cmp->in(2) == phi )
295 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
297 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
298 // Final value for iterator should be: trip_count * stride + init_trip.
299 const Type *limit_t = limit->bottom_type();
300 const Type *init_t = init_trip->bottom_type();
301 Node *one_p = gvn->intcon( 1);
302 Node *one_m = gvn->intcon(-1);
304 Node *trip_count = NULL;
305 Node *hook = new (C, 6) Node(6);
306 switch( bt ) {
307 case BoolTest::eq:
308 return NULL; // Bail out, but this loop trips at most twice!
309 case BoolTest::ne: // Ahh, the case we desire
310 if( stride_con == 1 )
311 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
312 else if( stride_con == -1 )
313 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
314 else
315 return NULL; // Odd stride; must prove we hit limit exactly
316 set_subtree_ctrl( trip_count );
317 //_loop.map(trip_count->_idx,loop(limit));
318 break;
319 case BoolTest::le: // Maybe convert to '<' case
320 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
321 set_subtree_ctrl( limit );
322 hook->init_req(4, limit);
324 bt = BoolTest::lt;
325 // Make the new limit be in the same loop nest as the old limit
326 //_loop.map(limit->_idx,limit_loop);
327 // Fall into next case
328 case BoolTest::lt: { // Maybe convert to '!=' case
329 if( stride_con < 0 ) return NULL; // Count down loop rolls through MAXINT
330 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
331 set_subtree_ctrl( range );
332 hook->init_req(0, range);
334 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
335 set_subtree_ctrl( bias );
336 hook->init_req(1, bias);
338 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
339 set_subtree_ctrl( bias1 );
340 hook->init_req(2, bias1);
342 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
343 set_subtree_ctrl( trip_count );
344 hook->init_req(3, trip_count);
345 break;
346 }
348 case BoolTest::ge: // Maybe convert to '>' case
349 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
350 set_subtree_ctrl( limit );
351 hook->init_req(4 ,limit);
353 bt = BoolTest::gt;
354 // Make the new limit be in the same loop nest as the old limit
355 //_loop.map(limit->_idx,limit_loop);
356 // Fall into next case
357 case BoolTest::gt: { // Maybe convert to '!=' case
358 if( stride_con > 0 ) return NULL; // count up loop rolls through MININT
359 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
360 set_subtree_ctrl( range );
361 hook->init_req(0, range);
363 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
364 set_subtree_ctrl( bias );
365 hook->init_req(1, bias);
367 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
368 set_subtree_ctrl( bias1 );
369 hook->init_req(2, bias1);
371 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
372 set_subtree_ctrl( trip_count );
373 hook->init_req(3, trip_count);
374 break;
375 }
376 }
378 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
379 set_subtree_ctrl( span );
380 hook->init_req(5, span);
382 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
383 set_subtree_ctrl( limit );
385 // Build a canonical trip test.
386 // Clone code, as old values may be in use.
387 incr = incr->clone();
388 incr->set_req(1,phi);
389 incr->set_req(2,stride);
390 incr = _igvn.register_new_node_with_optimizer(incr);
391 set_early_ctrl( incr );
392 _igvn.hash_delete(phi);
393 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
395 // If phi type is more restrictive than Int, raise to
396 // Int to prevent (almost) infinite recursion in igvn
397 // which can only handle integer types for constants or minint..maxint.
398 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
399 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
400 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
401 nphi = _igvn.register_new_node_with_optimizer(nphi);
402 set_ctrl(nphi, get_ctrl(phi));
403 _igvn.subsume_node(phi, nphi);
404 phi = nphi->as_Phi();
405 }
406 cmp = cmp->clone();
407 cmp->set_req(1,incr);
408 cmp->set_req(2,limit);
409 cmp = _igvn.register_new_node_with_optimizer(cmp);
410 set_ctrl(cmp, iff->in(0));
412 Node *tmp = test->clone();
413 assert( tmp->is_Bool(), "" );
414 test = (BoolNode*)tmp;
415 (*(BoolTest*)&test->_test)._test = bt; //BoolTest::ne;
416 test->set_req(1,cmp);
417 _igvn.register_new_node_with_optimizer(test);
418 set_ctrl(test, iff->in(0));
419 // If the exit test is dead, STOP!
420 if( test == NULL ) return NULL;
421 _igvn.hash_delete(iff);
422 iff->set_req_X( 1, test, &_igvn );
424 // Replace the old IfNode with a new LoopEndNode
425 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), iff->in(1), cl_prob, iff->as_If()->_fcnt ));
426 IfNode *le = lex->as_If();
427 uint dd = dom_depth(iff);
428 set_idom(le, le->in(0), dd); // Update dominance for loop exit
429 set_loop(le, loop);
431 // Get the loop-exit control
432 Node *if_f = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
434 // Need to swap loop-exit and loop-back control?
435 if( iftrue_op == Op_IfFalse ) {
436 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
437 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
439 loop->_tail = back_control = ift2;
440 set_loop(ift2, loop);
441 set_loop(iff2, get_loop(if_f));
443 // Lazy update of 'get_ctrl' mechanism.
444 lazy_replace_proj( if_f , iff2 );
445 lazy_replace_proj( iftrue, ift2 );
447 // Swap names
448 if_f = iff2;
449 iftrue = ift2;
450 } else {
451 _igvn.hash_delete(if_f );
452 _igvn.hash_delete(iftrue);
453 if_f ->set_req_X( 0, le, &_igvn );
454 iftrue->set_req_X( 0, le, &_igvn );
455 }
457 set_idom(iftrue, le, dd+1);
458 set_idom(if_f, le, dd+1);
460 // Now setup a new CountedLoopNode to replace the existing LoopNode
461 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
462 // The following assert is approximately true, and defines the intention
463 // of can_be_counted_loop. It fails, however, because phase->type
464 // is not yet initialized for this loop and its parts.
465 //assert(l->can_be_counted_loop(this), "sanity");
466 _igvn.register_new_node_with_optimizer(l);
467 set_loop(l, loop);
468 loop->_head = l;
469 // Fix all data nodes placed at the old loop head.
470 // Uses the lazy-update mechanism of 'get_ctrl'.
471 lazy_replace( x, l );
472 set_idom(l, init_control, dom_depth(x));
474 // Check for immediately preceding SafePoint and remove
475 Node *sfpt2 = le->in(0);
476 if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
477 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
479 // Free up intermediate goo
480 _igvn.remove_dead_node(hook);
482 C->print_method("After CountedLoop", 3);
484 // Return trip counter
485 return trip_count;
486 }
489 //------------------------------Ideal------------------------------------------
490 // Return a node which is more "ideal" than the current node.
491 // Attempt to convert into a counted-loop.
492 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
493 if (!can_be_counted_loop(phase)) {
494 phase->C->set_major_progress();
495 }
496 return RegionNode::Ideal(phase, can_reshape);
497 }
500 //=============================================================================
501 //------------------------------Ideal------------------------------------------
502 // Return a node which is more "ideal" than the current node.
503 // Attempt to convert into a counted-loop.
504 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
505 return RegionNode::Ideal(phase, can_reshape);
506 }
508 //------------------------------dump_spec--------------------------------------
509 // Dump special per-node info
510 #ifndef PRODUCT
511 void CountedLoopNode::dump_spec(outputStream *st) const {
512 LoopNode::dump_spec(st);
513 if( stride_is_con() ) {
514 st->print("stride: %d ",stride_con());
515 } else {
516 st->print("stride: not constant ");
517 }
518 if( is_pre_loop () ) st->print("pre of N%d" , _main_idx );
519 if( is_main_loop() ) st->print("main of N%d", _idx );
520 if( is_post_loop() ) st->print("post of N%d", _main_idx );
521 }
522 #endif
524 //=============================================================================
525 int CountedLoopEndNode::stride_con() const {
526 return stride()->bottom_type()->is_int()->get_con();
527 }
530 //----------------------match_incr_with_optional_truncation--------------------
531 // Match increment with optional truncation:
532 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
533 // Return NULL for failure. Success returns the increment node.
534 Node* CountedLoopNode::match_incr_with_optional_truncation(
535 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
536 // Quick cutouts:
537 if (expr == NULL || expr->req() != 3) return false;
539 Node *t1 = NULL;
540 Node *t2 = NULL;
541 const TypeInt* trunc_t = TypeInt::INT;
542 Node* n1 = expr;
543 int n1op = n1->Opcode();
545 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
546 if (n1op == Op_AndI &&
547 n1->in(2)->is_Con() &&
548 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
549 // %%% This check should match any mask of 2**K-1.
550 t1 = n1;
551 n1 = t1->in(1);
552 n1op = n1->Opcode();
553 trunc_t = TypeInt::CHAR;
554 } else if (n1op == Op_RShiftI &&
555 n1->in(1) != NULL &&
556 n1->in(1)->Opcode() == Op_LShiftI &&
557 n1->in(2) == n1->in(1)->in(2) &&
558 n1->in(2)->is_Con()) {
559 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
560 // %%% This check should match any shift in [1..31].
561 if (shift == 16 || shift == 8) {
562 t1 = n1;
563 t2 = t1->in(1);
564 n1 = t2->in(1);
565 n1op = n1->Opcode();
566 if (shift == 16) {
567 trunc_t = TypeInt::SHORT;
568 } else if (shift == 8) {
569 trunc_t = TypeInt::BYTE;
570 }
571 }
572 }
574 // If (maybe after stripping) it is an AddI, we won:
575 if (n1op == Op_AddI) {
576 *trunc1 = t1;
577 *trunc2 = t2;
578 *trunc_type = trunc_t;
579 return n1;
580 }
582 // failed
583 return NULL;
584 }
587 //------------------------------filtered_type--------------------------------
588 // Return a type based on condition control flow
589 // A successful return will be a type that is restricted due
590 // to a series of dominating if-tests, such as:
591 // if (i < 10) {
592 // if (i > 0) {
593 // here: "i" type is [1..10)
594 // }
595 // }
596 // or a control flow merge
597 // if (i < 10) {
598 // do {
599 // phi( , ) -- at top of loop type is [min_int..10)
600 // i = ?
601 // } while ( i < 10)
602 //
603 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
604 assert(n && n->bottom_type()->is_int(), "must be int");
605 const TypeInt* filtered_t = NULL;
606 if (!n->is_Phi()) {
607 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
608 filtered_t = filtered_type_from_dominators(n, n_ctrl);
610 } else {
611 Node* phi = n->as_Phi();
612 Node* region = phi->in(0);
613 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
614 if (region && region != C->top()) {
615 for (uint i = 1; i < phi->req(); i++) {
616 Node* val = phi->in(i);
617 Node* use_c = region->in(i);
618 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
619 if (val_t != NULL) {
620 if (filtered_t == NULL) {
621 filtered_t = val_t;
622 } else {
623 filtered_t = filtered_t->meet(val_t)->is_int();
624 }
625 }
626 }
627 }
628 }
629 const TypeInt* n_t = _igvn.type(n)->is_int();
630 if (filtered_t != NULL) {
631 n_t = n_t->join(filtered_t)->is_int();
632 }
633 return n_t;
634 }
637 //------------------------------filtered_type_from_dominators--------------------------------
638 // Return a possibly more restrictive type for val based on condition control flow of dominators
639 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
640 if (val->is_Con()) {
641 return val->bottom_type()->is_int();
642 }
643 uint if_limit = 10; // Max number of dominating if's visited
644 const TypeInt* rtn_t = NULL;
646 if (use_ctrl && use_ctrl != C->top()) {
647 Node* val_ctrl = get_ctrl(val);
648 uint val_dom_depth = dom_depth(val_ctrl);
649 Node* pred = use_ctrl;
650 uint if_cnt = 0;
651 while (if_cnt < if_limit) {
652 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
653 if_cnt++;
654 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
655 if (if_t != NULL) {
656 if (rtn_t == NULL) {
657 rtn_t = if_t;
658 } else {
659 rtn_t = rtn_t->join(if_t)->is_int();
660 }
661 }
662 }
663 pred = idom(pred);
664 if (pred == NULL || pred == C->top()) {
665 break;
666 }
667 // Stop if going beyond definition block of val
668 if (dom_depth(pred) < val_dom_depth) {
669 break;
670 }
671 }
672 }
673 return rtn_t;
674 }
677 //------------------------------dump_spec--------------------------------------
678 // Dump special per-node info
679 #ifndef PRODUCT
680 void CountedLoopEndNode::dump_spec(outputStream *st) const {
681 if( in(TestValue)->is_Bool() ) {
682 BoolTest bt( test_trip()); // Added this for g++.
684 st->print("[");
685 bt.dump_on(st);
686 st->print("]");
687 }
688 st->print(" ");
689 IfNode::dump_spec(st);
690 }
691 #endif
693 //=============================================================================
694 //------------------------------is_member--------------------------------------
695 // Is 'l' a member of 'this'?
696 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
697 while( l->_nest > _nest ) l = l->_parent;
698 return l == this;
699 }
701 //------------------------------set_nest---------------------------------------
702 // Set loop tree nesting depth. Accumulate _has_call bits.
703 int IdealLoopTree::set_nest( uint depth ) {
704 _nest = depth;
705 int bits = _has_call;
706 if( _child ) bits |= _child->set_nest(depth+1);
707 if( bits ) _has_call = 1;
708 if( _next ) bits |= _next ->set_nest(depth );
709 return bits;
710 }
712 //------------------------------split_fall_in----------------------------------
713 // Split out multiple fall-in edges from the loop header. Move them to a
714 // private RegionNode before the loop. This becomes the loop landing pad.
715 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
716 PhaseIterGVN &igvn = phase->_igvn;
717 uint i;
719 // Make a new RegionNode to be the landing pad.
720 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
721 phase->set_loop(landing_pad,_parent);
722 // Gather all the fall-in control paths into the landing pad
723 uint icnt = fall_in_cnt;
724 uint oreq = _head->req();
725 for( i = oreq-1; i>0; i-- )
726 if( !phase->is_member( this, _head->in(i) ) )
727 landing_pad->set_req(icnt--,_head->in(i));
729 // Peel off PhiNode edges as well
730 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
731 Node *oj = _head->fast_out(j);
732 if( oj->is_Phi() ) {
733 PhiNode* old_phi = oj->as_Phi();
734 assert( old_phi->region() == _head, "" );
735 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
736 Node *p = PhiNode::make_blank(landing_pad, old_phi);
737 uint icnt = fall_in_cnt;
738 for( i = oreq-1; i>0; i-- ) {
739 if( !phase->is_member( this, _head->in(i) ) ) {
740 p->init_req(icnt--, old_phi->in(i));
741 // Go ahead and clean out old edges from old phi
742 old_phi->del_req(i);
743 }
744 }
745 // Search for CSE's here, because ZKM.jar does a lot of
746 // loop hackery and we need to be a little incremental
747 // with the CSE to avoid O(N^2) node blow-up.
748 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
749 if( p2 ) { // Found CSE
750 p->destruct(); // Recover useless new node
751 p = p2; // Use old node
752 } else {
753 igvn.register_new_node_with_optimizer(p, old_phi);
754 }
755 // Make old Phi refer to new Phi.
756 old_phi->add_req(p);
757 // Check for the special case of making the old phi useless and
758 // disappear it. In JavaGrande I have a case where this useless
759 // Phi is the loop limit and prevents recognizing a CountedLoop
760 // which in turn prevents removing an empty loop.
761 Node *id_old_phi = old_phi->Identity( &igvn );
762 if( id_old_phi != old_phi ) { // Found a simple identity?
763 // Note that I cannot call 'subsume_node' here, because
764 // that will yank the edge from old_phi to the Region and
765 // I'm mid-iteration over the Region's uses.
766 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
767 Node* use = old_phi->last_out(i);
768 igvn.hash_delete(use);
769 igvn._worklist.push(use);
770 uint uses_found = 0;
771 for (uint j = 0; j < use->len(); j++) {
772 if (use->in(j) == old_phi) {
773 if (j < use->req()) use->set_req (j, id_old_phi);
774 else use->set_prec(j, id_old_phi);
775 uses_found++;
776 }
777 }
778 i -= uses_found; // we deleted 1 or more copies of this edge
779 }
780 }
781 igvn._worklist.push(old_phi);
782 }
783 }
784 // Finally clean out the fall-in edges from the RegionNode
785 for( i = oreq-1; i>0; i-- ) {
786 if( !phase->is_member( this, _head->in(i) ) ) {
787 _head->del_req(i);
788 }
789 }
790 // Transform landing pad
791 igvn.register_new_node_with_optimizer(landing_pad, _head);
792 // Insert landing pad into the header
793 _head->add_req(landing_pad);
794 }
796 //------------------------------split_outer_loop-------------------------------
797 // Split out the outermost loop from this shared header.
798 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
799 PhaseIterGVN &igvn = phase->_igvn;
801 // Find index of outermost loop; it should also be my tail.
802 uint outer_idx = 1;
803 while( _head->in(outer_idx) != _tail ) outer_idx++;
805 // Make a LoopNode for the outermost loop.
806 Node *ctl = _head->in(LoopNode::EntryControl);
807 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
808 outer = igvn.register_new_node_with_optimizer(outer, _head);
809 phase->set_created_loop_node();
810 // Outermost loop falls into '_head' loop
811 _head->set_req(LoopNode::EntryControl, outer);
812 _head->del_req(outer_idx);
813 // Split all the Phis up between '_head' loop and 'outer' loop.
814 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
815 Node *out = _head->fast_out(j);
816 if( out->is_Phi() ) {
817 PhiNode *old_phi = out->as_Phi();
818 assert( old_phi->region() == _head, "" );
819 Node *phi = PhiNode::make_blank(outer, old_phi);
820 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
821 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
822 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
823 // Make old Phi point to new Phi on the fall-in path
824 igvn.hash_delete(old_phi);
825 old_phi->set_req(LoopNode::EntryControl, phi);
826 old_phi->del_req(outer_idx);
827 igvn._worklist.push(old_phi);
828 }
829 }
831 // Use the new loop head instead of the old shared one
832 _head = outer;
833 phase->set_loop(_head, this);
834 }
836 //------------------------------fix_parent-------------------------------------
837 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
838 loop->_parent = parent;
839 if( loop->_child ) fix_parent( loop->_child, loop );
840 if( loop->_next ) fix_parent( loop->_next , parent );
841 }
843 //------------------------------estimate_path_freq-----------------------------
844 static float estimate_path_freq( Node *n ) {
845 // Try to extract some path frequency info
846 IfNode *iff;
847 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
848 uint nop = n->Opcode();
849 if( nop == Op_SafePoint ) { // Skip any safepoint
850 n = n->in(0);
851 continue;
852 }
853 if( nop == Op_CatchProj ) { // Get count from a prior call
854 // Assume call does not always throw exceptions: means the call-site
855 // count is also the frequency of the fall-through path.
856 assert( n->is_CatchProj(), "" );
857 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
858 return 0.0f; // Assume call exception path is rare
859 Node *call = n->in(0)->in(0)->in(0);
860 assert( call->is_Call(), "expect a call here" );
861 const JVMState *jvms = ((CallNode*)call)->jvms();
862 ciMethodData* methodData = jvms->method()->method_data();
863 if (!methodData->is_mature()) return 0.0f; // No call-site data
864 ciProfileData* data = methodData->bci_to_data(jvms->bci());
865 if ((data == NULL) || !data->is_CounterData()) {
866 // no call profile available, try call's control input
867 n = n->in(0);
868 continue;
869 }
870 return data->as_CounterData()->count()/FreqCountInvocations;
871 }
872 // See if there's a gating IF test
873 Node *n_c = n->in(0);
874 if( !n_c->is_If() ) break; // No estimate available
875 iff = n_c->as_If();
876 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
877 // Compute how much count comes on this path
878 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
879 // Have no count info. Skip dull uncommon-trap like branches.
880 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
881 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
882 break;
883 // Skip through never-taken branch; look for a real loop exit.
884 n = iff->in(0);
885 }
886 return 0.0f; // No estimate available
887 }
889 //------------------------------merge_many_backedges---------------------------
890 // Merge all the backedges from the shared header into a private Region.
891 // Feed that region as the one backedge to this loop.
892 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
893 uint i;
895 // Scan for the top 2 hottest backedges
896 float hotcnt = 0.0f;
897 float warmcnt = 0.0f;
898 uint hot_idx = 0;
899 // Loop starts at 2 because slot 1 is the fall-in path
900 for( i = 2; i < _head->req(); i++ ) {
901 float cnt = estimate_path_freq(_head->in(i));
902 if( cnt > hotcnt ) { // Grab hottest path
903 warmcnt = hotcnt;
904 hotcnt = cnt;
905 hot_idx = i;
906 } else if( cnt > warmcnt ) { // And 2nd hottest path
907 warmcnt = cnt;
908 }
909 }
911 // See if the hottest backedge is worthy of being an inner loop
912 // by being much hotter than the next hottest backedge.
913 if( hotcnt <= 0.0001 ||
914 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
916 // Peel out the backedges into a private merge point; peel
917 // them all except optionally hot_idx.
918 PhaseIterGVN &igvn = phase->_igvn;
920 Node *hot_tail = NULL;
921 // Make a Region for the merge point
922 Node *r = new (phase->C, 1) RegionNode(1);
923 for( i = 2; i < _head->req(); i++ ) {
924 if( i != hot_idx )
925 r->add_req( _head->in(i) );
926 else hot_tail = _head->in(i);
927 }
928 igvn.register_new_node_with_optimizer(r, _head);
929 // Plug region into end of loop _head, followed by hot_tail
930 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
931 _head->set_req(2, r);
932 if( hot_idx ) _head->add_req(hot_tail);
934 // Split all the Phis up between '_head' loop and the Region 'r'
935 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
936 Node *out = _head->fast_out(j);
937 if( out->is_Phi() ) {
938 PhiNode* n = out->as_Phi();
939 igvn.hash_delete(n); // Delete from hash before hacking edges
940 Node *hot_phi = NULL;
941 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
942 // Check all inputs for the ones to peel out
943 uint j = 1;
944 for( uint i = 2; i < n->req(); i++ ) {
945 if( i != hot_idx )
946 phi->set_req( j++, n->in(i) );
947 else hot_phi = n->in(i);
948 }
949 // Register the phi but do not transform until whole place transforms
950 igvn.register_new_node_with_optimizer(phi, n);
951 // Add the merge phi to the old Phi
952 while( n->req() > 3 ) n->del_req( n->req()-1 );
953 n->set_req(2, phi);
954 if( hot_idx ) n->add_req(hot_phi);
955 }
956 }
959 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
960 // of self loop tree. Turn self into a loop headed by _head and with
961 // tail being the new merge point.
962 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
963 phase->set_loop(_tail,ilt); // Adjust tail
964 _tail = r; // Self's tail is new merge point
965 phase->set_loop(r,this);
966 ilt->_child = _child; // New guy has my children
967 _child = ilt; // Self has new guy as only child
968 ilt->_parent = this; // new guy has self for parent
969 ilt->_nest = _nest; // Same nesting depth (for now)
971 // Starting with 'ilt', look for child loop trees using the same shared
972 // header. Flatten these out; they will no longer be loops in the end.
973 IdealLoopTree **pilt = &_child;
974 while( ilt ) {
975 if( ilt->_head == _head ) {
976 uint i;
977 for( i = 2; i < _head->req(); i++ )
978 if( _head->in(i) == ilt->_tail )
979 break; // Still a loop
980 if( i == _head->req() ) { // No longer a loop
981 // Flatten ilt. Hang ilt's "_next" list from the end of
982 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
983 IdealLoopTree **cp = &ilt->_child;
984 while( *cp ) cp = &(*cp)->_next; // Find end of child list
985 *cp = ilt->_next; // Hang next list at end of child list
986 *pilt = ilt->_child; // Move child up to replace ilt
987 ilt->_head = NULL; // Flag as a loop UNIONED into parent
988 ilt = ilt->_child; // Repeat using new ilt
989 continue; // do not advance over ilt->_child
990 }
991 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
992 phase->set_loop(_head,ilt);
993 }
994 pilt = &ilt->_child; // Advance to next
995 ilt = *pilt;
996 }
998 if( _child ) fix_parent( _child, this );
999 }
1001 //------------------------------beautify_loops---------------------------------
1002 // Split shared headers and insert loop landing pads.
1003 // Insert a LoopNode to replace the RegionNode.
1004 // Return TRUE if loop tree is structurally changed.
1005 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1006 bool result = false;
1007 // Cache parts in locals for easy
1008 PhaseIterGVN &igvn = phase->_igvn;
1010 phase->C->print_method("Before beautify loops", 3);
1012 igvn.hash_delete(_head); // Yank from hash before hacking edges
1014 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1015 int fall_in_cnt = 0;
1016 for( uint i = 1; i < _head->req(); i++ )
1017 if( !phase->is_member( this, _head->in(i) ) )
1018 fall_in_cnt++;
1019 assert( fall_in_cnt, "at least 1 fall-in path" );
1020 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1021 split_fall_in( phase, fall_in_cnt );
1023 // Swap inputs to the _head and all Phis to move the fall-in edge to
1024 // the left.
1025 fall_in_cnt = 1;
1026 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1027 fall_in_cnt++;
1028 if( fall_in_cnt > 1 ) {
1029 // Since I am just swapping inputs I do not need to update def-use info
1030 Node *tmp = _head->in(1);
1031 _head->set_req( 1, _head->in(fall_in_cnt) );
1032 _head->set_req( fall_in_cnt, tmp );
1033 // Swap also all Phis
1034 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1035 Node* phi = _head->fast_out(i);
1036 if( phi->is_Phi() ) {
1037 igvn.hash_delete(phi); // Yank from hash before hacking edges
1038 tmp = phi->in(1);
1039 phi->set_req( 1, phi->in(fall_in_cnt) );
1040 phi->set_req( fall_in_cnt, tmp );
1041 }
1042 }
1043 }
1044 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1045 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1047 // If I am a shared header (multiple backedges), peel off the many
1048 // backedges into a private merge point and use the merge point as
1049 // the one true backedge.
1050 if( _head->req() > 3 ) {
1051 // Merge the many backedges into a single backedge.
1052 merge_many_backedges( phase );
1053 result = true;
1054 }
1056 // If I am a shared header (multiple backedges), peel off myself loop.
1057 // I better be the outermost loop.
1058 if( _head->req() > 3 ) {
1059 split_outer_loop( phase );
1060 result = true;
1062 } else if( !_head->is_Loop() && !_irreducible ) {
1063 // Make a new LoopNode to replace the old loop head
1064 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1065 l = igvn.register_new_node_with_optimizer(l, _head);
1066 phase->set_created_loop_node();
1067 // Go ahead and replace _head
1068 phase->_igvn.subsume_node( _head, l );
1069 _head = l;
1070 phase->set_loop(_head, this);
1071 for (DUIterator_Fast imax, i = l->fast_outs(imax); i < imax; i++)
1072 phase->_igvn.add_users_to_worklist(l->fast_out(i));
1073 }
1075 // Now recursively beautify nested loops
1076 if( _child ) result |= _child->beautify_loops( phase );
1077 if( _next ) result |= _next ->beautify_loops( phase );
1078 return result;
1079 }
1081 //------------------------------allpaths_check_safepts----------------------------
1082 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1083 // encountered. Helper for check_safepts.
1084 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1085 assert(stack.size() == 0, "empty stack");
1086 stack.push(_tail);
1087 visited.Clear();
1088 visited.set(_tail->_idx);
1089 while (stack.size() > 0) {
1090 Node* n = stack.pop();
1091 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1092 // Terminate this path
1093 } else if (n->Opcode() == Op_SafePoint) {
1094 if (_phase->get_loop(n) != this) {
1095 if (_required_safept == NULL) _required_safept = new Node_List();
1096 _required_safept->push(n); // save the one closest to the tail
1097 }
1098 // Terminate this path
1099 } else {
1100 uint start = n->is_Region() ? 1 : 0;
1101 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1102 for (uint i = start; i < end; i++) {
1103 Node* in = n->in(i);
1104 assert(in->is_CFG(), "must be");
1105 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1106 stack.push(in);
1107 }
1108 }
1109 }
1110 }
1111 }
1113 //------------------------------check_safepts----------------------------
1114 // Given dominators, try to find loops with calls that must always be
1115 // executed (call dominates loop tail). These loops do not need non-call
1116 // safepoints (ncsfpt).
1117 //
1118 // A complication is that a safepoint in a inner loop may be needed
1119 // by an outer loop. In the following, the inner loop sees it has a
1120 // call (block 3) on every path from the head (block 2) to the
1121 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1122 // in block 2, _but_ this leaves the outer loop without a safepoint.
1123 //
1124 // entry 0
1125 // |
1126 // v
1127 // outer 1,2 +->1
1128 // | |
1129 // | v
1130 // | 2<---+ ncsfpt in 2
1131 // |_/|\ |
1132 // | v |
1133 // inner 2,3 / 3 | call in 3
1134 // / | |
1135 // v +--+
1136 // exit 4
1137 //
1138 //
1139 // This method creates a list (_required_safept) of ncsfpt nodes that must
1140 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1141 // is first looked for in the lists for the outer loops of the current loop.
1142 //
1143 // The insights into the problem:
1144 // A) counted loops are okay
1145 // B) innermost loops are okay (only an inner loop can delete
1146 // a ncsfpt needed by an outer loop)
1147 // C) a loop is immune from an inner loop deleting a safepoint
1148 // if the loop has a call on the idom-path
1149 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1150 // idom-path that is not in a nested loop
1151 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1152 // loop needs to be prevented from deletion by an inner loop
1153 //
1154 // There are two analyses:
1155 // 1) The first, and cheaper one, scans the loop body from
1156 // tail to head following the idom (immediate dominator)
1157 // chain, looking for the cases (C,D,E) above.
1158 // Since inner loops are scanned before outer loops, there is summary
1159 // information about inner loops. Inner loops can be skipped over
1160 // when the tail of an inner loop is encountered.
1161 //
1162 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1163 // the idom path (which is rare), scans all predecessor control paths
1164 // from the tail to the head, terminating a path when a call or sfpt
1165 // is encountered, to find the ncsfpt's that are closest to the tail.
1166 //
1167 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1168 // Bottom up traversal
1169 IdealLoopTree* ch = _child;
1170 while (ch != NULL) {
1171 ch->check_safepts(visited, stack);
1172 ch = ch->_next;
1173 }
1175 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1176 bool has_call = false; // call on dom-path
1177 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1178 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1179 // Scan the dom-path nodes from tail to head
1180 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1181 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1182 has_call = true;
1183 _has_sfpt = 1; // Then no need for a safept!
1184 break;
1185 } else if (n->Opcode() == Op_SafePoint) {
1186 if (_phase->get_loop(n) == this) {
1187 has_local_ncsfpt = true;
1188 break;
1189 }
1190 if (nonlocal_ncsfpt == NULL) {
1191 nonlocal_ncsfpt = n; // save the one closest to the tail
1192 }
1193 } else {
1194 IdealLoopTree* nlpt = _phase->get_loop(n);
1195 if (this != nlpt) {
1196 // If at an inner loop tail, see if the inner loop has already
1197 // recorded seeing a call on the dom-path (and stop.) If not,
1198 // jump to the head of the inner loop.
1199 assert(is_member(nlpt), "nested loop");
1200 Node* tail = nlpt->_tail;
1201 if (tail->in(0)->is_If()) tail = tail->in(0);
1202 if (n == tail) {
1203 // If inner loop has call on dom-path, so does outer loop
1204 if (nlpt->_has_sfpt) {
1205 has_call = true;
1206 _has_sfpt = 1;
1207 break;
1208 }
1209 // Skip to head of inner loop
1210 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1211 n = nlpt->_head;
1212 }
1213 }
1214 }
1215 }
1216 // Record safept's that this loop needs preserved when an
1217 // inner loop attempts to delete it's safepoints.
1218 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1219 if (nonlocal_ncsfpt != NULL) {
1220 if (_required_safept == NULL) _required_safept = new Node_List();
1221 _required_safept->push(nonlocal_ncsfpt);
1222 } else {
1223 // Failed to find a suitable safept on the dom-path. Now use
1224 // an all paths walk from tail to head, looking for safepoints to preserve.
1225 allpaths_check_safepts(visited, stack);
1226 }
1227 }
1228 }
1229 }
1231 //---------------------------is_deleteable_safept----------------------------
1232 // Is safept not required by an outer loop?
1233 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1234 assert(sfpt->Opcode() == Op_SafePoint, "");
1235 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1236 while (lp != NULL) {
1237 Node_List* sfpts = lp->_required_safept;
1238 if (sfpts != NULL) {
1239 for (uint i = 0; i < sfpts->size(); i++) {
1240 if (sfpt == sfpts->at(i))
1241 return false;
1242 }
1243 }
1244 lp = lp->_parent;
1245 }
1246 return true;
1247 }
1249 //------------------------------counted_loop-----------------------------------
1250 // Convert to counted loops where possible
1251 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1253 // For grins, set the inner-loop flag here
1254 if( !_child ) {
1255 if( _head->is_Loop() ) _head->as_Loop()->set_inner_loop();
1256 }
1258 if( _head->is_CountedLoop() ||
1259 phase->is_counted_loop( _head, this ) ) {
1260 _has_sfpt = 1; // Indicate we do not need a safepoint here
1262 // Look for a safepoint to remove
1263 for (Node* n = tail(); n != _head; n = phase->idom(n))
1264 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1265 phase->is_deleteable_safept(n))
1266 phase->lazy_replace(n,n->in(TypeFunc::Control));
1268 CountedLoopNode *cl = _head->as_CountedLoop();
1269 Node *incr = cl->incr();
1270 if( !incr ) return; // Dead loop?
1271 Node *init = cl->init_trip();
1272 Node *phi = cl->phi();
1273 // protect against stride not being a constant
1274 if( !cl->stride_is_con() ) return;
1275 int stride_con = cl->stride_con();
1277 // Look for induction variables
1279 // Visit all children, looking for Phis
1280 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1281 Node *out = cl->out(i);
1282 if (!out->is_Phi() || out == phi) continue; // Looking for other phis
1283 PhiNode* phi2 = out->as_Phi();
1284 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1285 // Look for induction variables of the form: X += constant
1286 if( phi2->region() != _head ||
1287 incr2->req() != 3 ||
1288 incr2->in(1) != phi2 ||
1289 incr2 == incr ||
1290 incr2->Opcode() != Op_AddI ||
1291 !incr2->in(2)->is_Con() )
1292 continue;
1294 // Check for parallel induction variable (parallel to trip counter)
1295 // via an affine function. In particular, count-down loops with
1296 // count-up array indices are common. We only RCE references off
1297 // the trip-counter, so we need to convert all these to trip-counter
1298 // expressions.
1299 Node *init2 = phi2->in( LoopNode::EntryControl );
1300 int stride_con2 = incr2->in(2)->get_int();
1302 // The general case here gets a little tricky. We want to find the
1303 // GCD of all possible parallel IV's and make a new IV using this
1304 // GCD for the loop. Then all possible IVs are simple multiples of
1305 // the GCD. In practice, this will cover very few extra loops.
1306 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1307 // where +/-1 is the common case, but other integer multiples are
1308 // also easy to handle.
1309 int ratio_con = stride_con2/stride_con;
1311 if( ratio_con * stride_con == stride_con2 ) { // Check for exact
1312 // Convert to using the trip counter. The parallel induction
1313 // variable differs from the trip counter by a loop-invariant
1314 // amount, the difference between their respective initial values.
1315 // It is scaled by the 'ratio_con'.
1316 Compile* C = phase->C;
1317 Node* ratio = phase->_igvn.intcon(ratio_con);
1318 phase->set_ctrl(ratio, C->root());
1319 Node* ratio_init = new (C, 3) MulINode(init, ratio);
1320 phase->_igvn.register_new_node_with_optimizer(ratio_init, init);
1321 phase->set_early_ctrl(ratio_init);
1322 Node* diff = new (C, 3) SubINode(init2, ratio_init);
1323 phase->_igvn.register_new_node_with_optimizer(diff, init2);
1324 phase->set_early_ctrl(diff);
1325 Node* ratio_idx = new (C, 3) MulINode(phi, ratio);
1326 phase->_igvn.register_new_node_with_optimizer(ratio_idx, phi);
1327 phase->set_ctrl(ratio_idx, cl);
1328 Node* add = new (C, 3) AddINode(ratio_idx, diff);
1329 phase->_igvn.register_new_node_with_optimizer(add);
1330 phase->set_ctrl(add, cl);
1331 phase->_igvn.hash_delete( phi2 );
1332 phase->_igvn.subsume_node( phi2, add );
1333 // Sometimes an induction variable is unused
1334 if (add->outcnt() == 0) {
1335 phase->_igvn.remove_dead_node(add);
1336 }
1337 --i; // deleted this phi; rescan starting with next position
1338 continue;
1339 }
1340 }
1341 } else if (_parent != NULL && !_irreducible) {
1342 // Not a counted loop.
1343 // Look for a safepoint on the idom-path to remove, preserving the first one
1344 bool found = false;
1345 Node* n = tail();
1346 for (; n != _head && !found; n = phase->idom(n)) {
1347 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1348 found = true; // Found one
1349 }
1350 // Skip past it and delete the others
1351 for (; n != _head; n = phase->idom(n)) {
1352 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1353 phase->is_deleteable_safept(n))
1354 phase->lazy_replace(n,n->in(TypeFunc::Control));
1355 }
1356 }
1358 // Recursively
1359 if( _child ) _child->counted_loop( phase );
1360 if( _next ) _next ->counted_loop( phase );
1361 }
1363 #ifndef PRODUCT
1364 //------------------------------dump_head--------------------------------------
1365 // Dump 1 liner for loop header info
1366 void IdealLoopTree::dump_head( ) const {
1367 for( uint i=0; i<_nest; i++ )
1368 tty->print(" ");
1369 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1370 if( _irreducible ) tty->print(" IRREDUCIBLE");
1371 if( _head->is_CountedLoop() ) {
1372 CountedLoopNode *cl = _head->as_CountedLoop();
1373 tty->print(" counted");
1374 if( cl->is_pre_loop () ) tty->print(" pre" );
1375 if( cl->is_main_loop() ) tty->print(" main");
1376 if( cl->is_post_loop() ) tty->print(" post");
1377 }
1378 tty->cr();
1379 }
1381 //------------------------------dump-------------------------------------------
1382 // Dump loops by loop tree
1383 void IdealLoopTree::dump( ) const {
1384 dump_head();
1385 if( _child ) _child->dump();
1386 if( _next ) _next ->dump();
1387 }
1389 #endif
1391 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1392 if (loop == root) {
1393 if (loop->_child != NULL) {
1394 log->begin_head("loop_tree");
1395 log->end_head();
1396 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1397 log->tail("loop_tree");
1398 assert(loop->_next == NULL, "what?");
1399 }
1400 } else {
1401 Node* head = loop->_head;
1402 log->begin_head("loop");
1403 log->print(" idx='%d' ", head->_idx);
1404 if (loop->_irreducible) log->print("irreducible='1' ");
1405 if (head->is_Loop()) {
1406 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1407 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1408 }
1409 if (head->is_CountedLoop()) {
1410 CountedLoopNode* cl = head->as_CountedLoop();
1411 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1412 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1413 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1414 }
1415 log->end_head();
1416 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1417 log->tail("loop");
1418 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1419 }
1420 }
1422 //=============================================================================
1423 //----------------------------build_and_optimize-------------------------------
1424 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1425 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1426 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs) {
1427 int old_progress = C->major_progress();
1429 // Reset major-progress flag for the driver's heuristics
1430 C->clear_major_progress();
1432 #ifndef PRODUCT
1433 // Capture for later assert
1434 uint unique = C->unique();
1435 _loop_invokes++;
1436 _loop_work += unique;
1437 #endif
1439 // True if the method has at least 1 irreducible loop
1440 _has_irreducible_loops = false;
1442 _created_loop_node = false;
1444 Arena *a = Thread::current()->resource_area();
1445 VectorSet visited(a);
1446 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1447 _nodes.map(C->unique(), NULL);
1448 memset(_nodes.adr(), 0, wordSize * C->unique());
1450 // Pre-build the top-level outermost loop tree entry
1451 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1452 // Do not need a safepoint at the top level
1453 _ltree_root->_has_sfpt = 1;
1455 // Empty pre-order array
1456 allocate_preorders();
1458 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1459 // IdealLoopTree entries. Data nodes are NOT walked.
1460 build_loop_tree();
1461 // Check for bailout, and return
1462 if (C->failing()) {
1463 return;
1464 }
1466 // No loops after all
1467 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
1469 // There should always be an outer loop containing the Root and Return nodes.
1470 // If not, we have a degenerate empty program. Bail out in this case.
1471 if (!has_node(C->root())) {
1472 if (!_verify_only) {
1473 C->clear_major_progress();
1474 C->record_method_not_compilable("empty program detected during loop optimization");
1475 }
1476 return;
1477 }
1479 // Nothing to do, so get out
1480 if( !C->has_loops() && !do_split_ifs && !_verify_me && !_verify_only ) {
1481 _igvn.optimize(); // Cleanup NeverBranches
1482 return;
1483 }
1485 // Set loop nesting depth
1486 _ltree_root->set_nest( 0 );
1488 // Split shared headers and insert loop landing pads.
1489 // Do not bother doing this on the Root loop of course.
1490 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
1491 if( _ltree_root->_child->beautify_loops( this ) ) {
1492 // Re-build loop tree!
1493 _ltree_root->_child = NULL;
1494 _nodes.clear();
1495 reallocate_preorders();
1496 build_loop_tree();
1497 // Check for bailout, and return
1498 if (C->failing()) {
1499 return;
1500 }
1501 // Reset loop nesting depth
1502 _ltree_root->set_nest( 0 );
1504 C->print_method("After beautify loops", 3);
1505 }
1506 }
1508 // Build Dominators for elision of NULL checks & loop finding.
1509 // Since nodes do not have a slot for immediate dominator, make
1510 // a persistent side array for that info indexed on node->_idx.
1511 _idom_size = C->unique();
1512 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1513 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1514 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1515 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1517 Dominators();
1519 if (!_verify_only) {
1520 // As a side effect, Dominators removed any unreachable CFG paths
1521 // into RegionNodes. It doesn't do this test against Root, so
1522 // we do it here.
1523 for( uint i = 1; i < C->root()->req(); i++ ) {
1524 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1525 _igvn.hash_delete(C->root());
1526 C->root()->del_req(i);
1527 _igvn._worklist.push(C->root());
1528 i--; // Rerun same iteration on compressed edges
1529 }
1530 }
1532 // Given dominators, try to find inner loops with calls that must
1533 // always be executed (call dominates loop tail). These loops do
1534 // not need a separate safepoint.
1535 Node_List cisstack(a);
1536 _ltree_root->check_safepts(visited, cisstack);
1537 }
1539 // Walk the DATA nodes and place into loops. Find earliest control
1540 // node. For CFG nodes, the _nodes array starts out and remains
1541 // holding the associated IdealLoopTree pointer. For DATA nodes, the
1542 // _nodes array holds the earliest legal controlling CFG node.
1544 // Allocate stack with enough space to avoid frequent realloc
1545 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
1546 Node_Stack nstack( a, stack_size );
1548 visited.Clear();
1549 Node_List worklist(a);
1550 // Don't need C->root() on worklist since
1551 // it will be processed among C->top() inputs
1552 worklist.push( C->top() );
1553 visited.set( C->top()->_idx ); // Set C->top() as visited now
1554 build_loop_early( visited, worklist, nstack );
1556 // Given early legal placement, try finding counted loops. This placement
1557 // is good enough to discover most loop invariants.
1558 if( !_verify_me && !_verify_only )
1559 _ltree_root->counted_loop( this );
1561 // Find latest loop placement. Find ideal loop placement.
1562 visited.Clear();
1563 init_dom_lca_tags();
1564 // Need C->root() on worklist when processing outs
1565 worklist.push( C->root() );
1566 NOT_PRODUCT( C->verify_graph_edges(); )
1567 worklist.push( C->top() );
1568 build_loop_late( visited, worklist, nstack );
1570 if (_verify_only) {
1571 // restore major progress flag
1572 for (int i = 0; i < old_progress; i++)
1573 C->set_major_progress();
1574 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
1575 assert(_igvn._worklist.size() == 0, "shouldn't push anything");
1576 return;
1577 }
1579 // clear out the dead code
1580 while(_deadlist.size()) {
1581 _igvn.remove_globally_dead_node(_deadlist.pop());
1582 }
1584 #ifndef PRODUCT
1585 C->verify_graph_edges();
1586 if( _verify_me ) { // Nested verify pass?
1587 // Check to see if the verify mode is broken
1588 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
1589 return;
1590 }
1591 if( VerifyLoopOptimizations ) verify();
1592 #endif
1594 if (ReassociateInvariants) {
1595 // Reassociate invariants and prep for split_thru_phi
1596 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1597 IdealLoopTree* lpt = iter.current();
1598 if (!lpt->is_counted() || !lpt->is_inner()) continue;
1600 lpt->reassociate_invariants(this);
1602 // Because RCE opportunities can be masked by split_thru_phi,
1603 // look for RCE candidates and inhibit split_thru_phi
1604 // on just their loop-phi's for this pass of loop opts
1605 if( SplitIfBlocks && do_split_ifs ) {
1606 if (lpt->policy_range_check(this)) {
1607 lpt->_rce_candidate = 1; // = true
1608 }
1609 }
1610 }
1611 }
1613 // Check for aggressive application of split-if and other transforms
1614 // that require basic-block info (like cloning through Phi's)
1615 if( SplitIfBlocks && do_split_ifs ) {
1616 visited.Clear();
1617 split_if_with_blocks( visited, nstack );
1618 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
1619 }
1621 // Perform iteration-splitting on inner loops. Split iterations to avoid
1622 // range checks or one-shot null checks.
1624 // If split-if's didn't hack the graph too bad (no CFG changes)
1625 // then do loop opts.
1626 if( C->has_loops() && !C->major_progress() ) {
1627 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
1628 _ltree_root->_child->iteration_split( this, worklist );
1629 // No verify after peeling! GCM has hoisted code out of the loop.
1630 // After peeling, the hoisted code could sink inside the peeled area.
1631 // The peeling code does not try to recompute the best location for
1632 // all the code before the peeled area, so the verify pass will always
1633 // complain about it.
1634 }
1635 // Do verify graph edges in any case
1636 NOT_PRODUCT( C->verify_graph_edges(); );
1638 if( !do_split_ifs ) {
1639 // We saw major progress in Split-If to get here. We forced a
1640 // pass with unrolling and not split-if, however more split-if's
1641 // might make progress. If the unrolling didn't make progress
1642 // then the major-progress flag got cleared and we won't try
1643 // another round of Split-If. In particular the ever-common
1644 // instance-of/check-cast pattern requires at least 2 rounds of
1645 // Split-If to clear out.
1646 C->set_major_progress();
1647 }
1649 // Repeat loop optimizations if new loops were seen
1650 if (created_loop_node()) {
1651 C->set_major_progress();
1652 }
1654 // Convert scalar to superword operations
1656 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
1657 // SuperWord transform
1658 SuperWord sw(this);
1659 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1660 IdealLoopTree* lpt = iter.current();
1661 if (lpt->is_counted()) {
1662 sw.transform_loop(lpt);
1663 }
1664 }
1665 }
1667 // Cleanup any modified bits
1668 _igvn.optimize();
1670 // disable assert until issue with split_flow_path is resolved (6742111)
1671 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
1672 // "shouldn't introduce irreducible loops");
1674 if (C->log() != NULL) {
1675 log_loop_tree(_ltree_root, _ltree_root, C->log());
1676 }
1677 }
1679 #ifndef PRODUCT
1680 //------------------------------print_statistics-------------------------------
1681 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
1682 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
1683 void PhaseIdealLoop::print_statistics() {
1684 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
1685 }
1687 //------------------------------verify-----------------------------------------
1688 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
1689 static int fail; // debug only, so its multi-thread dont care
1690 void PhaseIdealLoop::verify() const {
1691 int old_progress = C->major_progress();
1692 ResourceMark rm;
1693 PhaseIdealLoop loop_verify( _igvn, this );
1694 VectorSet visited(Thread::current()->resource_area());
1696 fail = 0;
1697 verify_compare( C->root(), &loop_verify, visited );
1698 assert( fail == 0, "verify loops failed" );
1699 // Verify loop structure is the same
1700 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
1701 // Reset major-progress. It was cleared by creating a verify version of
1702 // PhaseIdealLoop.
1703 for( int i=0; i<old_progress; i++ )
1704 C->set_major_progress();
1705 }
1707 //------------------------------verify_compare---------------------------------
1708 // Make sure me and the given PhaseIdealLoop agree on key data structures
1709 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
1710 if( !n ) return;
1711 if( visited.test_set( n->_idx ) ) return;
1712 if( !_nodes[n->_idx] ) { // Unreachable
1713 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
1714 return;
1715 }
1717 uint i;
1718 for( i = 0; i < n->req(); i++ )
1719 verify_compare( n->in(i), loop_verify, visited );
1721 // Check the '_nodes' block/loop structure
1722 i = n->_idx;
1723 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
1724 if( _nodes[i] != loop_verify->_nodes[i] &&
1725 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
1726 tty->print("Mismatched control setting for: ");
1727 n->dump();
1728 if( fail++ > 10 ) return;
1729 Node *c = get_ctrl_no_update(n);
1730 tty->print("We have it as: ");
1731 if( c->in(0) ) c->dump();
1732 else tty->print_cr("N%d",c->_idx);
1733 tty->print("Verify thinks: ");
1734 if( loop_verify->has_ctrl(n) )
1735 loop_verify->get_ctrl_no_update(n)->dump();
1736 else
1737 loop_verify->get_loop_idx(n)->dump();
1738 tty->cr();
1739 }
1740 } else { // We have a loop
1741 IdealLoopTree *us = get_loop_idx(n);
1742 if( loop_verify->has_ctrl(n) ) {
1743 tty->print("Mismatched loop setting for: ");
1744 n->dump();
1745 if( fail++ > 10 ) return;
1746 tty->print("We have it as: ");
1747 us->dump();
1748 tty->print("Verify thinks: ");
1749 loop_verify->get_ctrl_no_update(n)->dump();
1750 tty->cr();
1751 } else if (!C->major_progress()) {
1752 // Loop selection can be messed up if we did a major progress
1753 // operation, like split-if. Do not verify in that case.
1754 IdealLoopTree *them = loop_verify->get_loop_idx(n);
1755 if( us->_head != them->_head || us->_tail != them->_tail ) {
1756 tty->print("Unequals loops for: ");
1757 n->dump();
1758 if( fail++ > 10 ) return;
1759 tty->print("We have it as: ");
1760 us->dump();
1761 tty->print("Verify thinks: ");
1762 them->dump();
1763 tty->cr();
1764 }
1765 }
1766 }
1768 // Check for immediate dominators being equal
1769 if( i >= _idom_size ) {
1770 if( !n->is_CFG() ) return;
1771 tty->print("CFG Node with no idom: ");
1772 n->dump();
1773 return;
1774 }
1775 if( !n->is_CFG() ) return;
1776 if( n == C->root() ) return; // No IDOM here
1778 assert(n->_idx == i, "sanity");
1779 Node *id = idom_no_update(n);
1780 if( id != loop_verify->idom_no_update(n) ) {
1781 tty->print("Unequals idoms for: ");
1782 n->dump();
1783 if( fail++ > 10 ) return;
1784 tty->print("We have it as: ");
1785 id->dump();
1786 tty->print("Verify thinks: ");
1787 loop_verify->idom_no_update(n)->dump();
1788 tty->cr();
1789 }
1791 }
1793 //------------------------------verify_tree------------------------------------
1794 // Verify that tree structures match. Because the CFG can change, siblings
1795 // within the loop tree can be reordered. We attempt to deal with that by
1796 // reordering the verify's loop tree if possible.
1797 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
1798 assert( _parent == parent, "Badly formed loop tree" );
1800 // Siblings not in same order? Attempt to re-order.
1801 if( _head != loop->_head ) {
1802 // Find _next pointer to update
1803 IdealLoopTree **pp = &loop->_parent->_child;
1804 while( *pp != loop )
1805 pp = &((*pp)->_next);
1806 // Find proper sibling to be next
1807 IdealLoopTree **nn = &loop->_next;
1808 while( (*nn) && (*nn)->_head != _head )
1809 nn = &((*nn)->_next);
1811 // Check for no match.
1812 if( !(*nn) ) {
1813 // Annoyingly, irreducible loops can pick different headers
1814 // after a major_progress operation, so the rest of the loop
1815 // tree cannot be matched.
1816 if (_irreducible && Compile::current()->major_progress()) return;
1817 assert( 0, "failed to match loop tree" );
1818 }
1820 // Move (*nn) to (*pp)
1821 IdealLoopTree *hit = *nn;
1822 *nn = hit->_next;
1823 hit->_next = loop;
1824 *pp = loop;
1825 loop = hit;
1826 // Now try again to verify
1827 }
1829 assert( _head == loop->_head , "mismatched loop head" );
1830 Node *tail = _tail; // Inline a non-updating version of
1831 while( !tail->in(0) ) // the 'tail()' call.
1832 tail = tail->in(1);
1833 assert( tail == loop->_tail, "mismatched loop tail" );
1835 // Counted loops that are guarded should be able to find their guards
1836 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
1837 CountedLoopNode *cl = _head->as_CountedLoop();
1838 Node *init = cl->init_trip();
1839 Node *ctrl = cl->in(LoopNode::EntryControl);
1840 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1841 Node *iff = ctrl->in(0);
1842 assert( iff->Opcode() == Op_If, "" );
1843 Node *bol = iff->in(1);
1844 assert( bol->Opcode() == Op_Bool, "" );
1845 Node *cmp = bol->in(1);
1846 assert( cmp->Opcode() == Op_CmpI, "" );
1847 Node *add = cmp->in(1);
1848 Node *opaq;
1849 if( add->Opcode() == Op_Opaque1 ) {
1850 opaq = add;
1851 } else {
1852 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
1853 assert( add == init, "" );
1854 opaq = cmp->in(2);
1855 }
1856 assert( opaq->Opcode() == Op_Opaque1, "" );
1858 }
1860 if (_child != NULL) _child->verify_tree(loop->_child, this);
1861 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
1862 // Innermost loops need to verify loop bodies,
1863 // but only if no 'major_progress'
1864 int fail = 0;
1865 if (!Compile::current()->major_progress() && _child == NULL) {
1866 for( uint i = 0; i < _body.size(); i++ ) {
1867 Node *n = _body.at(i);
1868 if (n->outcnt() == 0) continue; // Ignore dead
1869 uint j;
1870 for( j = 0; j < loop->_body.size(); j++ )
1871 if( loop->_body.at(j) == n )
1872 break;
1873 if( j == loop->_body.size() ) { // Not found in loop body
1874 // Last ditch effort to avoid assertion: Its possible that we
1875 // have some users (so outcnt not zero) but are still dead.
1876 // Try to find from root.
1877 if (Compile::current()->root()->find(n->_idx)) {
1878 fail++;
1879 tty->print("We have that verify does not: ");
1880 n->dump();
1881 }
1882 }
1883 }
1884 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
1885 Node *n = loop->_body.at(i2);
1886 if (n->outcnt() == 0) continue; // Ignore dead
1887 uint j;
1888 for( j = 0; j < _body.size(); j++ )
1889 if( _body.at(j) == n )
1890 break;
1891 if( j == _body.size() ) { // Not found in loop body
1892 // Last ditch effort to avoid assertion: Its possible that we
1893 // have some users (so outcnt not zero) but are still dead.
1894 // Try to find from root.
1895 if (Compile::current()->root()->find(n->_idx)) {
1896 fail++;
1897 tty->print("Verify has that we do not: ");
1898 n->dump();
1899 }
1900 }
1901 }
1902 assert( !fail, "loop body mismatch" );
1903 }
1904 }
1906 #endif
1908 //------------------------------set_idom---------------------------------------
1909 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
1910 uint idx = d->_idx;
1911 if (idx >= _idom_size) {
1912 uint newsize = _idom_size<<1;
1913 while( idx >= newsize ) {
1914 newsize <<= 1;
1915 }
1916 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
1917 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
1918 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
1919 _idom_size = newsize;
1920 }
1921 _idom[idx] = n;
1922 _dom_depth[idx] = dom_depth;
1923 }
1925 //------------------------------recompute_dom_depth---------------------------------------
1926 // The dominator tree is constructed with only parent pointers.
1927 // This recomputes the depth in the tree by first tagging all
1928 // nodes as "no depth yet" marker. The next pass then runs up
1929 // the dom tree from each node marked "no depth yet", and computes
1930 // the depth on the way back down.
1931 void PhaseIdealLoop::recompute_dom_depth() {
1932 uint no_depth_marker = C->unique();
1933 uint i;
1934 // Initialize depth to "no depth yet"
1935 for (i = 0; i < _idom_size; i++) {
1936 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
1937 _dom_depth[i] = no_depth_marker;
1938 }
1939 }
1940 if (_dom_stk == NULL) {
1941 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
1942 if (init_size < 10) init_size = 10;
1943 _dom_stk = new (C->node_arena()) GrowableArray<uint>(C->node_arena(), init_size, 0, 0);
1944 }
1945 // Compute new depth for each node.
1946 for (i = 0; i < _idom_size; i++) {
1947 uint j = i;
1948 // Run up the dom tree to find a node with a depth
1949 while (_dom_depth[j] == no_depth_marker) {
1950 _dom_stk->push(j);
1951 j = _idom[j]->_idx;
1952 }
1953 // Compute the depth on the way back down this tree branch
1954 uint dd = _dom_depth[j] + 1;
1955 while (_dom_stk->length() > 0) {
1956 uint j = _dom_stk->pop();
1957 _dom_depth[j] = dd;
1958 dd++;
1959 }
1960 }
1961 }
1963 //------------------------------sort-------------------------------------------
1964 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
1965 // loop tree, not the root.
1966 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
1967 if( !innermost ) return loop; // New innermost loop
1969 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
1970 assert( loop_preorder, "not yet post-walked loop" );
1971 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
1972 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
1974 // Insert at start of list
1975 while( l ) { // Insertion sort based on pre-order
1976 if( l == loop ) return innermost; // Already on list!
1977 int l_preorder = get_preorder(l->_head); // Cache pre-order number
1978 assert( l_preorder, "not yet post-walked l" );
1979 // Check header pre-order number to figure proper nesting
1980 if( loop_preorder > l_preorder )
1981 break; // End of insertion
1982 // If headers tie (e.g., shared headers) check tail pre-order numbers.
1983 // Since I split shared headers, you'd think this could not happen.
1984 // BUT: I must first do the preorder numbering before I can discover I
1985 // have shared headers, so the split headers all get the same preorder
1986 // number as the RegionNode they split from.
1987 if( loop_preorder == l_preorder &&
1988 get_preorder(loop->_tail) < get_preorder(l->_tail) )
1989 break; // Also check for shared headers (same pre#)
1990 pp = &l->_parent; // Chain up list
1991 l = *pp;
1992 }
1993 // Link into list
1994 // Point predecessor to me
1995 *pp = loop;
1996 // Point me to successor
1997 IdealLoopTree *p = loop->_parent;
1998 loop->_parent = l; // Point me to successor
1999 if( p ) sort( p, innermost ); // Insert my parents into list as well
2000 return innermost;
2001 }
2003 //------------------------------build_loop_tree--------------------------------
2004 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2005 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2006 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2007 // tightest enclosing IdealLoopTree for post-walked.
2008 //
2009 // During my forward walk I do a short 1-layer lookahead to see if I can find
2010 // a loop backedge with that doesn't have any work on the backedge. This
2011 // helps me construct nested loops with shared headers better.
2012 //
2013 // Once I've done the forward recursion, I do the post-work. For each child
2014 // I check to see if there is a backedge. Backedges define a loop! I
2015 // insert an IdealLoopTree at the target of the backedge.
2016 //
2017 // During the post-work I also check to see if I have several children
2018 // belonging to different loops. If so, then this Node is a decision point
2019 // where control flow can choose to change loop nests. It is at this
2020 // decision point where I can figure out how loops are nested. At this
2021 // time I can properly order the different loop nests from my children.
2022 // Note that there may not be any backedges at the decision point!
2023 //
2024 // Since the decision point can be far removed from the backedges, I can't
2025 // order my loops at the time I discover them. Thus at the decision point
2026 // I need to inspect loop header pre-order numbers to properly nest my
2027 // loops. This means I need to sort my childrens' loops by pre-order.
2028 // The sort is of size number-of-control-children, which generally limits
2029 // it to size 2 (i.e., I just choose between my 2 target loops).
2030 void PhaseIdealLoop::build_loop_tree() {
2031 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2032 GrowableArray <Node *> bltstack(C->unique() >> 1);
2033 Node *n = C->root();
2034 bltstack.push(n);
2035 int pre_order = 1;
2036 int stack_size;
2038 while ( ( stack_size = bltstack.length() ) != 0 ) {
2039 n = bltstack.top(); // Leave node on stack
2040 if ( !is_visited(n) ) {
2041 // ---- Pre-pass Work ----
2042 // Pre-walked but not post-walked nodes need a pre_order number.
2044 set_preorder_visited( n, pre_order ); // set as visited
2046 // ---- Scan over children ----
2047 // Scan first over control projections that lead to loop headers.
2048 // This helps us find inner-to-outer loops with shared headers better.
2050 // Scan children's children for loop headers.
2051 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2052 Node* m = n->raw_out(i); // Child
2053 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2054 // Scan over children's children to find loop
2055 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2056 Node* l = m->fast_out(j);
2057 if( is_visited(l) && // Been visited?
2058 !is_postvisited(l) && // But not post-visited
2059 get_preorder(l) < pre_order ) { // And smaller pre-order
2060 // Found! Scan the DFS down this path before doing other paths
2061 bltstack.push(m);
2062 break;
2063 }
2064 }
2065 }
2066 }
2067 pre_order++;
2068 }
2069 else if ( !is_postvisited(n) ) {
2070 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2071 // such as com.sun.rsasign.am::a.
2072 // For non-recursive version, first, process current children.
2073 // On next iteration, check if additional children were added.
2074 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2075 Node* u = n->raw_out(k);
2076 if ( u->is_CFG() && !is_visited(u) ) {
2077 bltstack.push(u);
2078 }
2079 }
2080 if ( bltstack.length() == stack_size ) {
2081 // There were no additional children, post visit node now
2082 (void)bltstack.pop(); // Remove node from stack
2083 pre_order = build_loop_tree_impl( n, pre_order );
2084 // Check for bailout
2085 if (C->failing()) {
2086 return;
2087 }
2088 // Check to grow _preorders[] array for the case when
2089 // build_loop_tree_impl() adds new nodes.
2090 check_grow_preorders();
2091 }
2092 }
2093 else {
2094 (void)bltstack.pop(); // Remove post-visited node from stack
2095 }
2096 }
2097 }
2099 //------------------------------build_loop_tree_impl---------------------------
2100 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2101 // ---- Post-pass Work ----
2102 // Pre-walked but not post-walked nodes need a pre_order number.
2104 // Tightest enclosing loop for this Node
2105 IdealLoopTree *innermost = NULL;
2107 // For all children, see if any edge is a backedge. If so, make a loop
2108 // for it. Then find the tightest enclosing loop for the self Node.
2109 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2110 Node* m = n->fast_out(i); // Child
2111 if( n == m ) continue; // Ignore control self-cycles
2112 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2114 IdealLoopTree *l; // Child's loop
2115 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2116 // Found a backedge
2117 assert( get_preorder(m) < pre_order, "should be backedge" );
2118 // Check for the RootNode, which is already a LoopNode and is allowed
2119 // to have multiple "backedges".
2120 if( m == C->root()) { // Found the root?
2121 l = _ltree_root; // Root is the outermost LoopNode
2122 } else { // Else found a nested loop
2123 // Insert a LoopNode to mark this loop.
2124 l = new IdealLoopTree(this, m, n);
2125 } // End of Else found a nested loop
2126 if( !has_loop(m) ) // If 'm' does not already have a loop set
2127 set_loop(m, l); // Set loop header to loop now
2129 } else { // Else not a nested loop
2130 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2131 l = get_loop(m); // Get previously determined loop
2132 // If successor is header of a loop (nest), move up-loop till it
2133 // is a member of some outer enclosing loop. Since there are no
2134 // shared headers (I've split them already) I only need to go up
2135 // at most 1 level.
2136 while( l && l->_head == m ) // Successor heads loop?
2137 l = l->_parent; // Move up 1 for me
2138 // If this loop is not properly parented, then this loop
2139 // has no exit path out, i.e. its an infinite loop.
2140 if( !l ) {
2141 // Make loop "reachable" from root so the CFG is reachable. Basically
2142 // insert a bogus loop exit that is never taken. 'm', the loop head,
2143 // points to 'n', one (of possibly many) fall-in paths. There may be
2144 // many backedges as well.
2146 // Here I set the loop to be the root loop. I could have, after
2147 // inserting a bogus loop exit, restarted the recursion and found my
2148 // new loop exit. This would make the infinite loop a first-class
2149 // loop and it would then get properly optimized. What's the use of
2150 // optimizing an infinite loop?
2151 l = _ltree_root; // Oops, found infinite loop
2153 if (!_verify_only) {
2154 // Insert the NeverBranch between 'm' and it's control user.
2155 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2156 _igvn.register_new_node_with_optimizer(iff);
2157 set_loop(iff, l);
2158 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2159 _igvn.register_new_node_with_optimizer(if_t);
2160 set_loop(if_t, l);
2162 Node* cfg = NULL; // Find the One True Control User of m
2163 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2164 Node* x = m->fast_out(j);
2165 if (x->is_CFG() && x != m && x != iff)
2166 { cfg = x; break; }
2167 }
2168 assert(cfg != NULL, "must find the control user of m");
2169 uint k = 0; // Probably cfg->in(0)
2170 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2171 cfg->set_req( k, if_t ); // Now point to NeverBranch
2173 // Now create the never-taken loop exit
2174 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2175 _igvn.register_new_node_with_optimizer(if_f);
2176 set_loop(if_f, l);
2177 // Find frame ptr for Halt. Relies on the optimizer
2178 // V-N'ing. Easier and quicker than searching through
2179 // the program structure.
2180 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2181 _igvn.register_new_node_with_optimizer(frame);
2182 // Halt & Catch Fire
2183 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2184 _igvn.register_new_node_with_optimizer(halt);
2185 set_loop(halt, l);
2186 C->root()->add_req(halt);
2187 }
2188 set_loop(C->root(), _ltree_root);
2189 }
2190 }
2191 // Weeny check for irreducible. This child was already visited (this
2192 // IS the post-work phase). Is this child's loop header post-visited
2193 // as well? If so, then I found another entry into the loop.
2194 if (!_verify_only) {
2195 while( is_postvisited(l->_head) ) {
2196 // found irreducible
2197 l->_irreducible = 1; // = true
2198 l = l->_parent;
2199 _has_irreducible_loops = true;
2200 // Check for bad CFG here to prevent crash, and bailout of compile
2201 if (l == NULL) {
2202 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2203 return pre_order;
2204 }
2205 }
2206 }
2208 // This Node might be a decision point for loops. It is only if
2209 // it's children belong to several different loops. The sort call
2210 // does a trivial amount of work if there is only 1 child or all
2211 // children belong to the same loop. If however, the children
2212 // belong to different loops, the sort call will properly set the
2213 // _parent pointers to show how the loops nest.
2214 //
2215 // In any case, it returns the tightest enclosing loop.
2216 innermost = sort( l, innermost );
2217 }
2219 // Def-use info will have some dead stuff; dead stuff will have no
2220 // loop decided on.
2222 // Am I a loop header? If so fix up my parent's child and next ptrs.
2223 if( innermost && innermost->_head == n ) {
2224 assert( get_loop(n) == innermost, "" );
2225 IdealLoopTree *p = innermost->_parent;
2226 IdealLoopTree *l = innermost;
2227 while( p && l->_head == n ) {
2228 l->_next = p->_child; // Put self on parents 'next child'
2229 p->_child = l; // Make self as first child of parent
2230 l = p; // Now walk up the parent chain
2231 p = l->_parent;
2232 }
2233 } else {
2234 // Note that it is possible for a LoopNode to reach here, if the
2235 // backedge has been made unreachable (hence the LoopNode no longer
2236 // denotes a Loop, and will eventually be removed).
2238 // Record tightest enclosing loop for self. Mark as post-visited.
2239 set_loop(n, innermost);
2240 // Also record has_call flag early on
2241 if( innermost ) {
2242 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2243 // Do not count uncommon calls
2244 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2245 Node *iff = n->in(0)->in(0);
2246 if( !iff->is_If() ||
2247 (n->in(0)->Opcode() == Op_IfFalse &&
2248 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2249 (iff->as_If()->_prob >= 0.01) )
2250 innermost->_has_call = 1;
2251 }
2252 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2253 // Disable loop optimizations if the loop has a scalar replaceable
2254 // allocation. This disabling may cause a potential performance lost
2255 // if the allocation is not eliminated for some reason.
2256 innermost->_allow_optimizations = false;
2257 innermost->_has_call = 1; // = true
2258 }
2259 }
2260 }
2262 // Flag as post-visited now
2263 set_postvisited(n);
2264 return pre_order;
2265 }
2268 //------------------------------build_loop_early-------------------------------
2269 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2270 // First pass computes the earliest controlling node possible. This is the
2271 // controlling input with the deepest dominating depth.
2272 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2273 while (worklist.size() != 0) {
2274 // Use local variables nstack_top_n & nstack_top_i to cache values
2275 // on nstack's top.
2276 Node *nstack_top_n = worklist.pop();
2277 uint nstack_top_i = 0;
2278 //while_nstack_nonempty:
2279 while (true) {
2280 // Get parent node and next input's index from stack's top.
2281 Node *n = nstack_top_n;
2282 uint i = nstack_top_i;
2283 uint cnt = n->req(); // Count of inputs
2284 if (i == 0) { // Pre-process the node.
2285 if( has_node(n) && // Have either loop or control already?
2286 !has_ctrl(n) ) { // Have loop picked out already?
2287 // During "merge_many_backedges" we fold up several nested loops
2288 // into a single loop. This makes the members of the original
2289 // loop bodies pointing to dead loops; they need to move up
2290 // to the new UNION'd larger loop. I set the _head field of these
2291 // dead loops to NULL and the _parent field points to the owning
2292 // loop. Shades of UNION-FIND algorithm.
2293 IdealLoopTree *ilt;
2294 while( !(ilt = get_loop(n))->_head ) {
2295 // Normally I would use a set_loop here. But in this one special
2296 // case, it is legal (and expected) to change what loop a Node
2297 // belongs to.
2298 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2299 }
2300 // Remove safepoints ONLY if I've already seen I don't need one.
2301 // (the old code here would yank a 2nd safepoint after seeing a
2302 // first one, even though the 1st did not dominate in the loop body
2303 // and thus could be avoided indefinitely)
2304 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2305 is_deleteable_safept(n)) {
2306 Node *in = n->in(TypeFunc::Control);
2307 lazy_replace(n,in); // Pull safepoint now
2308 // Carry on with the recursion "as if" we are walking
2309 // only the control input
2310 if( !visited.test_set( in->_idx ) ) {
2311 worklist.push(in); // Visit this guy later, using worklist
2312 }
2313 // Get next node from nstack:
2314 // - skip n's inputs processing by setting i > cnt;
2315 // - we also will not call set_early_ctrl(n) since
2316 // has_node(n) == true (see the condition above).
2317 i = cnt + 1;
2318 }
2319 }
2320 } // if (i == 0)
2322 // Visit all inputs
2323 bool done = true; // Assume all n's inputs will be processed
2324 while (i < cnt) {
2325 Node *in = n->in(i);
2326 ++i;
2327 if (in == NULL) continue;
2328 if (in->pinned() && !in->is_CFG())
2329 set_ctrl(in, in->in(0));
2330 int is_visited = visited.test_set( in->_idx );
2331 if (!has_node(in)) { // No controlling input yet?
2332 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2333 assert( !is_visited, "visit only once" );
2334 nstack.push(n, i); // Save parent node and next input's index.
2335 nstack_top_n = in; // Process current input now.
2336 nstack_top_i = 0;
2337 done = false; // Not all n's inputs processed.
2338 break; // continue while_nstack_nonempty;
2339 } else if (!is_visited) {
2340 // This guy has a location picked out for him, but has not yet
2341 // been visited. Happens to all CFG nodes, for instance.
2342 // Visit him using the worklist instead of recursion, to break
2343 // cycles. Since he has a location already we do not need to
2344 // find his location before proceeding with the current Node.
2345 worklist.push(in); // Visit this guy later, using worklist
2346 }
2347 }
2348 if (done) {
2349 // All of n's inputs have been processed, complete post-processing.
2351 // Compute earliest point this Node can go.
2352 // CFG, Phi, pinned nodes already know their controlling input.
2353 if (!has_node(n)) {
2354 // Record earliest legal location
2355 set_early_ctrl( n );
2356 }
2357 if (nstack.is_empty()) {
2358 // Finished all nodes on stack.
2359 // Process next node on the worklist.
2360 break;
2361 }
2362 // Get saved parent node and next input's index.
2363 nstack_top_n = nstack.node();
2364 nstack_top_i = nstack.index();
2365 nstack.pop();
2366 }
2367 } // while (true)
2368 }
2369 }
2371 //------------------------------dom_lca_internal--------------------------------
2372 // Pair-wise LCA
2373 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2374 if( !n1 ) return n2; // Handle NULL original LCA
2375 assert( n1->is_CFG(), "" );
2376 assert( n2->is_CFG(), "" );
2377 // find LCA of all uses
2378 uint d1 = dom_depth(n1);
2379 uint d2 = dom_depth(n2);
2380 while (n1 != n2) {
2381 if (d1 > d2) {
2382 n1 = idom(n1);
2383 d1 = dom_depth(n1);
2384 } else if (d1 < d2) {
2385 n2 = idom(n2);
2386 d2 = dom_depth(n2);
2387 } else {
2388 // Here d1 == d2. Due to edits of the dominator-tree, sections
2389 // of the tree might have the same depth. These sections have
2390 // to be searched more carefully.
2392 // Scan up all the n1's with equal depth, looking for n2.
2393 Node *t1 = idom(n1);
2394 while (dom_depth(t1) == d1) {
2395 if (t1 == n2) return n2;
2396 t1 = idom(t1);
2397 }
2398 // Scan up all the n2's with equal depth, looking for n1.
2399 Node *t2 = idom(n2);
2400 while (dom_depth(t2) == d2) {
2401 if (t2 == n1) return n1;
2402 t2 = idom(t2);
2403 }
2404 // Move up to a new dominator-depth value as well as up the dom-tree.
2405 n1 = t1;
2406 n2 = t2;
2407 d1 = dom_depth(n1);
2408 d2 = dom_depth(n2);
2409 }
2410 }
2411 return n1;
2412 }
2414 //------------------------------compute_idom-----------------------------------
2415 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2416 // IDOMs are correct.
2417 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2418 assert( region->is_Region(), "" );
2419 Node *LCA = NULL;
2420 for( uint i = 1; i < region->req(); i++ ) {
2421 if( region->in(i) != C->top() )
2422 LCA = dom_lca( LCA, region->in(i) );
2423 }
2424 return LCA;
2425 }
2427 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
2428 bool had_error = false;
2429 #ifdef ASSERT
2430 if (early != C->root()) {
2431 // Make sure that there's a dominance path from use to LCA
2432 Node* d = use;
2433 while (d != LCA) {
2434 d = idom(d);
2435 if (d == C->root()) {
2436 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
2437 n->dump();
2438 use->dump();
2439 had_error = true;
2440 break;
2441 }
2442 }
2443 }
2444 #endif
2445 return had_error;
2446 }
2449 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
2450 // Compute LCA over list of uses
2451 bool had_error = false;
2452 Node *LCA = NULL;
2453 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2454 Node* c = n->fast_out(i);
2455 if (_nodes[c->_idx] == NULL)
2456 continue; // Skip the occasional dead node
2457 if( c->is_Phi() ) { // For Phis, we must land above on the path
2458 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2459 if( c->in(j) == n ) { // Found matching input?
2460 Node *use = c->in(0)->in(j);
2461 if (_verify_only && use->is_top()) continue;
2462 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2463 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2464 }
2465 }
2466 } else {
2467 // For CFG data-users, use is in the block just prior
2468 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2469 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2470 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2471 }
2472 }
2473 assert(!had_error, "bad dominance");
2474 return LCA;
2475 }
2477 //------------------------------get_late_ctrl----------------------------------
2478 // Compute latest legal control.
2479 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2480 assert(early != NULL, "early control should not be NULL");
2482 Node* LCA = compute_lca_of_uses(n, early);
2483 #ifdef ASSERT
2484 if (LCA == C->root() && LCA != early) {
2485 // def doesn't dominate uses so print some useful debugging output
2486 compute_lca_of_uses(n, early, true);
2487 }
2488 #endif
2490 // if this is a load, check for anti-dependent stores
2491 // We use a conservative algorithm to identify potential interfering
2492 // instructions and for rescheduling the load. The users of the memory
2493 // input of this load are examined. Any use which is not a load and is
2494 // dominated by early is considered a potentially interfering store.
2495 // This can produce false positives.
2496 if (n->is_Load() && LCA != early) {
2497 Node_List worklist;
2499 Node *mem = n->in(MemNode::Memory);
2500 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2501 Node* s = mem->fast_out(i);
2502 worklist.push(s);
2503 }
2504 while(worklist.size() != 0 && LCA != early) {
2505 Node* s = worklist.pop();
2506 if (s->is_Load()) {
2507 continue;
2508 } else if (s->is_MergeMem()) {
2509 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
2510 Node* s1 = s->fast_out(i);
2511 worklist.push(s1);
2512 }
2513 } else {
2514 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
2515 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
2516 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
2517 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
2518 }
2519 }
2520 }
2521 }
2523 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
2524 return LCA;
2525 }
2527 // true if CFG node d dominates CFG node n
2528 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
2529 if (d == n)
2530 return true;
2531 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
2532 uint dd = dom_depth(d);
2533 while (dom_depth(n) >= dd) {
2534 if (n == d)
2535 return true;
2536 n = idom(n);
2537 }
2538 return false;
2539 }
2541 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
2542 // Pair-wise LCA with tags.
2543 // Tag each index with the node 'tag' currently being processed
2544 // before advancing up the dominator chain using idom().
2545 // Later calls that find a match to 'tag' know that this path has already
2546 // been considered in the current LCA (which is input 'n1' by convention).
2547 // Since get_late_ctrl() is only called once for each node, the tag array
2548 // does not need to be cleared between calls to get_late_ctrl().
2549 // Algorithm trades a larger constant factor for better asymptotic behavior
2550 //
2551 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
2552 uint d1 = dom_depth(n1);
2553 uint d2 = dom_depth(n2);
2555 do {
2556 if (d1 > d2) {
2557 // current lca is deeper than n2
2558 _dom_lca_tags.map(n1->_idx, tag);
2559 n1 = idom(n1);
2560 d1 = dom_depth(n1);
2561 } else if (d1 < d2) {
2562 // n2 is deeper than current lca
2563 Node *memo = _dom_lca_tags[n2->_idx];
2564 if( memo == tag ) {
2565 return n1; // Return the current LCA
2566 }
2567 _dom_lca_tags.map(n2->_idx, tag);
2568 n2 = idom(n2);
2569 d2 = dom_depth(n2);
2570 } else {
2571 // Here d1 == d2. Due to edits of the dominator-tree, sections
2572 // of the tree might have the same depth. These sections have
2573 // to be searched more carefully.
2575 // Scan up all the n1's with equal depth, looking for n2.
2576 _dom_lca_tags.map(n1->_idx, tag);
2577 Node *t1 = idom(n1);
2578 while (dom_depth(t1) == d1) {
2579 if (t1 == n2) return n2;
2580 _dom_lca_tags.map(t1->_idx, tag);
2581 t1 = idom(t1);
2582 }
2583 // Scan up all the n2's with equal depth, looking for n1.
2584 _dom_lca_tags.map(n2->_idx, tag);
2585 Node *t2 = idom(n2);
2586 while (dom_depth(t2) == d2) {
2587 if (t2 == n1) return n1;
2588 _dom_lca_tags.map(t2->_idx, tag);
2589 t2 = idom(t2);
2590 }
2591 // Move up to a new dominator-depth value as well as up the dom-tree.
2592 n1 = t1;
2593 n2 = t2;
2594 d1 = dom_depth(n1);
2595 d2 = dom_depth(n2);
2596 }
2597 } while (n1 != n2);
2598 return n1;
2599 }
2601 //------------------------------init_dom_lca_tags------------------------------
2602 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2603 // Intended use does not involve any growth for the array, so it could
2604 // be of fixed size.
2605 void PhaseIdealLoop::init_dom_lca_tags() {
2606 uint limit = C->unique() + 1;
2607 _dom_lca_tags.map( limit, NULL );
2608 #ifdef ASSERT
2609 for( uint i = 0; i < limit; ++i ) {
2610 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2611 }
2612 #endif // ASSERT
2613 }
2615 //------------------------------clear_dom_lca_tags------------------------------
2616 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2617 // Intended use does not involve any growth for the array, so it could
2618 // be of fixed size.
2619 void PhaseIdealLoop::clear_dom_lca_tags() {
2620 uint limit = C->unique() + 1;
2621 _dom_lca_tags.map( limit, NULL );
2622 _dom_lca_tags.clear();
2623 #ifdef ASSERT
2624 for( uint i = 0; i < limit; ++i ) {
2625 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2626 }
2627 #endif // ASSERT
2628 }
2630 //------------------------------build_loop_late--------------------------------
2631 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2632 // Second pass finds latest legal placement, and ideal loop placement.
2633 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2634 while (worklist.size() != 0) {
2635 Node *n = worklist.pop();
2636 // Only visit once
2637 if (visited.test_set(n->_idx)) continue;
2638 uint cnt = n->outcnt();
2639 uint i = 0;
2640 while (true) {
2641 assert( _nodes[n->_idx], "no dead nodes" );
2642 // Visit all children
2643 if (i < cnt) {
2644 Node* use = n->raw_out(i);
2645 ++i;
2646 // Check for dead uses. Aggressively prune such junk. It might be
2647 // dead in the global sense, but still have local uses so I cannot
2648 // easily call 'remove_dead_node'.
2649 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
2650 // Due to cycles, we might not hit the same fixed point in the verify
2651 // pass as we do in the regular pass. Instead, visit such phis as
2652 // simple uses of the loop head.
2653 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
2654 if( !visited.test(use->_idx) )
2655 worklist.push(use);
2656 } else if( !visited.test_set(use->_idx) ) {
2657 nstack.push(n, i); // Save parent and next use's index.
2658 n = use; // Process all children of current use.
2659 cnt = use->outcnt();
2660 i = 0;
2661 }
2662 } else {
2663 // Do not visit around the backedge of loops via data edges.
2664 // push dead code onto a worklist
2665 _deadlist.push(use);
2666 }
2667 } else {
2668 // All of n's children have been processed, complete post-processing.
2669 build_loop_late_post(n);
2670 if (nstack.is_empty()) {
2671 // Finished all nodes on stack.
2672 // Process next node on the worklist.
2673 break;
2674 }
2675 // Get saved parent node and next use's index. Visit the rest of uses.
2676 n = nstack.node();
2677 cnt = n->outcnt();
2678 i = nstack.index();
2679 nstack.pop();
2680 }
2681 }
2682 }
2683 }
2685 //------------------------------build_loop_late_post---------------------------
2686 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2687 // Second pass finds latest legal placement, and ideal loop placement.
2688 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
2690 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
2691 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
2692 }
2694 // CFG and pinned nodes already handled
2695 if( n->in(0) ) {
2696 if( n->in(0)->is_top() ) return; // Dead?
2698 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
2699 // _must_ be pinned (they have to observe their control edge of course).
2700 // Unlike Stores (which modify an unallocable resource, the memory
2701 // state), Mods/Loads can float around. So free them up.
2702 bool pinned = true;
2703 switch( n->Opcode() ) {
2704 case Op_DivI:
2705 case Op_DivF:
2706 case Op_DivD:
2707 case Op_ModI:
2708 case Op_ModF:
2709 case Op_ModD:
2710 case Op_LoadB: // Same with Loads; they can sink
2711 case Op_LoadUS: // during loop optimizations.
2712 case Op_LoadD:
2713 case Op_LoadF:
2714 case Op_LoadI:
2715 case Op_LoadKlass:
2716 case Op_LoadNKlass:
2717 case Op_LoadL:
2718 case Op_LoadS:
2719 case Op_LoadP:
2720 case Op_LoadN:
2721 case Op_LoadRange:
2722 case Op_LoadD_unaligned:
2723 case Op_LoadL_unaligned:
2724 case Op_StrComp: // Does a bunch of load-like effects
2725 case Op_StrEquals:
2726 case Op_StrIndexOf:
2727 case Op_AryEq:
2728 pinned = false;
2729 }
2730 if( pinned ) {
2731 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
2732 if( !chosen_loop->_child ) // Inner loop?
2733 chosen_loop->_body.push(n); // Collect inner loops
2734 return;
2735 }
2736 } else { // No slot zero
2737 if( n->is_CFG() ) { // CFG with no slot 0 is dead
2738 _nodes.map(n->_idx,0); // No block setting, it's globally dead
2739 return;
2740 }
2741 assert(!n->is_CFG() || n->outcnt() == 0, "");
2742 }
2744 // Do I have a "safe range" I can select over?
2745 Node *early = get_ctrl(n);// Early location already computed
2747 // Compute latest point this Node can go
2748 Node *LCA = get_late_ctrl( n, early );
2749 // LCA is NULL due to uses being dead
2750 if( LCA == NULL ) {
2751 #ifdef ASSERT
2752 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
2753 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
2754 }
2755 #endif
2756 _nodes.map(n->_idx, 0); // This node is useless
2757 _deadlist.push(n);
2758 return;
2759 }
2760 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
2762 Node *legal = LCA; // Walk 'legal' up the IDOM chain
2763 Node *least = legal; // Best legal position so far
2764 while( early != legal ) { // While not at earliest legal
2765 // Find least loop nesting depth
2766 legal = idom(legal); // Bump up the IDOM tree
2767 // Check for lower nesting depth
2768 if( get_loop(legal)->_nest < get_loop(least)->_nest )
2769 least = legal;
2770 }
2771 assert(early == legal || legal != C->root(), "bad dominance of inputs");
2773 // Try not to place code on a loop entry projection
2774 // which can inhibit range check elimination.
2775 if (least != early) {
2776 Node* ctrl_out = least->unique_ctrl_out();
2777 if (ctrl_out && ctrl_out->is_CountedLoop() &&
2778 least == ctrl_out->in(LoopNode::EntryControl)) {
2779 Node* least_dom = idom(least);
2780 if (get_loop(least_dom)->is_member(get_loop(least))) {
2781 least = least_dom;
2782 }
2783 }
2784 }
2786 #ifdef ASSERT
2787 // If verifying, verify that 'verify_me' has a legal location
2788 // and choose it as our location.
2789 if( _verify_me ) {
2790 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
2791 Node *legal = LCA;
2792 while( early != legal ) { // While not at earliest legal
2793 if( legal == v_ctrl ) break; // Check for prior good location
2794 legal = idom(legal) ;// Bump up the IDOM tree
2795 }
2796 // Check for prior good location
2797 if( legal == v_ctrl ) least = legal; // Keep prior if found
2798 }
2799 #endif
2801 // Assign discovered "here or above" point
2802 least = find_non_split_ctrl(least);
2803 set_ctrl(n, least);
2805 // Collect inner loop bodies
2806 IdealLoopTree *chosen_loop = get_loop(least);
2807 if( !chosen_loop->_child ) // Inner loop?
2808 chosen_loop->_body.push(n);// Collect inner loops
2809 }
2811 #ifndef PRODUCT
2812 //------------------------------dump-------------------------------------------
2813 void PhaseIdealLoop::dump( ) const {
2814 ResourceMark rm;
2815 Arena* arena = Thread::current()->resource_area();
2816 Node_Stack stack(arena, C->unique() >> 2);
2817 Node_List rpo_list;
2818 VectorSet visited(arena);
2819 visited.set(C->top()->_idx);
2820 rpo( C->root(), stack, visited, rpo_list );
2821 // Dump root loop indexed by last element in PO order
2822 dump( _ltree_root, rpo_list.size(), rpo_list );
2823 }
2825 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
2826 loop->dump_head();
2828 // Now scan for CFG nodes in the same loop
2829 for( uint j=idx; j > 0; j-- ) {
2830 Node *n = rpo_list[j-1];
2831 if( !_nodes[n->_idx] ) // Skip dead nodes
2832 continue;
2833 if( get_loop(n) != loop ) { // Wrong loop nest
2834 if( get_loop(n)->_head == n && // Found nested loop?
2835 get_loop(n)->_parent == loop )
2836 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
2837 continue;
2838 }
2840 // Dump controlling node
2841 for( uint x = 0; x < loop->_nest; x++ )
2842 tty->print(" ");
2843 tty->print("C");
2844 if( n == C->root() ) {
2845 n->dump();
2846 } else {
2847 Node* cached_idom = idom_no_update(n);
2848 Node *computed_idom = n->in(0);
2849 if( n->is_Region() ) {
2850 computed_idom = compute_idom(n);
2851 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
2852 // any MultiBranch ctrl node), so apply a similar transform to
2853 // the cached idom returned from idom_no_update.
2854 cached_idom = find_non_split_ctrl(cached_idom);
2855 }
2856 tty->print(" ID:%d",computed_idom->_idx);
2857 n->dump();
2858 if( cached_idom != computed_idom ) {
2859 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
2860 computed_idom->_idx, cached_idom->_idx);
2861 }
2862 }
2863 // Dump nodes it controls
2864 for( uint k = 0; k < _nodes.Size(); k++ ) {
2865 // (k < C->unique() && get_ctrl(find(k)) == n)
2866 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
2867 Node *m = C->root()->find(k);
2868 if( m && m->outcnt() > 0 ) {
2869 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
2870 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
2871 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
2872 }
2873 for( uint j = 0; j < loop->_nest; j++ )
2874 tty->print(" ");
2875 tty->print(" ");
2876 m->dump();
2877 }
2878 }
2879 }
2880 }
2881 }
2883 // Collect a R-P-O for the whole CFG.
2884 // Result list is in post-order (scan backwards for RPO)
2885 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
2886 stk.push(start, 0);
2887 visited.set(start->_idx);
2889 while (stk.is_nonempty()) {
2890 Node* m = stk.node();
2891 uint idx = stk.index();
2892 if (idx < m->outcnt()) {
2893 stk.set_index(idx + 1);
2894 Node* n = m->raw_out(idx);
2895 if (n->is_CFG() && !visited.test_set(n->_idx)) {
2896 stk.push(n, 0);
2897 }
2898 } else {
2899 rpo_list.push(m);
2900 stk.pop();
2901 }
2902 }
2903 }
2904 #endif
2907 //=============================================================================
2908 //------------------------------LoopTreeIterator-----------------------------------
2910 // Advance to next loop tree using a preorder, left-to-right traversal.
2911 void LoopTreeIterator::next() {
2912 assert(!done(), "must not be done.");
2913 if (_curnt->_child != NULL) {
2914 _curnt = _curnt->_child;
2915 } else if (_curnt->_next != NULL) {
2916 _curnt = _curnt->_next;
2917 } else {
2918 while (_curnt != _root && _curnt->_next == NULL) {
2919 _curnt = _curnt->_parent;
2920 }
2921 if (_curnt == _root) {
2922 _curnt = NULL;
2923 assert(done(), "must be done.");
2924 } else {
2925 assert(_curnt->_next != NULL, "must be more to do");
2926 _curnt = _curnt->_next;
2927 }
2928 }
2929 }