1 /*
2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/divnode.hpp"
38 #include "opto/idealGraphPrinter.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/mulnode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/superword.hpp"
43 #include "utilities/macros.hpp"
44 #if INCLUDE_SHENANDOAHGC
45 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
46 #endif
47
48 //=============================================================================
49 //------------------------------is_loop_iv-------------------------------------
50 // Determine if a node is Counted loop induction variable.
51 // The method is declared in node.hpp.
52 const Node* Node::is_loop_iv() const {
53 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
54 this->as_Phi()->region()->is_CountedLoop() &&
55 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
56 return this;
57 } else {
58 return NULL;
59 }
60 }
61
62 //=============================================================================
63 //------------------------------dump_spec--------------------------------------
64 // Dump special per-node info
65 #ifndef PRODUCT
66 void LoopNode::dump_spec(outputStream *st) const {
67 if (is_inner_loop()) st->print( "inner " );
68 if (is_partial_peel_loop()) st->print( "partial_peel " );
69 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
70 }
71 #endif
72
73 //------------------------------is_valid_counted_loop-------------------------
74 bool LoopNode::is_valid_counted_loop() const {
75 if (is_CountedLoop()) {
76 CountedLoopNode* l = as_CountedLoop();
77 CountedLoopEndNode* le = l->loopexit_or_null();
78 if (le != NULL &&
79 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
80 Node* phi = l->phi();
81 Node* exit = le->proj_out_or_null(0 /* false */);
82 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
83 phi != NULL && phi->is_Phi() &&
84 phi->in(LoopNode::LoopBackControl) == l->incr() &&
85 le->loopnode() == l && le->stride_is_con()) {
86 return true;
87 }
88 }
89 }
90 return false;
91 }
92
93 //------------------------------get_early_ctrl---------------------------------
94 // Compute earliest legal control
95 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
96 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
97 uint i;
98 Node *early;
99 if (n->in(0) && !n->is_expensive()) {
100 early = n->in(0);
101 if (!early->is_CFG()) // Might be a non-CFG multi-def
102 early = get_ctrl(early); // So treat input as a straight data input
103 i = 1;
104 } else {
105 early = get_ctrl(n->in(1));
106 i = 2;
107 }
108 uint e_d = dom_depth(early);
109 assert( early, "" );
110 for (; i < n->req(); i++) {
111 Node *cin = get_ctrl(n->in(i));
112 assert( cin, "" );
113 // Keep deepest dominator depth
114 uint c_d = dom_depth(cin);
115 if (c_d > e_d) { // Deeper guy?
116 early = cin; // Keep deepest found so far
117 e_d = c_d;
118 } else if (c_d == e_d && // Same depth?
119 early != cin) { // If not equal, must use slower algorithm
120 // If same depth but not equal, one _must_ dominate the other
121 // and we want the deeper (i.e., dominated) guy.
122 Node *n1 = early;
123 Node *n2 = cin;
124 while (1) {
125 n1 = idom(n1); // Walk up until break cycle
126 n2 = idom(n2);
127 if (n1 == cin || // Walked early up to cin
128 dom_depth(n2) < c_d)
129 break; // early is deeper; keep him
130 if (n2 == early || // Walked cin up to early
131 dom_depth(n1) < c_d) {
132 early = cin; // cin is deeper; keep him
133 break;
134 }
135 }
136 e_d = dom_depth(early); // Reset depth register cache
137 }
138 }
139
140 // Return earliest legal location
141 assert(early == find_non_split_ctrl(early), "unexpected early control");
142
143 if (n->is_expensive() && !_verify_only && !_verify_me) {
144 assert(n->in(0), "should have control input");
145 early = get_early_ctrl_for_expensive(n, early);
146 }
147
148 return early;
149 }
150
151 //------------------------------get_early_ctrl_for_expensive---------------------------------
152 // Move node up the dominator tree as high as legal while still beneficial
153 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
154 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
155 assert(OptimizeExpensiveOps, "optimization off?");
156
157 Node* ctl = n->in(0);
158 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
159 uint min_dom_depth = dom_depth(earliest);
160 #ifdef ASSERT
161 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
162 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
163 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
164 }
165 #endif
166 if (dom_depth(ctl) < min_dom_depth) {
167 return earliest;
168 }
169
170 while (1) {
171 Node *next = ctl;
172 // Moving the node out of a loop on the projection of a If
173 // confuses loop predication. So once we hit a Loop in a If branch
174 // that doesn't branch to an UNC, we stop. The code that process
175 // expensive nodes will notice the loop and skip over it to try to
176 // move the node further up.
177 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
178 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
179 break;
180 }
181 next = idom(ctl->in(1)->in(0));
182 } else if (ctl->is_Proj()) {
183 // We only move it up along a projection if the projection is
184 // the single control projection for its parent: same code path,
185 // if it's a If with UNC or fallthrough of a call.
186 Node* parent_ctl = ctl->in(0);
187 if (parent_ctl == NULL) {
188 break;
189 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
190 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
191 } else if (parent_ctl->is_If()) {
192 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
193 break;
194 }
195 assert(idom(ctl) == parent_ctl, "strange");
196 next = idom(parent_ctl);
197 } else if (ctl->is_CatchProj()) {
198 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
199 break;
200 }
201 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
202 next = parent_ctl->in(0)->in(0)->in(0);
203 } else {
204 // Check if parent control has a single projection (this
205 // control is the only possible successor of the parent
206 // control). If so, we can try to move the node above the
207 // parent control.
208 int nb_ctl_proj = 0;
209 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
210 Node *p = parent_ctl->fast_out(i);
211 if (p->is_Proj() && p->is_CFG()) {
212 nb_ctl_proj++;
213 if (nb_ctl_proj > 1) {
214 break;
215 }
216 }
217 }
218
219 if (nb_ctl_proj > 1) {
220 break;
221 }
222 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
223 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
224 assert(idom(ctl) == parent_ctl, "strange");
225 next = idom(parent_ctl);
226 }
227 } else {
228 next = idom(ctl);
229 }
230 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
231 break;
232 }
233 ctl = next;
234 }
235
236 if (ctl != n->in(0)) {
237 _igvn.replace_input_of(n, 0, ctl);
238 _igvn.hash_insert(n);
239 }
240
241 return ctl;
242 }
243
244
245 //------------------------------set_early_ctrl---------------------------------
246 // Set earliest legal control
247 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
248 Node *early = get_early_ctrl(n);
249
250 // Record earliest legal location
251 set_ctrl(n, early);
252 }
253
254 //------------------------------set_subtree_ctrl-------------------------------
255 // set missing _ctrl entries on new nodes
256 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
257 // Already set? Get out.
258 if( _nodes[n->_idx] ) return;
259 // Recursively set _nodes array to indicate where the Node goes
260 uint i;
261 for( i = 0; i < n->req(); ++i ) {
262 Node *m = n->in(i);
263 if( m && m != C->root() )
264 set_subtree_ctrl( m );
265 }
266
267 // Fixup self
268 set_early_ctrl( n );
269 }
270
271 // Create a skeleton strip mined outer loop: a Loop head before the
272 // inner strip mined loop, a safepoint and an exit condition guarded
273 // by an opaque node after the inner strip mined loop with a backedge
274 // to the loop head. The inner strip mined loop is left as it is. Only
275 // once loop optimizations are over, do we adjust the inner loop exit
276 // condition to limit its number of iterations, set the outer loop
277 // exit condition and add Phis to the outer loop head. Some loop
278 // optimizations that operate on the inner strip mined loop need to be
279 // aware of the outer strip mined loop: loop unswitching needs to
280 // clone the outer loop as well as the inner, unrolling needs to only
281 // clone the inner loop etc. No optimizations need to change the outer
282 // strip mined loop as it is only a skeleton.
283 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
284 IdealLoopTree* loop, float cl_prob, float le_fcnt,
285 Node*& entry_control, Node*& iffalse) {
286 Node* outer_test = _igvn.intcon(0);
287 set_ctrl(outer_test, C->root());
288 Node *orig = iffalse;
289 iffalse = iffalse->clone();
290 _igvn.register_new_node_with_optimizer(iffalse);
291 set_idom(iffalse, idom(orig), dom_depth(orig));
292
293 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
294 Node *outer_ift = new IfTrueNode (outer_le);
295 Node* outer_iff = orig;
296 _igvn.replace_input_of(outer_iff, 0, outer_le);
297
298 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
299 entry_control = outer_l;
300
301 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
302 IdealLoopTree* parent = loop->_parent;
303 IdealLoopTree* sibling = parent->_child;
304 if (sibling == loop) {
305 parent->_child = outer_ilt;
306 } else {
307 while (sibling->_next != loop) {
308 sibling = sibling->_next;
309 }
310 sibling->_next = outer_ilt;
311 }
312 outer_ilt->_next = loop->_next;
313 outer_ilt->_parent = parent;
314 outer_ilt->_child = loop;
315 outer_ilt->_nest = loop->_nest;
316 loop->_parent = outer_ilt;
317 loop->_next = NULL;
318 loop->_nest++;
319
320 set_loop(iffalse, outer_ilt);
321 register_control(outer_le, outer_ilt, iffalse);
322 register_control(outer_ift, outer_ilt, outer_le);
323 set_idom(outer_iff, outer_le, dom_depth(outer_le));
324 _igvn.register_new_node_with_optimizer(outer_l);
325 set_loop(outer_l, outer_ilt);
326 set_idom(outer_l, init_control, dom_depth(init_control)+1);
327
328 return outer_ilt;
329 }
330
331 void PhaseIdealLoop::insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol) {
332 Node* new_predicate_proj = create_new_if_for_predicate(limit_check_proj, NULL,
333 Deoptimization::Reason_loop_limit_check,
334 Op_If);
335 Node* iff = new_predicate_proj->in(0);
336 assert(iff->Opcode() == Op_If, "bad graph shape");
337 Node* conv = iff->in(1);
338 assert(conv->Opcode() == Op_Conv2B, "bad graph shape");
339 Node* opaq = conv->in(1);
340 assert(opaq->Opcode() == Op_Opaque1, "bad graph shape");
341 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
342 bol = _igvn.register_new_node_with_optimizer(bol);
343 set_subtree_ctrl(bol);
344 _igvn.replace_input_of(iff, 1, bol);
345
346 #ifndef PRODUCT
347 // report that the loop predication has been actually performed
348 // for this loop
349 if (TraceLoopLimitCheck) {
350 tty->print_cr("Counted Loop Limit Check generated:");
351 debug_only( bol->dump(2); )
352 }
353 #endif
354 }
355
356 //------------------------------is_counted_loop--------------------------------
357 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) {
358 PhaseGVN *gvn = &_igvn;
359
360 // Counted loop head must be a good RegionNode with only 3 not NULL
361 // control input edges: Self, Entry, LoopBack.
362 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
363 return false;
364 }
365 Node *init_control = x->in(LoopNode::EntryControl);
366 Node *back_control = x->in(LoopNode::LoopBackControl);
367 if (init_control == NULL || back_control == NULL) // Partially dead
368 return false;
369 // Must also check for TOP when looking for a dead loop
370 if (init_control->is_top() || back_control->is_top())
371 return false;
372
373 // Allow funny placement of Safepoint
374 if (back_control->Opcode() == Op_SafePoint) {
375 if (LoopStripMiningIter != 0) {
376 // Leaving the safepoint on the backedge and creating a
377 // CountedLoop will confuse optimizations. We can't move the
378 // safepoint around because its jvm state wouldn't match a new
379 // location. Give up on that loop.
380 return false;
381 }
382 back_control = back_control->in(TypeFunc::Control);
383 }
384
385 // Controlling test for loop
386 Node *iftrue = back_control;
387 uint iftrue_op = iftrue->Opcode();
388 if (iftrue_op != Op_IfTrue &&
389 iftrue_op != Op_IfFalse)
390 // I have a weird back-control. Probably the loop-exit test is in
391 // the middle of the loop and I am looking at some trailing control-flow
392 // merge point. To fix this I would have to partially peel the loop.
393 return false; // Obscure back-control
394
395 // Get boolean guarding loop-back test
396 Node *iff = iftrue->in(0);
397 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
398 return false;
399 BoolNode *test = iff->in(1)->as_Bool();
400 BoolTest::mask bt = test->_test._test;
401 float cl_prob = iff->as_If()->_prob;
402 if (iftrue_op == Op_IfFalse) {
403 bt = BoolTest(bt).negate();
404 cl_prob = 1.0 - cl_prob;
405 }
406 // Get backedge compare
407 Node *cmp = test->in(1);
408 int cmp_op = cmp->Opcode();
409 if (cmp_op != Op_CmpI)
410 return false; // Avoid pointer & float compares
411
412 // Find the trip-counter increment & limit. Limit must be loop invariant.
413 Node *incr = cmp->in(1);
414 Node *limit = cmp->in(2);
415
416 // ---------
417 // need 'loop()' test to tell if limit is loop invariant
418 // ---------
419
420 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
421 Node *tmp = incr; // Then reverse order into the CmpI
422 incr = limit;
423 limit = tmp;
424 bt = BoolTest(bt).commute(); // And commute the exit test
425 }
426 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
427 return false;
428 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
429 return false;
430
431 Node* phi_incr = NULL;
432 // Trip-counter increment must be commutative & associative.
433 if (incr->Opcode() == Op_CastII) {
434 incr = incr->in(1);
435 }
436 if (incr->is_Phi()) {
437 if (incr->as_Phi()->region() != x || incr->req() != 3)
438 return false; // Not simple trip counter expression
439 phi_incr = incr;
440 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
441 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
442 return false;
443 }
444
445 Node* trunc1 = NULL;
446 Node* trunc2 = NULL;
447 const TypeInt* iv_trunc_t = NULL;
448 Node* orig_incr = incr;
449 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
450 return false; // Funny increment opcode
451 }
452 assert(incr->Opcode() == Op_AddI, "wrong increment code");
453
454 const TypeInt* limit_t = gvn->type(limit)->is_int();
455 if (trunc1 != NULL) {
456 // When there is a truncation, we must be sure that after the truncation
457 // the trip counter will end up higher than the limit, otherwise we are looking
458 // at an endless loop. Can happen with range checks.
459
460 // Example:
461 // int i = 0;
462 // while (true)
463 // sum + = array[i];
464 // i++;
465 // i = i && 0x7fff;
466 // }
467 //
468 // If the array is shorter than 0x8000 this exits through a AIOOB
469 // - Counted loop transformation is ok
470 // If the array is longer then this is an endless loop
471 // - No transformation can be done.
472
473 const TypeInt* incr_t = gvn->type(orig_incr)->is_int();
474 if (limit_t->_hi > incr_t->_hi) {
475 // if the limit can have a higher value than the increment (before the phi)
476 return false;
477 }
478 }
479
480 // Get merge point
481 Node *xphi = incr->in(1);
482 Node *stride = incr->in(2);
483 if (!stride->is_Con()) { // Oops, swap these
484 if (!xphi->is_Con()) // Is the other guy a constant?
485 return false; // Nope, unknown stride, bail out
486 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
487 xphi = stride;
488 stride = tmp;
489 }
490 if (xphi->Opcode() == Op_CastII) {
491 xphi = xphi->in(1);
492 }
493 // Stride must be constant
494 int stride_con = stride->get_int();
495 if (stride_con == 0)
496 return false; // missed some peephole opt
497
498 if (!xphi->is_Phi())
499 return false; // Too much math on the trip counter
500 if (phi_incr != NULL && phi_incr != xphi)
501 return false;
502 PhiNode *phi = xphi->as_Phi();
503
504 // Phi must be of loop header; backedge must wrap to increment
505 if (phi->region() != x)
506 return false;
507 if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
508 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
509 return false;
510 }
511 Node *init_trip = phi->in(LoopNode::EntryControl);
512
513 // If iv trunc type is smaller than int, check for possible wrap.
514 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
515 assert(trunc1 != NULL, "must have found some truncation");
516
517 // Get a better type for the phi (filtered thru if's)
518 const TypeInt* phi_ft = filtered_type(phi);
519
520 // Can iv take on a value that will wrap?
521 //
522 // Ensure iv's limit is not within "stride" of the wrap value.
523 //
524 // Example for "short" type
525 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
526 // If the stride is +10, then the last value of the induction
527 // variable before the increment (phi_ft->_hi) must be
528 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
529 // ensure no truncation occurs after the increment.
530
531 if (stride_con > 0) {
532 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
533 iv_trunc_t->_lo > phi_ft->_lo) {
534 return false; // truncation may occur
535 }
536 } else if (stride_con < 0) {
537 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
538 iv_trunc_t->_hi < phi_ft->_hi) {
539 return false; // truncation may occur
540 }
541 }
542 // No possibility of wrap so truncation can be discarded
543 // Promote iv type to Int
544 } else {
545 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
546 }
547
548 // If the condition is inverted and we will be rolling
549 // through MININT to MAXINT, then bail out.
550 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
551 // Odd stride
552 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
553 // Count down loop rolls through MAXINT
554 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
555 // Count up loop rolls through MININT
556 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
557 return false; // Bail out
558 }
559
560 const TypeInt* init_t = gvn->type(init_trip)->is_int();
561
562 if (stride_con > 0) {
563 jlong init_p = (jlong)init_t->_lo + stride_con;
564 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
565 return false; // cyclic loop or this loop trips only once
566 } else {
567 jlong init_p = (jlong)init_t->_hi + stride_con;
568 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
569 return false; // cyclic loop or this loop trips only once
570 }
571
572 if (phi_incr != NULL && bt != BoolTest::ne) {
573 // check if there is a possiblity of IV overflowing after the first increment
574 if (stride_con > 0) {
575 if (init_t->_hi > max_jint - stride_con) {
576 return false;
577 }
578 } else {
579 if (init_t->_lo < min_jint - stride_con) {
580 return false;
581 }
582 }
583 }
584
585 // =================================================
586 // ---- SUCCESS! Found A Trip-Counted Loop! -----
587 //
588 assert(x->Opcode() == Op_Loop, "regular loops only");
589 C->print_method(PHASE_BEFORE_CLOOPS, 3);
590
591 Node *hook = new Node(6);
592
593 // ===================================================
594 // Generate loop limit check to avoid integer overflow
595 // in cases like next (cyclic loops):
596 //
597 // for (i=0; i <= max_jint; i++) {}
598 // for (i=0; i < max_jint; i+=2) {}
599 //
600 //
601 // Limit check predicate depends on the loop test:
602 //
603 // for(;i != limit; i++) --> limit <= (max_jint)
604 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
605 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
606 //
607
608 // Check if limit is excluded to do more precise int overflow check.
609 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
610 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
611
612 // If compare points directly to the phi we need to adjust
613 // the compare so that it points to the incr. Limit have
614 // to be adjusted to keep trip count the same and the
615 // adjusted limit should be checked for int overflow.
616 if (phi_incr != NULL) {
617 stride_m += stride_con;
618 }
619
620 if (limit->is_Con()) {
621 int limit_con = limit->get_int();
622 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
623 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
624 // Bailout: it could be integer overflow.
625 return false;
626 }
627 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
628 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
629 // Limit's type may satisfy the condition, for example,
630 // when it is an array length.
631 } else {
632 // Generate loop's limit check.
633 // Loop limit check predicate should be near the loop.
634 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
635 if (!limit_check_proj) {
636 // The limit check predicate is not generated if this method trapped here before.
637 #ifdef ASSERT
638 if (TraceLoopLimitCheck) {
639 tty->print("missing loop limit check:");
640 loop->dump_head();
641 x->dump(1);
642 }
643 #endif
644 return false;
645 }
646
647 IfNode* check_iff = limit_check_proj->in(0)->as_If();
648
649 if (!is_dominator(get_ctrl(limit), check_iff->in(0))) {
650 return false;
651 }
652
653 Node* cmp_limit;
654 Node* bol;
655
656 if (stride_con > 0) {
657 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
658 bol = new BoolNode(cmp_limit, BoolTest::le);
659 } else {
660 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
661 bol = new BoolNode(cmp_limit, BoolTest::ge);
662 }
663
664 insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
665 }
666
667 // Now we need to canonicalize loop condition.
668 if (bt == BoolTest::ne) {
669 assert(stride_con == 1 || stride_con == -1, "simple increment only");
670 if (stride_con > 0 && init_t->_hi < limit_t->_lo) {
671 // 'ne' can be replaced with 'lt' only when init < limit.
672 bt = BoolTest::lt;
673 } else if (stride_con < 0 && init_t->_lo > limit_t->_hi) {
674 // 'ne' can be replaced with 'gt' only when init > limit.
675 bt = BoolTest::gt;
676 } else {
677 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
678 if (!limit_check_proj) {
679 // The limit check predicate is not generated if this method trapped here before.
680 #ifdef ASSERT
681 if (TraceLoopLimitCheck) {
682 tty->print("missing loop limit check:");
683 loop->dump_head();
684 x->dump(1);
685 }
686 #endif
687 return false;
688 }
689 IfNode* check_iff = limit_check_proj->in(0)->as_If();
690
691 if (!is_dominator(get_ctrl(limit), check_iff->in(0)) ||
692 !is_dominator(get_ctrl(init_trip), check_iff->in(0))) {
693 return false;
694 }
695
696 Node* cmp_limit;
697 Node* bol;
698
699 if (stride_con > 0) {
700 cmp_limit = new CmpINode(init_trip, limit);
701 bol = new BoolNode(cmp_limit, BoolTest::lt);
702 } else {
703 cmp_limit = new CmpINode(init_trip, limit);
704 bol = new BoolNode(cmp_limit, BoolTest::gt);
705 }
706
707 insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
708
709 if (stride_con > 0) {
710 // 'ne' can be replaced with 'lt' only when init < limit.
711 bt = BoolTest::lt;
712 } else if (stride_con < 0) {
713 // 'ne' can be replaced with 'gt' only when init > limit.
714 bt = BoolTest::gt;
715 }
716 }
717 }
718
719 if (phi_incr != NULL) {
720 // If compare points directly to the phi we need to adjust
721 // the compare so that it points to the incr. Limit have
722 // to be adjusted to keep trip count the same and we
723 // should avoid int overflow.
724 //
725 // i = init; do {} while(i++ < limit);
726 // is converted to
727 // i = init; do {} while(++i < limit+1);
728 //
729 limit = gvn->transform(new AddINode(limit, stride));
730 }
731
732 if (incl_limit) {
733 // The limit check guaranties that 'limit <= (max_jint - stride)' so
734 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
735 //
736 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
737 limit = gvn->transform(new AddINode(limit, one));
738 if (bt == BoolTest::le)
739 bt = BoolTest::lt;
740 else if (bt == BoolTest::ge)
741 bt = BoolTest::gt;
742 else
743 ShouldNotReachHere();
744 }
745 set_subtree_ctrl( limit );
746
747 if (LoopStripMiningIter == 0) {
748 // Check for SafePoint on backedge and remove
749 Node *sfpt = x->in(LoopNode::LoopBackControl);
750 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
751 lazy_replace( sfpt, iftrue );
752 if (loop->_safepts != NULL) {
753 loop->_safepts->yank(sfpt);
754 }
755 loop->_tail = iftrue;
756 }
757 }
758
759 // Build a canonical trip test.
760 // Clone code, as old values may be in use.
761 incr = incr->clone();
762 incr->set_req(1,phi);
763 incr->set_req(2,stride);
764 incr = _igvn.register_new_node_with_optimizer(incr);
765 set_early_ctrl( incr );
766 _igvn.rehash_node_delayed(phi);
767 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
768
769 // If phi type is more restrictive than Int, raise to
770 // Int to prevent (almost) infinite recursion in igvn
771 // which can only handle integer types for constants or minint..maxint.
772 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
773 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
774 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
775 nphi = _igvn.register_new_node_with_optimizer(nphi);
776 set_ctrl(nphi, get_ctrl(phi));
777 _igvn.replace_node(phi, nphi);
778 phi = nphi->as_Phi();
779 }
780 cmp = cmp->clone();
781 cmp->set_req(1,incr);
782 cmp->set_req(2,limit);
783 cmp = _igvn.register_new_node_with_optimizer(cmp);
784 set_ctrl(cmp, iff->in(0));
785
786 test = test->clone()->as_Bool();
787 (*(BoolTest*)&test->_test)._test = bt;
788 test->set_req(1,cmp);
789 _igvn.register_new_node_with_optimizer(test);
790 set_ctrl(test, iff->in(0));
791
792 // Replace the old IfNode with a new LoopEndNode
793 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
794 IfNode *le = lex->as_If();
795 uint dd = dom_depth(iff);
796 set_idom(le, le->in(0), dd); // Update dominance for loop exit
797 set_loop(le, loop);
798
799 // Get the loop-exit control
800 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
801
802 // Need to swap loop-exit and loop-back control?
803 if (iftrue_op == Op_IfFalse) {
804 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
805 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
806
807 loop->_tail = back_control = ift2;
808 set_loop(ift2, loop);
809 set_loop(iff2, get_loop(iffalse));
810
811 // Lazy update of 'get_ctrl' mechanism.
812 lazy_replace(iffalse, iff2);
813 lazy_replace(iftrue, ift2);
814
815 // Swap names
816 iffalse = iff2;
817 iftrue = ift2;
818 } else {
819 _igvn.rehash_node_delayed(iffalse);
820 _igvn.rehash_node_delayed(iftrue);
821 iffalse->set_req_X( 0, le, &_igvn );
822 iftrue ->set_req_X( 0, le, &_igvn );
823 }
824
825 set_idom(iftrue, le, dd+1);
826 set_idom(iffalse, le, dd+1);
827 assert(iff->outcnt() == 0, "should be dead now");
828 lazy_replace( iff, le ); // fix 'get_ctrl'
829
830 Node *sfpt2 = le->in(0);
831
832 Node* entry_control = init_control;
833 bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL &&
834 sfpt2->Opcode() == Op_SafePoint && !loop->_has_call;
835 IdealLoopTree* outer_ilt = NULL;
836 if (strip_mine_loop) {
837 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
838 cl_prob, le->_fcnt, entry_control,
839 iffalse);
840 }
841
842 // Now setup a new CountedLoopNode to replace the existing LoopNode
843 CountedLoopNode *l = new CountedLoopNode(entry_control, back_control);
844 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
845 // The following assert is approximately true, and defines the intention
846 // of can_be_counted_loop. It fails, however, because phase->type
847 // is not yet initialized for this loop and its parts.
848 //assert(l->can_be_counted_loop(this), "sanity");
849 _igvn.register_new_node_with_optimizer(l);
850 set_loop(l, loop);
851 loop->_head = l;
852 // Fix all data nodes placed at the old loop head.
853 // Uses the lazy-update mechanism of 'get_ctrl'.
854 lazy_replace( x, l );
855 set_idom(l, entry_control, dom_depth(entry_control) + 1);
856
857 if (LoopStripMiningIter == 0 || strip_mine_loop) {
858 // Check for immediately preceding SafePoint and remove
859 if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) {
860 if (strip_mine_loop) {
861 Node* outer_le = outer_ilt->_tail->in(0);
862 Node* sfpt = sfpt2->clone();
863 sfpt->set_req(0, iffalse);
864 outer_le->set_req(0, sfpt);
865 register_control(sfpt, outer_ilt, iffalse);
866 set_idom(outer_le, sfpt, dom_depth(sfpt));
867 }
868 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
869 if (loop->_safepts != NULL) {
870 loop->_safepts->yank(sfpt2);
871 }
872 }
873 }
874
875 // Free up intermediate goo
876 _igvn.remove_dead_node(hook);
877
878 #ifdef ASSERT
879 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
880 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
881 #endif
882 #ifndef PRODUCT
883 if (TraceLoopOpts) {
884 tty->print("Counted ");
885 loop->dump_head();
886 }
887 #endif
888
889 C->print_method(PHASE_AFTER_CLOOPS, 3);
890
891 // Capture bounds of the loop in the induction variable Phi before
892 // subsequent transformation (iteration splitting) obscures the
893 // bounds
894 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
895
896 if (strip_mine_loop) {
897 l->mark_strip_mined();
898 l->verify_strip_mined(1);
899 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
900 loop = outer_ilt;
901 }
902
903 return true;
904 }
905
906 //----------------------exact_limit-------------------------------------------
907 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
908 assert(loop->_head->is_CountedLoop(), "");
909 CountedLoopNode *cl = loop->_head->as_CountedLoop();
910 assert(cl->is_valid_counted_loop(), "");
911
912 if (ABS(cl->stride_con()) == 1 ||
913 cl->limit()->Opcode() == Op_LoopLimit) {
914 // Old code has exact limit (it could be incorrect in case of int overflow).
915 // Loop limit is exact with stride == 1. And loop may already have exact limit.
916 return cl->limit();
917 }
918 Node *limit = NULL;
919 #ifdef ASSERT
920 BoolTest::mask bt = cl->loopexit()->test_trip();
921 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
922 #endif
923 if (cl->has_exact_trip_count()) {
924 // Simple case: loop has constant boundaries.
925 // Use jlongs to avoid integer overflow.
926 int stride_con = cl->stride_con();
927 jlong init_con = cl->init_trip()->get_int();
928 jlong limit_con = cl->limit()->get_int();
929 julong trip_cnt = cl->trip_count();
930 jlong final_con = init_con + trip_cnt*stride_con;
931 int final_int = (int)final_con;
932 // The final value should be in integer range since the loop
933 // is counted and the limit was checked for overflow.
934 assert(final_con == (jlong)final_int, "final value should be integer");
935 limit = _igvn.intcon(final_int);
936 } else {
937 // Create new LoopLimit node to get exact limit (final iv value).
938 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
939 register_new_node(limit, cl->in(LoopNode::EntryControl));
940 }
941 assert(limit != NULL, "sanity");
942 return limit;
943 }
944
945 //------------------------------Ideal------------------------------------------
946 // Return a node which is more "ideal" than the current node.
947 // Attempt to convert into a counted-loop.
948 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
949 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
950 phase->C->set_major_progress();
951 }
952 return RegionNode::Ideal(phase, can_reshape);
953 }
954
955 void LoopNode::verify_strip_mined(int expect_skeleton) const {
956 #ifdef ASSERT
957 const OuterStripMinedLoopNode* outer = NULL;
958 const CountedLoopNode* inner = NULL;
959 if (is_strip_mined()) {
960 assert(is_CountedLoop(), "no Loop should be marked strip mined");
961 inner = as_CountedLoop();
962 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
963 } else if (is_OuterStripMinedLoop()) {
964 outer = this->as_OuterStripMinedLoop();
965 inner = outer->unique_ctrl_out()->as_CountedLoop();
966 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
967 }
968 if (inner != NULL || outer != NULL) {
969 assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
970 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
971 Node* outer_le = outer_tail->in(0);
972 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
973 Node* sfpt = outer_le->in(0);
974 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
975 Node* inner_out = sfpt->in(0);
976 if (inner_out->outcnt() != 1) {
977 ResourceMark rm;
978 Unique_Node_List wq;
979
980 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
981 Node* u = inner_out->fast_out(i);
982 if (u == sfpt) {
983 continue;
984 }
985 wq.clear();
986 wq.push(u);
987 bool found_sfpt = false;
988 for (uint next = 0; next < wq.size() && !found_sfpt; next++) {
989 Node *n = wq.at(next);
990 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) {
991 Node* u = n->fast_out(i);
992 if (u == sfpt) {
993 found_sfpt = true;
994 }
995 if (!u->is_CFG()) {
996 wq.push(u);
997 }
998 }
999 }
1000 assert(found_sfpt, "no node in loop that's not input to safepoint");
1001 }
1002 }
1003 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
1004 if (has_skeleton) {
1005 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
1006 assert(cle == inner->loopexit_or_null(), "mismatch");
1007 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
1008 assert(outer->outcnt() == 2, "only phis");
1009 } else {
1010 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
1011 uint phis = 0;
1012 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
1013 Node* u = inner->fast_out(i);
1014 if (u->is_Phi()) {
1015 phis++;
1016 }
1017 }
1018 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
1019 Node* u = outer->fast_out(i);
1020 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
1021 }
1022 Node* c = inner_out;
1023 uint stores = 0;
1024 for (;;) {
1025 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
1026 Node* u = c->fast_out(i);
1027 if (u->is_Store()) {
1028 stores++;
1029 }
1030 }
1031 if (c->in(0)->is_CountedLoopEnd()) {
1032 break;
1033 }
1034 #if INCLUDE_SHENANDOAHGC
1035 assert(UseShenandoahGC, "only for shenandoah barriers");
1036 assert(c->is_Region() && c->req() == 3, "region that ends barrier");
1037 uint j = 1;
1038 uint req = c->req();
1039 for (; j < req; j++) {
1040 Node* in = c->in(j);
1041 if (in->is_IfProj() && ShenandoahWriteBarrierNode::is_heap_stable_test(in->in(0))) {
1042 c = in->in(0)->in(0);
1043 break;
1044 }
1045 }
1046 assert(j < req, "should have found heap stable test");
1047 #endif
1048 }
1049 assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
1050 }
1051 assert(sfpt->outcnt() == 1, "no data node");
1052 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
1053 }
1054 #endif
1055 }
1056
1057 //=============================================================================
1058 //------------------------------Ideal------------------------------------------
1059 // Return a node which is more "ideal" than the current node.
1060 // Attempt to convert into a counted-loop.
1061 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1062 return RegionNode::Ideal(phase, can_reshape);
1063 }
1064
1065 //------------------------------dump_spec--------------------------------------
1066 // Dump special per-node info
1067 #ifndef PRODUCT
1068 void CountedLoopNode::dump_spec(outputStream *st) const {
1069 LoopNode::dump_spec(st);
1070 if (stride_is_con()) {
1071 st->print("stride: %d ",stride_con());
1072 }
1073 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
1074 if (is_main_loop()) st->print("main of N%d", _idx);
1075 if (is_post_loop()) st->print("post of N%d", _main_idx);
1076 if (is_strip_mined()) st->print(" strip mined");
1077 }
1078 #endif
1079
1080 //=============================================================================
1081 int CountedLoopEndNode::stride_con() const {
1082 return stride()->bottom_type()->is_int()->get_con();
1083 }
1084
1085 //=============================================================================
1086 //------------------------------Value-----------------------------------------
1087 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
1088 const Type* init_t = phase->type(in(Init));
1089 const Type* limit_t = phase->type(in(Limit));
1090 const Type* stride_t = phase->type(in(Stride));
1091 // Either input is TOP ==> the result is TOP
1092 if (init_t == Type::TOP) return Type::TOP;
1093 if (limit_t == Type::TOP) return Type::TOP;
1094 if (stride_t == Type::TOP) return Type::TOP;
1095
1096 int stride_con = stride_t->is_int()->get_con();
1097 if (stride_con == 1)
1098 return NULL; // Identity
1099
1100 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
1101 // Use jlongs to avoid integer overflow.
1102 jlong init_con = init_t->is_int()->get_con();
1103 jlong limit_con = limit_t->is_int()->get_con();
1104 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
1105 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
1106 jlong final_con = init_con + stride_con*trip_count;
1107 int final_int = (int)final_con;
1108 // The final value should be in integer range since the loop
1109 // is counted and the limit was checked for overflow.
1110 assert(final_con == (jlong)final_int, "final value should be integer");
1111 return TypeInt::make(final_int);
1112 }
1113
1114 return bottom_type(); // TypeInt::INT
1115 }
1116
1117 //------------------------------Ideal------------------------------------------
1118 // Return a node which is more "ideal" than the current node.
1119 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1120 if (phase->type(in(Init)) == Type::TOP ||
1121 phase->type(in(Limit)) == Type::TOP ||
1122 phase->type(in(Stride)) == Type::TOP)
1123 return NULL; // Dead
1124
1125 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1126 if (stride_con == 1)
1127 return NULL; // Identity
1128
1129 if (in(Init)->is_Con() && in(Limit)->is_Con())
1130 return NULL; // Value
1131
1132 // Delay following optimizations until all loop optimizations
1133 // done to keep Ideal graph simple.
1134 if (!can_reshape || phase->C->major_progress())
1135 return NULL;
1136
1137 const TypeInt* init_t = phase->type(in(Init) )->is_int();
1138 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
1139 int stride_p;
1140 jlong lim, ini;
1141 julong max;
1142 if (stride_con > 0) {
1143 stride_p = stride_con;
1144 lim = limit_t->_hi;
1145 ini = init_t->_lo;
1146 max = (julong)max_jint;
1147 } else {
1148 stride_p = -stride_con;
1149 lim = init_t->_hi;
1150 ini = limit_t->_lo;
1151 max = (julong)min_jint;
1152 }
1153 julong range = lim - ini + stride_p;
1154 if (range <= max) {
1155 // Convert to integer expression if it is not overflow.
1156 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
1157 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
1158 Node *bias = phase->transform(new AddINode(range, stride_m));
1159 Node *trip = phase->transform(new DivINode(0, bias, in(Stride)));
1160 Node *span = phase->transform(new MulINode(trip, in(Stride)));
1161 return new AddINode(span, in(Init)); // exact limit
1162 }
1163
1164 if (is_power_of_2(stride_p) || // divisor is 2^n
1165 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
1166 // Convert to long expression to avoid integer overflow
1167 // and let igvn optimizer convert this division.
1168 //
1169 Node* init = phase->transform( new ConvI2LNode(in(Init)));
1170 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
1171 Node* stride = phase->longcon(stride_con);
1172 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
1173
1174 Node *range = phase->transform(new SubLNode(limit, init));
1175 Node *bias = phase->transform(new AddLNode(range, stride_m));
1176 Node *span;
1177 if (stride_con > 0 && is_power_of_2(stride_p)) {
1178 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
1179 // and avoid generating rounding for division. Zero trip guard should
1180 // guarantee that init < limit but sometimes the guard is missing and
1181 // we can get situation when init > limit. Note, for the empty loop
1182 // optimization zero trip guard is generated explicitly which leaves
1183 // only RCE predicate where exact limit is used and the predicate
1184 // will simply fail forcing recompilation.
1185 Node* neg_stride = phase->longcon(-stride_con);
1186 span = phase->transform(new AndLNode(bias, neg_stride));
1187 } else {
1188 Node *trip = phase->transform(new DivLNode(0, bias, stride));
1189 span = phase->transform(new MulLNode(trip, stride));
1190 }
1191 // Convert back to int
1192 Node *span_int = phase->transform(new ConvL2INode(span));
1193 return new AddINode(span_int, in(Init)); // exact limit
1194 }
1195
1196 return NULL; // No progress
1197 }
1198
1199 //------------------------------Identity---------------------------------------
1200 // If stride == 1 return limit node.
1201 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1202 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1203 if (stride_con == 1 || stride_con == -1)
1204 return in(Limit);
1205 return this;
1206 }
1207
1208 //=============================================================================
1209 //----------------------match_incr_with_optional_truncation--------------------
1210 // Match increment with optional truncation:
1211 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1212 // Return NULL for failure. Success returns the increment node.
1213 Node* CountedLoopNode::match_incr_with_optional_truncation(
1214 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1215 // Quick cutouts:
1216 if (expr == NULL || expr->req() != 3) return NULL;
1217
1218 Node *t1 = NULL;
1219 Node *t2 = NULL;
1220 const TypeInt* trunc_t = TypeInt::INT;
1221 Node* n1 = expr;
1222 int n1op = n1->Opcode();
1223
1224 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1225 if (n1op == Op_AndI &&
1226 n1->in(2)->is_Con() &&
1227 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1228 // %%% This check should match any mask of 2**K-1.
1229 t1 = n1;
1230 n1 = t1->in(1);
1231 n1op = n1->Opcode();
1232 trunc_t = TypeInt::CHAR;
1233 } else if (n1op == Op_RShiftI &&
1234 n1->in(1) != NULL &&
1235 n1->in(1)->Opcode() == Op_LShiftI &&
1236 n1->in(2) == n1->in(1)->in(2) &&
1237 n1->in(2)->is_Con()) {
1238 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1239 // %%% This check should match any shift in [1..31].
1240 if (shift == 16 || shift == 8) {
1241 t1 = n1;
1242 t2 = t1->in(1);
1243 n1 = t2->in(1);
1244 n1op = n1->Opcode();
1245 if (shift == 16) {
1246 trunc_t = TypeInt::SHORT;
1247 } else if (shift == 8) {
1248 trunc_t = TypeInt::BYTE;
1249 }
1250 }
1251 }
1252
1253 // If (maybe after stripping) it is an AddI, we won:
1254 if (n1op == Op_AddI) {
1255 *trunc1 = t1;
1256 *trunc2 = t2;
1257 *trunc_type = trunc_t;
1258 return n1;
1259 }
1260
1261 // failed
1262 return NULL;
1263 }
1264
1265 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
1266 if (is_strip_mined()) {
1267 verify_strip_mined(expect_skeleton);
1268 return in(EntryControl)->as_Loop();
1269 }
1270 return this;
1271 }
1272
1273 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
1274 assert(is_strip_mined(), "not a strip mined loop");
1275 Node* c = in(EntryControl);
1276 if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) {
1277 return NULL;
1278 }
1279 return c->as_OuterStripMinedLoop();
1280 }
1281
1282 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
1283 Node* c = in(LoopBackControl);
1284 if (c == NULL || c->is_top()) {
1285 return NULL;
1286 }
1287 return c->as_IfTrue();
1288 }
1289
1290 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
1291 LoopNode* l = outer_loop();
1292 if (l == NULL) {
1293 return NULL;
1294 }
1295 return l->outer_loop_tail();
1296 }
1297
1298 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
1299 IfTrueNode* proj = outer_loop_tail();
1300 if (proj == NULL) {
1301 return NULL;
1302 }
1303 Node* c = proj->in(0);
1304 if (c == NULL || c->is_top() || c->outcnt() != 2) {
1305 return NULL;
1306 }
1307 return c->as_OuterStripMinedLoopEnd();
1308 }
1309
1310 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
1311 LoopNode* l = outer_loop();
1312 if (l == NULL) {
1313 return NULL;
1314 }
1315 return l->outer_loop_end();
1316 }
1317
1318 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
1319 IfNode* le = outer_loop_end();
1320 if (le == NULL) {
1321 return NULL;
1322 }
1323 Node* c = le->proj_out_or_null(false);
1324 if (c == NULL) {
1325 return NULL;
1326 }
1327 return c->as_IfFalse();
1328 }
1329
1330 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
1331 LoopNode* l = outer_loop();
1332 if (l == NULL) {
1333 return NULL;
1334 }
1335 return l->outer_loop_exit();
1336 }
1337
1338 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
1339 IfNode* le = outer_loop_end();
1340 if (le == NULL) {
1341 return NULL;
1342 }
1343 Node* c = le->in(0);
1344 if (c == NULL || c->is_top()) {
1345 return NULL;
1346 }
1347 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
1348 return c->as_SafePoint();
1349 }
1350
1351 SafePointNode* CountedLoopNode::outer_safepoint() const {
1352 LoopNode* l = outer_loop();
1353 if (l == NULL) {
1354 return NULL;
1355 }
1356 return l->outer_safepoint();
1357 }
1358
1359 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
1360 while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
1361 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
1362 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
1363 ctrl = ctrl->in(0)->in(0);
1364 }
1365
1366 return ctrl;
1367 }
1368
1369 Node* CountedLoopNode::skip_predicates() {
1370 if (is_main_loop()) {
1371 Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
1372
1373 return skip_predicates_from_entry(ctrl);
1374 }
1375 return in(LoopNode::EntryControl);
1376 }
1377
1378 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
1379 // Look for the outer & inner strip mined loop, reduce number of
1380 // iterations of the inner loop, set exit condition of outer loop,
1381 // construct required phi nodes for outer loop.
1382 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
1383 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
1384 Node* inner_iv_phi = inner_cl->phi();
1385 if (inner_iv_phi == NULL) {
1386 IfNode* outer_le = outer_loop_end();
1387 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1388 igvn->replace_node(outer_le, iff);
1389 inner_cl->clear_strip_mined();
1390 return;
1391 }
1392 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
1393
1394 int stride = inner_cl->stride_con();
1395 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride);
1396 int scaled_iters = (int)scaled_iters_long;
1397 int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride);
1398 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
1399 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
1400 assert(iter_estimate > 0, "broken");
1401 if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) {
1402 // Remove outer loop and safepoint (too few iterations)
1403 Node* outer_sfpt = outer_safepoint();
1404 Node* outer_out = outer_loop_exit();
1405 igvn->replace_node(outer_out, outer_sfpt->in(0));
1406 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
1407 inner_cl->clear_strip_mined();
1408 return;
1409 }
1410 if (iter_estimate <= scaled_iters_long) {
1411 // We would only go through one iteration of
1412 // the outer loop: drop the outer loop but
1413 // keep the safepoint so we don't run for
1414 // too long without a safepoint
1415 IfNode* outer_le = outer_loop_end();
1416 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1417 igvn->replace_node(outer_le, iff);
1418 inner_cl->clear_strip_mined();
1419 return;
1420 }
1421
1422 Node* cle_tail = inner_cle->proj_out(true);
1423 ResourceMark rm;
1424 Node_List old_new;
1425 if (cle_tail->outcnt() > 1) {
1426 // Look for nodes on backedge of inner loop and clone them
1427 Unique_Node_List backedge_nodes;
1428 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
1429 Node* u = cle_tail->fast_out(i);
1430 if (u != inner_cl) {
1431 assert(!u->is_CFG(), "control flow on the backedge?");
1432 backedge_nodes.push(u);
1433 }
1434 }
1435 uint last = igvn->C->unique();
1436 for (uint next = 0; next < backedge_nodes.size(); next++) {
1437 Node* n = backedge_nodes.at(next);
1438 old_new.map(n->_idx, n->clone());
1439 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1440 Node* u = n->fast_out(i);
1441 assert(!u->is_CFG(), "broken");
1442 if (u->_idx >= last) {
1443 continue;
1444 }
1445 if (!u->is_Phi()) {
1446 backedge_nodes.push(u);
1447 } else {
1448 assert(u->in(0) == inner_cl, "strange phi on the backedge");
1449 }
1450 }
1451 }
1452 // Put the clones on the outer loop backedge
1453 Node* le_tail = outer_loop_tail();
1454 for (uint next = 0; next < backedge_nodes.size(); next++) {
1455 Node *n = old_new[backedge_nodes.at(next)->_idx];
1456 for (uint i = 1; i < n->req(); i++) {
1457 if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
1458 n->set_req(i, old_new[n->in(i)->_idx]);
1459 }
1460 }
1461 if (n->in(0) != NULL && n->in(0) == cle_tail) {
1462 n->set_req(0, le_tail);
1463 }
1464 igvn->register_new_node_with_optimizer(n);
1465 }
1466 }
1467
1468 Node* iv_phi = NULL;
1469 // Make a clone of each phi in the inner loop
1470 // for the outer loop
1471 for (uint i = 0; i < inner_cl->outcnt(); i++) {
1472 Node* u = inner_cl->raw_out(i);
1473 if (u->is_Phi()) {
1474 assert(u->in(0) == inner_cl, "inconsistent");
1475 Node* phi = u->clone();
1476 phi->set_req(0, this);
1477 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
1478 if (be != NULL) {
1479 phi->set_req(LoopNode::LoopBackControl, be);
1480 }
1481 phi = igvn->transform(phi);
1482 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
1483 if (u == inner_iv_phi) {
1484 iv_phi = phi;
1485 }
1486 }
1487 }
1488 Node* cle_out = inner_cle->proj_out(false);
1489 if (cle_out->outcnt() > 1) {
1490 // Look for chains of stores that were sunk
1491 // out of the inner loop and are in the outer loop
1492 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
1493 Node* u = cle_out->fast_out(i);
1494 if (u->is_Store()) {
1495 Node* first = u;
1496 for(;;) {
1497 Node* next = first->in(MemNode::Memory);
1498 if (!next->is_Store() || next->in(0) != cle_out) {
1499 break;
1500 }
1501 first = next;
1502 }
1503 Node* last = u;
1504 for(;;) {
1505 Node* next = NULL;
1506 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
1507 Node* uu = last->fast_out(j);
1508 if (uu->is_Store() && uu->in(0) == cle_out) {
1509 assert(next == NULL, "only one in the outer loop");
1510 next = uu;
1511 }
1512 }
1513 if (next == NULL) {
1514 break;
1515 }
1516 last = next;
1517 }
1518 Node* phi = NULL;
1519 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1520 Node* uu = fast_out(j);
1521 if (uu->is_Phi()) {
1522 Node* be = uu->in(LoopNode::LoopBackControl);
1523 if (be->is_Store() && old_new[be->_idx] != NULL) {
1524 assert(false, "store on the backedge + sunk stores: unsupported");
1525 // drop outer loop
1526 IfNode* outer_le = outer_loop_end();
1527 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1528 igvn->replace_node(outer_le, iff);
1529 inner_cl->clear_strip_mined();
1530 return;
1531 }
1532 if (be == last || be == first->in(MemNode::Memory)) {
1533 assert(phi == NULL, "only one phi");
1534 phi = uu;
1535 }
1536 }
1537 }
1538 #ifdef ASSERT
1539 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1540 Node* uu = fast_out(j);
1541 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) {
1542 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
1543 assert(phi == uu, "what's that phi?");
1544 } else if (uu->adr_type() == TypePtr::BOTTOM) {
1545 Node* n = uu->in(LoopNode::LoopBackControl);
1546 uint limit = igvn->C->live_nodes();
1547 uint i = 0;
1548 while (n != uu) {
1549 i++;
1550 assert(i < limit, "infinite loop");
1551 if (n->is_Proj()) {
1552 n = n->in(0);
1553 } else if (n->is_SafePoint() || n->is_MemBar()) {
1554 n = n->in(TypeFunc::Memory);
1555 } else if (n->is_Phi()) {
1556 n = n->in(1);
1557 } else if (n->is_MergeMem()) {
1558 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
1559 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
1560 n = n->in(MemNode::Memory);
1561 } else {
1562 n->dump();
1563 ShouldNotReachHere();
1564 }
1565 }
1566 }
1567 }
1568 }
1569 #endif
1570 if (phi == NULL) {
1571 // If the an entire chains was sunk, the
1572 // inner loop has no phi for that memory
1573 // slice, create one for the outer loop
1574 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY,
1575 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
1576 phi->set_req(LoopNode::LoopBackControl, last);
1577 phi = igvn->transform(phi);
1578 igvn->replace_input_of(first, MemNode::Memory, phi);
1579 } else {
1580 // Or fix the outer loop fix to include
1581 // that chain of stores.
1582 Node* be = phi->in(LoopNode::LoopBackControl);
1583 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported");
1584 if (be == first->in(MemNode::Memory)) {
1585 if (be == phi->in(LoopNode::LoopBackControl)) {
1586 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
1587 } else {
1588 igvn->replace_input_of(be, MemNode::Memory, last);
1589 }
1590 } else {
1591 #ifdef ASSERT
1592 if (be == phi->in(LoopNode::LoopBackControl)) {
1593 assert(phi->in(LoopNode::LoopBackControl) == last, "");
1594 } else {
1595 assert(be->in(MemNode::Memory) == last, "");
1596 }
1597 #endif
1598 }
1599 }
1600 }
1601 }
1602 }
1603
1604 if (iv_phi != NULL) {
1605 // Now adjust the inner loop's exit condition
1606 Node* limit = inner_cl->limit();
1607 Node* sub = NULL;
1608 if (stride > 0) {
1609 sub = igvn->transform(new SubINode(limit, iv_phi));
1610 } else {
1611 sub = igvn->transform(new SubINode(iv_phi, limit));
1612 }
1613 Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters)));
1614 Node* new_limit = NULL;
1615 if (stride > 0) {
1616 new_limit = igvn->transform(new AddINode(min, iv_phi));
1617 } else {
1618 new_limit = igvn->transform(new SubINode(iv_phi, min));
1619 }
1620 Node* cmp = inner_cle->cmp_node()->clone();
1621 igvn->replace_input_of(cmp, 2, new_limit);
1622 Node* bol = inner_cle->in(CountedLoopEndNode::TestValue)->clone();
1623 cmp->set_req(2, limit);
1624 bol->set_req(1, igvn->transform(cmp));
1625 igvn->replace_input_of(outer_loop_end(), 1, igvn->transform(bol));
1626 } else {
1627 assert(false, "should be able to adjust outer loop");
1628 IfNode* outer_le = outer_loop_end();
1629 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1630 igvn->replace_node(outer_le, iff);
1631 inner_cl->clear_strip_mined();
1632 }
1633 }
1634
1635 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
1636 if (!in(0)) return Type::TOP;
1637 if (phase->type(in(0)) == Type::TOP)
1638 return Type::TOP;
1639
1640 return TypeTuple::IFBOTH;
1641 }
1642
1643 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1644 if (remove_dead_region(phase, can_reshape)) return this;
1645
1646 return NULL;
1647 }
1648
1649 //------------------------------filtered_type--------------------------------
1650 // Return a type based on condition control flow
1651 // A successful return will be a type that is restricted due
1652 // to a series of dominating if-tests, such as:
1653 // if (i < 10) {
1654 // if (i > 0) {
1655 // here: "i" type is [1..10)
1656 // }
1657 // }
1658 // or a control flow merge
1659 // if (i < 10) {
1660 // do {
1661 // phi( , ) -- at top of loop type is [min_int..10)
1662 // i = ?
1663 // } while ( i < 10)
1664 //
1665 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1666 assert(n && n->bottom_type()->is_int(), "must be int");
1667 const TypeInt* filtered_t = NULL;
1668 if (!n->is_Phi()) {
1669 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1670 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1671
1672 } else {
1673 Node* phi = n->as_Phi();
1674 Node* region = phi->in(0);
1675 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1676 if (region && region != C->top()) {
1677 for (uint i = 1; i < phi->req(); i++) {
1678 Node* val = phi->in(i);
1679 Node* use_c = region->in(i);
1680 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1681 if (val_t != NULL) {
1682 if (filtered_t == NULL) {
1683 filtered_t = val_t;
1684 } else {
1685 filtered_t = filtered_t->meet(val_t)->is_int();
1686 }
1687 }
1688 }
1689 }
1690 }
1691 const TypeInt* n_t = _igvn.type(n)->is_int();
1692 if (filtered_t != NULL) {
1693 n_t = n_t->join(filtered_t)->is_int();
1694 }
1695 return n_t;
1696 }
1697
1698
1699 //------------------------------filtered_type_from_dominators--------------------------------
1700 // Return a possibly more restrictive type for val based on condition control flow of dominators
1701 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1702 if (val->is_Con()) {
1703 return val->bottom_type()->is_int();
1704 }
1705 uint if_limit = 10; // Max number of dominating if's visited
1706 const TypeInt* rtn_t = NULL;
1707
1708 if (use_ctrl && use_ctrl != C->top()) {
1709 Node* val_ctrl = get_ctrl(val);
1710 uint val_dom_depth = dom_depth(val_ctrl);
1711 Node* pred = use_ctrl;
1712 uint if_cnt = 0;
1713 while (if_cnt < if_limit) {
1714 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1715 if_cnt++;
1716 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1717 if (if_t != NULL) {
1718 if (rtn_t == NULL) {
1719 rtn_t = if_t;
1720 } else {
1721 rtn_t = rtn_t->join(if_t)->is_int();
1722 }
1723 }
1724 }
1725 pred = idom(pred);
1726 if (pred == NULL || pred == C->top()) {
1727 break;
1728 }
1729 // Stop if going beyond definition block of val
1730 if (dom_depth(pred) < val_dom_depth) {
1731 break;
1732 }
1733 }
1734 }
1735 return rtn_t;
1736 }
1737
1738
1739 //------------------------------dump_spec--------------------------------------
1740 // Dump special per-node info
1741 #ifndef PRODUCT
1742 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1743 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1744 BoolTest bt( test_trip()); // Added this for g++.
1745
1746 st->print("[");
1747 bt.dump_on(st);
1748 st->print("]");
1749 }
1750 st->print(" ");
1751 IfNode::dump_spec(st);
1752 }
1753 #endif
1754
1755 //=============================================================================
1756 //------------------------------is_member--------------------------------------
1757 // Is 'l' a member of 'this'?
1758 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1759 while( l->_nest > _nest ) l = l->_parent;
1760 return l == this;
1761 }
1762
1763 //------------------------------set_nest---------------------------------------
1764 // Set loop tree nesting depth. Accumulate _has_call bits.
1765 int IdealLoopTree::set_nest( uint depth ) {
1766 _nest = depth;
1767 int bits = _has_call;
1768 if( _child ) bits |= _child->set_nest(depth+1);
1769 if( bits ) _has_call = 1;
1770 if( _next ) bits |= _next ->set_nest(depth );
1771 return bits;
1772 }
1773
1774 //------------------------------split_fall_in----------------------------------
1775 // Split out multiple fall-in edges from the loop header. Move them to a
1776 // private RegionNode before the loop. This becomes the loop landing pad.
1777 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1778 PhaseIterGVN &igvn = phase->_igvn;
1779 uint i;
1780
1781 // Make a new RegionNode to be the landing pad.
1782 Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1783 phase->set_loop(landing_pad,_parent);
1784 // Gather all the fall-in control paths into the landing pad
1785 uint icnt = fall_in_cnt;
1786 uint oreq = _head->req();
1787 for( i = oreq-1; i>0; i-- )
1788 if( !phase->is_member( this, _head->in(i) ) )
1789 landing_pad->set_req(icnt--,_head->in(i));
1790
1791 // Peel off PhiNode edges as well
1792 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1793 Node *oj = _head->fast_out(j);
1794 if( oj->is_Phi() ) {
1795 PhiNode* old_phi = oj->as_Phi();
1796 assert( old_phi->region() == _head, "" );
1797 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1798 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1799 uint icnt = fall_in_cnt;
1800 for( i = oreq-1; i>0; i-- ) {
1801 if( !phase->is_member( this, _head->in(i) ) ) {
1802 p->init_req(icnt--, old_phi->in(i));
1803 // Go ahead and clean out old edges from old phi
1804 old_phi->del_req(i);
1805 }
1806 }
1807 // Search for CSE's here, because ZKM.jar does a lot of
1808 // loop hackery and we need to be a little incremental
1809 // with the CSE to avoid O(N^2) node blow-up.
1810 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1811 if( p2 ) { // Found CSE
1812 p->destruct(); // Recover useless new node
1813 p = p2; // Use old node
1814 } else {
1815 igvn.register_new_node_with_optimizer(p, old_phi);
1816 }
1817 // Make old Phi refer to new Phi.
1818 old_phi->add_req(p);
1819 // Check for the special case of making the old phi useless and
1820 // disappear it. In JavaGrande I have a case where this useless
1821 // Phi is the loop limit and prevents recognizing a CountedLoop
1822 // which in turn prevents removing an empty loop.
1823 Node *id_old_phi = igvn.apply_identity(old_phi);
1824 if( id_old_phi != old_phi ) { // Found a simple identity?
1825 // Note that I cannot call 'replace_node' here, because
1826 // that will yank the edge from old_phi to the Region and
1827 // I'm mid-iteration over the Region's uses.
1828 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1829 Node* use = old_phi->last_out(i);
1830 igvn.rehash_node_delayed(use);
1831 uint uses_found = 0;
1832 for (uint j = 0; j < use->len(); j++) {
1833 if (use->in(j) == old_phi) {
1834 if (j < use->req()) use->set_req (j, id_old_phi);
1835 else use->set_prec(j, id_old_phi);
1836 uses_found++;
1837 }
1838 }
1839 i -= uses_found; // we deleted 1 or more copies of this edge
1840 }
1841 }
1842 igvn._worklist.push(old_phi);
1843 }
1844 }
1845 // Finally clean out the fall-in edges from the RegionNode
1846 for( i = oreq-1; i>0; i-- ) {
1847 if( !phase->is_member( this, _head->in(i) ) ) {
1848 _head->del_req(i);
1849 }
1850 }
1851 igvn.rehash_node_delayed(_head);
1852 // Transform landing pad
1853 igvn.register_new_node_with_optimizer(landing_pad, _head);
1854 // Insert landing pad into the header
1855 _head->add_req(landing_pad);
1856 }
1857
1858 //------------------------------split_outer_loop-------------------------------
1859 // Split out the outermost loop from this shared header.
1860 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1861 PhaseIterGVN &igvn = phase->_igvn;
1862
1863 // Find index of outermost loop; it should also be my tail.
1864 uint outer_idx = 1;
1865 while( _head->in(outer_idx) != _tail ) outer_idx++;
1866
1867 // Make a LoopNode for the outermost loop.
1868 Node *ctl = _head->in(LoopNode::EntryControl);
1869 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1870 outer = igvn.register_new_node_with_optimizer(outer, _head);
1871 phase->set_created_loop_node();
1872
1873 // Outermost loop falls into '_head' loop
1874 _head->set_req(LoopNode::EntryControl, outer);
1875 _head->del_req(outer_idx);
1876 // Split all the Phis up between '_head' loop and 'outer' loop.
1877 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1878 Node *out = _head->fast_out(j);
1879 if( out->is_Phi() ) {
1880 PhiNode *old_phi = out->as_Phi();
1881 assert( old_phi->region() == _head, "" );
1882 Node *phi = PhiNode::make_blank(outer, old_phi);
1883 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1884 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1885 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1886 // Make old Phi point to new Phi on the fall-in path
1887 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1888 old_phi->del_req(outer_idx);
1889 }
1890 }
1891
1892 // Use the new loop head instead of the old shared one
1893 _head = outer;
1894 phase->set_loop(_head, this);
1895 }
1896
1897 //------------------------------fix_parent-------------------------------------
1898 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1899 loop->_parent = parent;
1900 if( loop->_child ) fix_parent( loop->_child, loop );
1901 if( loop->_next ) fix_parent( loop->_next , parent );
1902 }
1903
1904 //------------------------------estimate_path_freq-----------------------------
1905 static float estimate_path_freq( Node *n ) {
1906 // Try to extract some path frequency info
1907 IfNode *iff;
1908 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1909 uint nop = n->Opcode();
1910 if( nop == Op_SafePoint ) { // Skip any safepoint
1911 n = n->in(0);
1912 continue;
1913 }
1914 if( nop == Op_CatchProj ) { // Get count from a prior call
1915 // Assume call does not always throw exceptions: means the call-site
1916 // count is also the frequency of the fall-through path.
1917 assert( n->is_CatchProj(), "" );
1918 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1919 return 0.0f; // Assume call exception path is rare
1920 Node *call = n->in(0)->in(0)->in(0);
1921 assert( call->is_Call(), "expect a call here" );
1922 const JVMState *jvms = ((CallNode*)call)->jvms();
1923 ciMethodData* methodData = jvms->method()->method_data();
1924 if (!methodData->is_mature()) return 0.0f; // No call-site data
1925 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1926 if ((data == NULL) || !data->is_CounterData()) {
1927 // no call profile available, try call's control input
1928 n = n->in(0);
1929 continue;
1930 }
1931 return data->as_CounterData()->count()/FreqCountInvocations;
1932 }
1933 // See if there's a gating IF test
1934 Node *n_c = n->in(0);
1935 if( !n_c->is_If() ) break; // No estimate available
1936 iff = n_c->as_If();
1937 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1938 // Compute how much count comes on this path
1939 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1940 // Have no count info. Skip dull uncommon-trap like branches.
1941 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1942 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1943 break;
1944 // Skip through never-taken branch; look for a real loop exit.
1945 n = iff->in(0);
1946 }
1947 return 0.0f; // No estimate available
1948 }
1949
1950 //------------------------------merge_many_backedges---------------------------
1951 // Merge all the backedges from the shared header into a private Region.
1952 // Feed that region as the one backedge to this loop.
1953 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1954 uint i;
1955
1956 // Scan for the top 2 hottest backedges
1957 float hotcnt = 0.0f;
1958 float warmcnt = 0.0f;
1959 uint hot_idx = 0;
1960 // Loop starts at 2 because slot 1 is the fall-in path
1961 for( i = 2; i < _head->req(); i++ ) {
1962 float cnt = estimate_path_freq(_head->in(i));
1963 if( cnt > hotcnt ) { // Grab hottest path
1964 warmcnt = hotcnt;
1965 hotcnt = cnt;
1966 hot_idx = i;
1967 } else if( cnt > warmcnt ) { // And 2nd hottest path
1968 warmcnt = cnt;
1969 }
1970 }
1971
1972 // See if the hottest backedge is worthy of being an inner loop
1973 // by being much hotter than the next hottest backedge.
1974 if( hotcnt <= 0.0001 ||
1975 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1976
1977 // Peel out the backedges into a private merge point; peel
1978 // them all except optionally hot_idx.
1979 PhaseIterGVN &igvn = phase->_igvn;
1980
1981 Node *hot_tail = NULL;
1982 // Make a Region for the merge point
1983 Node *r = new RegionNode(1);
1984 for( i = 2; i < _head->req(); i++ ) {
1985 if( i != hot_idx )
1986 r->add_req( _head->in(i) );
1987 else hot_tail = _head->in(i);
1988 }
1989 igvn.register_new_node_with_optimizer(r, _head);
1990 // Plug region into end of loop _head, followed by hot_tail
1991 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1992 igvn.replace_input_of(_head, 2, r);
1993 if( hot_idx ) _head->add_req(hot_tail);
1994
1995 // Split all the Phis up between '_head' loop and the Region 'r'
1996 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1997 Node *out = _head->fast_out(j);
1998 if( out->is_Phi() ) {
1999 PhiNode* n = out->as_Phi();
2000 igvn.hash_delete(n); // Delete from hash before hacking edges
2001 Node *hot_phi = NULL;
2002 Node *phi = new PhiNode(r, n->type(), n->adr_type());
2003 // Check all inputs for the ones to peel out
2004 uint j = 1;
2005 for( uint i = 2; i < n->req(); i++ ) {
2006 if( i != hot_idx )
2007 phi->set_req( j++, n->in(i) );
2008 else hot_phi = n->in(i);
2009 }
2010 // Register the phi but do not transform until whole place transforms
2011 igvn.register_new_node_with_optimizer(phi, n);
2012 // Add the merge phi to the old Phi
2013 while( n->req() > 3 ) n->del_req( n->req()-1 );
2014 igvn.replace_input_of(n, 2, phi);
2015 if( hot_idx ) n->add_req(hot_phi);
2016 }
2017 }
2018
2019
2020 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
2021 // of self loop tree. Turn self into a loop headed by _head and with
2022 // tail being the new merge point.
2023 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
2024 phase->set_loop(_tail,ilt); // Adjust tail
2025 _tail = r; // Self's tail is new merge point
2026 phase->set_loop(r,this);
2027 ilt->_child = _child; // New guy has my children
2028 _child = ilt; // Self has new guy as only child
2029 ilt->_parent = this; // new guy has self for parent
2030 ilt->_nest = _nest; // Same nesting depth (for now)
2031
2032 // Starting with 'ilt', look for child loop trees using the same shared
2033 // header. Flatten these out; they will no longer be loops in the end.
2034 IdealLoopTree **pilt = &_child;
2035 while( ilt ) {
2036 if( ilt->_head == _head ) {
2037 uint i;
2038 for( i = 2; i < _head->req(); i++ )
2039 if( _head->in(i) == ilt->_tail )
2040 break; // Still a loop
2041 if( i == _head->req() ) { // No longer a loop
2042 // Flatten ilt. Hang ilt's "_next" list from the end of
2043 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
2044 IdealLoopTree **cp = &ilt->_child;
2045 while( *cp ) cp = &(*cp)->_next; // Find end of child list
2046 *cp = ilt->_next; // Hang next list at end of child list
2047 *pilt = ilt->_child; // Move child up to replace ilt
2048 ilt->_head = NULL; // Flag as a loop UNIONED into parent
2049 ilt = ilt->_child; // Repeat using new ilt
2050 continue; // do not advance over ilt->_child
2051 }
2052 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
2053 phase->set_loop(_head,ilt);
2054 }
2055 pilt = &ilt->_child; // Advance to next
2056 ilt = *pilt;
2057 }
2058
2059 if( _child ) fix_parent( _child, this );
2060 }
2061
2062 //------------------------------beautify_loops---------------------------------
2063 // Split shared headers and insert loop landing pads.
2064 // Insert a LoopNode to replace the RegionNode.
2065 // Return TRUE if loop tree is structurally changed.
2066 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
2067 bool result = false;
2068 // Cache parts in locals for easy
2069 PhaseIterGVN &igvn = phase->_igvn;
2070
2071 igvn.hash_delete(_head); // Yank from hash before hacking edges
2072
2073 // Check for multiple fall-in paths. Peel off a landing pad if need be.
2074 int fall_in_cnt = 0;
2075 for( uint i = 1; i < _head->req(); i++ )
2076 if( !phase->is_member( this, _head->in(i) ) )
2077 fall_in_cnt++;
2078 assert( fall_in_cnt, "at least 1 fall-in path" );
2079 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
2080 split_fall_in( phase, fall_in_cnt );
2081
2082 // Swap inputs to the _head and all Phis to move the fall-in edge to
2083 // the left.
2084 fall_in_cnt = 1;
2085 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
2086 fall_in_cnt++;
2087 if( fall_in_cnt > 1 ) {
2088 // Since I am just swapping inputs I do not need to update def-use info
2089 Node *tmp = _head->in(1);
2090 igvn.rehash_node_delayed(_head);
2091 _head->set_req( 1, _head->in(fall_in_cnt) );
2092 _head->set_req( fall_in_cnt, tmp );
2093 // Swap also all Phis
2094 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
2095 Node* phi = _head->fast_out(i);
2096 if( phi->is_Phi() ) {
2097 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
2098 tmp = phi->in(1);
2099 phi->set_req( 1, phi->in(fall_in_cnt) );
2100 phi->set_req( fall_in_cnt, tmp );
2101 }
2102 }
2103 }
2104 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
2105 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
2106
2107 // If I am a shared header (multiple backedges), peel off the many
2108 // backedges into a private merge point and use the merge point as
2109 // the one true backedge.
2110 if( _head->req() > 3 ) {
2111 // Merge the many backedges into a single backedge but leave
2112 // the hottest backedge as separate edge for the following peel.
2113 merge_many_backedges( phase );
2114 result = true;
2115 }
2116
2117 // If I have one hot backedge, peel off myself loop.
2118 // I better be the outermost loop.
2119 if (_head->req() > 3 && !_irreducible) {
2120 split_outer_loop( phase );
2121 result = true;
2122
2123 } else if (!_head->is_Loop() && !_irreducible) {
2124 // Make a new LoopNode to replace the old loop head
2125 Node *l = new LoopNode( _head->in(1), _head->in(2) );
2126 l = igvn.register_new_node_with_optimizer(l, _head);
2127 phase->set_created_loop_node();
2128 // Go ahead and replace _head
2129 phase->_igvn.replace_node( _head, l );
2130 _head = l;
2131 phase->set_loop(_head, this);
2132 }
2133
2134 // Now recursively beautify nested loops
2135 if( _child ) result |= _child->beautify_loops( phase );
2136 if( _next ) result |= _next ->beautify_loops( phase );
2137 return result;
2138 }
2139
2140 //------------------------------allpaths_check_safepts----------------------------
2141 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
2142 // encountered. Helper for check_safepts.
2143 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
2144 assert(stack.size() == 0, "empty stack");
2145 stack.push(_tail);
2146 visited.Clear();
2147 visited.set(_tail->_idx);
2148 while (stack.size() > 0) {
2149 Node* n = stack.pop();
2150 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2151 // Terminate this path
2152 } else if (n->Opcode() == Op_SafePoint) {
2153 if (_phase->get_loop(n) != this) {
2154 if (_required_safept == NULL) _required_safept = new Node_List();
2155 _required_safept->push(n); // save the one closest to the tail
2156 }
2157 // Terminate this path
2158 } else {
2159 uint start = n->is_Region() ? 1 : 0;
2160 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
2161 for (uint i = start; i < end; i++) {
2162 Node* in = n->in(i);
2163 assert(in->is_CFG(), "must be");
2164 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
2165 stack.push(in);
2166 }
2167 }
2168 }
2169 }
2170 }
2171
2172 //------------------------------check_safepts----------------------------
2173 // Given dominators, try to find loops with calls that must always be
2174 // executed (call dominates loop tail). These loops do not need non-call
2175 // safepoints (ncsfpt).
2176 //
2177 // A complication is that a safepoint in a inner loop may be needed
2178 // by an outer loop. In the following, the inner loop sees it has a
2179 // call (block 3) on every path from the head (block 2) to the
2180 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
2181 // in block 2, _but_ this leaves the outer loop without a safepoint.
2182 //
2183 // entry 0
2184 // |
2185 // v
2186 // outer 1,2 +->1
2187 // | |
2188 // | v
2189 // | 2<---+ ncsfpt in 2
2190 // |_/|\ |
2191 // | v |
2192 // inner 2,3 / 3 | call in 3
2193 // / | |
2194 // v +--+
2195 // exit 4
2196 //
2197 //
2198 // This method creates a list (_required_safept) of ncsfpt nodes that must
2199 // be protected is created for each loop. When a ncsfpt maybe deleted, it
2200 // is first looked for in the lists for the outer loops of the current loop.
2201 //
2202 // The insights into the problem:
2203 // A) counted loops are okay
2204 // B) innermost loops are okay (only an inner loop can delete
2205 // a ncsfpt needed by an outer loop)
2206 // C) a loop is immune from an inner loop deleting a safepoint
2207 // if the loop has a call on the idom-path
2208 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
2209 // idom-path that is not in a nested loop
2210 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
2211 // loop needs to be prevented from deletion by an inner loop
2212 //
2213 // There are two analyses:
2214 // 1) The first, and cheaper one, scans the loop body from
2215 // tail to head following the idom (immediate dominator)
2216 // chain, looking for the cases (C,D,E) above.
2217 // Since inner loops are scanned before outer loops, there is summary
2218 // information about inner loops. Inner loops can be skipped over
2219 // when the tail of an inner loop is encountered.
2220 //
2221 // 2) The second, invoked if the first fails to find a call or ncsfpt on
2222 // the idom path (which is rare), scans all predecessor control paths
2223 // from the tail to the head, terminating a path when a call or sfpt
2224 // is encountered, to find the ncsfpt's that are closest to the tail.
2225 //
2226 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
2227 // Bottom up traversal
2228 IdealLoopTree* ch = _child;
2229 if (_child) _child->check_safepts(visited, stack);
2230 if (_next) _next ->check_safepts(visited, stack);
2231
2232 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
2233 bool has_call = false; // call on dom-path
2234 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
2235 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
2236 // Scan the dom-path nodes from tail to head
2237 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
2238 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2239 has_call = true;
2240 _has_sfpt = 1; // Then no need for a safept!
2241 break;
2242 } else if (n->Opcode() == Op_SafePoint) {
2243 if (_phase->get_loop(n) == this) {
2244 has_local_ncsfpt = true;
2245 break;
2246 }
2247 if (nonlocal_ncsfpt == NULL) {
2248 nonlocal_ncsfpt = n; // save the one closest to the tail
2249 }
2250 } else {
2251 IdealLoopTree* nlpt = _phase->get_loop(n);
2252 if (this != nlpt) {
2253 // If at an inner loop tail, see if the inner loop has already
2254 // recorded seeing a call on the dom-path (and stop.) If not,
2255 // jump to the head of the inner loop.
2256 assert(is_member(nlpt), "nested loop");
2257 Node* tail = nlpt->_tail;
2258 if (tail->in(0)->is_If()) tail = tail->in(0);
2259 if (n == tail) {
2260 // If inner loop has call on dom-path, so does outer loop
2261 if (nlpt->_has_sfpt) {
2262 has_call = true;
2263 _has_sfpt = 1;
2264 break;
2265 }
2266 // Skip to head of inner loop
2267 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
2268 n = nlpt->_head;
2269 }
2270 }
2271 }
2272 }
2273 // Record safept's that this loop needs preserved when an
2274 // inner loop attempts to delete it's safepoints.
2275 if (_child != NULL && !has_call && !has_local_ncsfpt) {
2276 if (nonlocal_ncsfpt != NULL) {
2277 if (_required_safept == NULL) _required_safept = new Node_List();
2278 _required_safept->push(nonlocal_ncsfpt);
2279 } else {
2280 // Failed to find a suitable safept on the dom-path. Now use
2281 // an all paths walk from tail to head, looking for safepoints to preserve.
2282 allpaths_check_safepts(visited, stack);
2283 }
2284 }
2285 }
2286 }
2287
2288 //---------------------------is_deleteable_safept----------------------------
2289 // Is safept not required by an outer loop?
2290 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
2291 assert(sfpt->Opcode() == Op_SafePoint, "");
2292 IdealLoopTree* lp = get_loop(sfpt)->_parent;
2293 while (lp != NULL) {
2294 Node_List* sfpts = lp->_required_safept;
2295 if (sfpts != NULL) {
2296 for (uint i = 0; i < sfpts->size(); i++) {
2297 if (sfpt == sfpts->at(i))
2298 return false;
2299 }
2300 }
2301 lp = lp->_parent;
2302 }
2303 return true;
2304 }
2305
2306 //---------------------------replace_parallel_iv-------------------------------
2307 // Replace parallel induction variable (parallel to trip counter)
2308 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
2309 assert(loop->_head->is_CountedLoop(), "");
2310 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2311 if (!cl->is_valid_counted_loop())
2312 return; // skip malformed counted loop
2313 Node *incr = cl->incr();
2314 if (incr == NULL)
2315 return; // Dead loop?
2316 Node *init = cl->init_trip();
2317 Node *phi = cl->phi();
2318 int stride_con = cl->stride_con();
2319
2320 // Visit all children, looking for Phis
2321 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
2322 Node *out = cl->out(i);
2323 // Look for other phis (secondary IVs). Skip dead ones
2324 if (!out->is_Phi() || out == phi || !has_node(out))
2325 continue;
2326 PhiNode* phi2 = out->as_Phi();
2327 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
2328 // Look for induction variables of the form: X += constant
2329 if (phi2->region() != loop->_head ||
2330 incr2->req() != 3 ||
2331 incr2->in(1) != phi2 ||
2332 incr2 == incr ||
2333 incr2->Opcode() != Op_AddI ||
2334 !incr2->in(2)->is_Con())
2335 continue;
2336
2337 // Check for parallel induction variable (parallel to trip counter)
2338 // via an affine function. In particular, count-down loops with
2339 // count-up array indices are common. We only RCE references off
2340 // the trip-counter, so we need to convert all these to trip-counter
2341 // expressions.
2342 Node *init2 = phi2->in( LoopNode::EntryControl );
2343 int stride_con2 = incr2->in(2)->get_int();
2344
2345 // The ratio of the two strides cannot be represented as an int
2346 // if stride_con2 is min_int and stride_con is -1.
2347 if (stride_con2 == min_jint && stride_con == -1) {
2348 continue;
2349 }
2350
2351 // The general case here gets a little tricky. We want to find the
2352 // GCD of all possible parallel IV's and make a new IV using this
2353 // GCD for the loop. Then all possible IVs are simple multiples of
2354 // the GCD. In practice, this will cover very few extra loops.
2355 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
2356 // where +/-1 is the common case, but other integer multiples are
2357 // also easy to handle.
2358 int ratio_con = stride_con2/stride_con;
2359
2360 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
2361 #ifndef PRODUCT
2362 if (TraceLoopOpts) {
2363 tty->print("Parallel IV: %d ", phi2->_idx);
2364 loop->dump_head();
2365 }
2366 #endif
2367 // Convert to using the trip counter. The parallel induction
2368 // variable differs from the trip counter by a loop-invariant
2369 // amount, the difference between their respective initial values.
2370 // It is scaled by the 'ratio_con'.
2371 Node* ratio = _igvn.intcon(ratio_con);
2372 set_ctrl(ratio, C->root());
2373 Node* ratio_init = new MulINode(init, ratio);
2374 _igvn.register_new_node_with_optimizer(ratio_init, init);
2375 set_early_ctrl(ratio_init);
2376 Node* diff = new SubINode(init2, ratio_init);
2377 _igvn.register_new_node_with_optimizer(diff, init2);
2378 set_early_ctrl(diff);
2379 Node* ratio_idx = new MulINode(phi, ratio);
2380 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
2381 set_ctrl(ratio_idx, cl);
2382 Node* add = new AddINode(ratio_idx, diff);
2383 _igvn.register_new_node_with_optimizer(add);
2384 set_ctrl(add, cl);
2385 _igvn.replace_node( phi2, add );
2386 // Sometimes an induction variable is unused
2387 if (add->outcnt() == 0) {
2388 _igvn.remove_dead_node(add);
2389 }
2390 --i; // deleted this phi; rescan starting with next position
2391 continue;
2392 }
2393 }
2394 }
2395
2396 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
2397 Node* keep = NULL;
2398 if (keep_one) {
2399 // Look for a safepoint on the idom-path.
2400 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
2401 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
2402 keep = i;
2403 break; // Found one
2404 }
2405 }
2406 }
2407
2408 // Don't remove any safepoints if it is requested to keep a single safepoint and
2409 // no safepoint was found on idom-path. It is not safe to remove any safepoint
2410 // in this case since there's no safepoint dominating all paths in the loop body.
2411 bool prune = !keep_one || keep != NULL;
2412
2413 // Delete other safepoints in this loop.
2414 Node_List* sfpts = _safepts;
2415 if (prune && sfpts != NULL) {
2416 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
2417 for (uint i = 0; i < sfpts->size(); i++) {
2418 Node* n = sfpts->at(i);
2419 assert(phase->get_loop(n) == this, "");
2420 if (n != keep && phase->is_deleteable_safept(n)) {
2421 phase->lazy_replace(n, n->in(TypeFunc::Control));
2422 }
2423 }
2424 }
2425 }
2426
2427 //------------------------------counted_loop-----------------------------------
2428 // Convert to counted loops where possible
2429 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
2430
2431 // For grins, set the inner-loop flag here
2432 if (!_child) {
2433 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
2434 }
2435
2436 IdealLoopTree* loop = this;
2437 if (_head->is_CountedLoop() ||
2438 phase->is_counted_loop(_head, loop)) {
2439
2440 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) {
2441 // Indicate we do not need a safepoint here
2442 _has_sfpt = 1;
2443 }
2444
2445 // Remove safepoints
2446 bool keep_one_sfpt = !(_has_call || _has_sfpt);
2447 remove_safepoints(phase, keep_one_sfpt);
2448
2449 // Look for induction variables
2450 phase->replace_parallel_iv(this);
2451
2452 } else if (_parent != NULL && !_irreducible) {
2453 // Not a counted loop. Keep one safepoint.
2454 bool keep_one_sfpt = true;
2455 remove_safepoints(phase, keep_one_sfpt);
2456 }
2457
2458 // Recursively
2459 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
2460 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops");
2461 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
2462 if (loop->_next) loop->_next ->counted_loop(phase);
2463 }
2464
2465 #ifndef PRODUCT
2466 //------------------------------dump_head--------------------------------------
2467 // Dump 1 liner for loop header info
2468 void IdealLoopTree::dump_head( ) const {
2469 for (uint i=0; i<_nest; i++)
2470 tty->print(" ");
2471 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
2472 if (_irreducible) tty->print(" IRREDUCIBLE");
2473 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
2474 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
2475 if (predicate != NULL ) {
2476 tty->print(" limit_check");
2477 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2478 }
2479 if (UseLoopPredicate) {
2480 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2481 if (entry != NULL) {
2482 tty->print(" predicated");
2483 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2484 }
2485 }
2486 if (UseProfiledLoopPredicate) {
2487 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
2488 if (entry != NULL) {
2489 tty->print(" profile_predicated");
2490 }
2491 }
2492 if (_head->is_CountedLoop()) {
2493 CountedLoopNode *cl = _head->as_CountedLoop();
2494 tty->print(" counted");
2495
2496 Node* init_n = cl->init_trip();
2497 if (init_n != NULL && init_n->is_Con())
2498 tty->print(" [%d,", cl->init_trip()->get_int());
2499 else
2500 tty->print(" [int,");
2501 Node* limit_n = cl->limit();
2502 if (limit_n != NULL && limit_n->is_Con())
2503 tty->print("%d),", cl->limit()->get_int());
2504 else
2505 tty->print("int),");
2506 int stride_con = cl->stride_con();
2507 if (stride_con > 0) tty->print("+");
2508 tty->print("%d", stride_con);
2509
2510 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
2511
2512 if (cl->is_pre_loop ()) tty->print(" pre" );
2513 if (cl->is_main_loop()) tty->print(" main");
2514 if (cl->is_post_loop()) tty->print(" post");
2515 if (cl->is_vectorized_loop()) tty->print(" vector");
2516 if (cl->range_checks_present()) tty->print(" rc ");
2517 if (cl->is_multiversioned()) tty->print(" multi ");
2518 }
2519 if (_has_call) tty->print(" has_call");
2520 if (_has_sfpt) tty->print(" has_sfpt");
2521 if (_rce_candidate) tty->print(" rce");
2522 if (_safepts != NULL && _safepts->size() > 0) {
2523 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
2524 }
2525 if (_required_safept != NULL && _required_safept->size() > 0) {
2526 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
2527 }
2528 if (Verbose) {
2529 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
2530 }
2531 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
2532 tty->print(" strip_mined");
2533 }
2534 tty->cr();
2535 }
2536
2537 //------------------------------dump-------------------------------------------
2538 // Dump loops by loop tree
2539 void IdealLoopTree::dump( ) const {
2540 dump_head();
2541 if (_child) _child->dump();
2542 if (_next) _next ->dump();
2543 }
2544
2545 #endif
2546
2547 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
2548 if (loop == root) {
2549 if (loop->_child != NULL) {
2550 log->begin_head("loop_tree");
2551 log->end_head();
2552 if( loop->_child ) log_loop_tree(root, loop->_child, log);
2553 log->tail("loop_tree");
2554 assert(loop->_next == NULL, "what?");
2555 }
2556 } else {
2557 Node* head = loop->_head;
2558 log->begin_head("loop");
2559 log->print(" idx='%d' ", head->_idx);
2560 if (loop->_irreducible) log->print("irreducible='1' ");
2561 if (head->is_Loop()) {
2562 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
2563 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
2564 }
2565 if (head->is_CountedLoop()) {
2566 CountedLoopNode* cl = head->as_CountedLoop();
2567 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
2568 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
2569 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
2570 }
2571 log->end_head();
2572 if( loop->_child ) log_loop_tree(root, loop->_child, log);
2573 log->tail("loop");
2574 if( loop->_next ) log_loop_tree(root, loop->_next, log);
2575 }
2576 }
2577
2578 //---------------------collect_potentially_useful_predicates-----------------------
2579 // Helper function to collect potentially useful predicates to prevent them from
2580 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2581 void PhaseIdealLoop::collect_potentially_useful_predicates(
2582 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2583 if (loop->_child) { // child
2584 collect_potentially_useful_predicates(loop->_child, useful_predicates);
2585 }
2586
2587 // self (only loops that we can apply loop predication may use their predicates)
2588 if (loop->_head->is_Loop() &&
2589 !loop->_irreducible &&
2590 !loop->tail()->is_top()) {
2591 LoopNode* lpn = loop->_head->as_Loop();
2592 Node* entry = lpn->in(LoopNode::EntryControl);
2593 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2594 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2595 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2596 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2597 entry = skip_loop_predicates(entry);
2598 }
2599 predicate_proj = find_predicate(entry); // Predicate
2600 if (predicate_proj != NULL ) {
2601 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2602 entry = skip_loop_predicates(entry);
2603 }
2604 if (UseProfiledLoopPredicate) {
2605 predicate_proj = find_predicate(entry); // Predicate
2606 if (predicate_proj != NULL ) {
2607 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2608 }
2609 }
2610 }
2611
2612 if (loop->_next) { // sibling
2613 collect_potentially_useful_predicates(loop->_next, useful_predicates);
2614 }
2615 }
2616
2617 //------------------------eliminate_useless_predicates-----------------------------
2618 // Eliminate all inserted predicates if they could not be used by loop predication.
2619 // Note: it will also eliminates loop limits check predicate since it also uses
2620 // Opaque1 node (see Parse::add_predicate()).
2621 void PhaseIdealLoop::eliminate_useless_predicates() {
2622 if (C->predicate_count() == 0)
2623 return; // no predicate left
2624
2625 Unique_Node_List useful_predicates; // to store useful predicates
2626 if (C->has_loops()) {
2627 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2628 }
2629
2630 for (int i = C->predicate_count(); i > 0; i--) {
2631 Node * n = C->predicate_opaque1_node(i-1);
2632 assert(n->Opcode() == Op_Opaque1, "must be");
2633 if (!useful_predicates.member(n)) { // not in the useful list
2634 _igvn.replace_node(n, n->in(1));
2635 }
2636 }
2637 }
2638
2639 //------------------------process_expensive_nodes-----------------------------
2640 // Expensive nodes have their control input set to prevent the GVN
2641 // from commoning them and as a result forcing the resulting node to
2642 // be in a more frequent path. Use CFG information here, to change the
2643 // control inputs so that some expensive nodes can be commoned while
2644 // not executed more frequently.
2645 bool PhaseIdealLoop::process_expensive_nodes() {
2646 assert(OptimizeExpensiveOps, "optimization off?");
2647
2648 // Sort nodes to bring similar nodes together
2649 C->sort_expensive_nodes();
2650
2651 bool progress = false;
2652
2653 for (int i = 0; i < C->expensive_count(); ) {
2654 Node* n = C->expensive_node(i);
2655 int start = i;
2656 // Find nodes similar to n
2657 i++;
2658 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2659 int end = i;
2660 // And compare them two by two
2661 for (int j = start; j < end; j++) {
2662 Node* n1 = C->expensive_node(j);
2663 if (is_node_unreachable(n1)) {
2664 continue;
2665 }
2666 for (int k = j+1; k < end; k++) {
2667 Node* n2 = C->expensive_node(k);
2668 if (is_node_unreachable(n2)) {
2669 continue;
2670 }
2671
2672 assert(n1 != n2, "should be pair of nodes");
2673
2674 Node* c1 = n1->in(0);
2675 Node* c2 = n2->in(0);
2676
2677 Node* parent_c1 = c1;
2678 Node* parent_c2 = c2;
2679
2680 // The call to get_early_ctrl_for_expensive() moves the
2681 // expensive nodes up but stops at loops that are in a if
2682 // branch. See whether we can exit the loop and move above the
2683 // If.
2684 if (c1->is_Loop()) {
2685 parent_c1 = c1->in(1);
2686 }
2687 if (c2->is_Loop()) {
2688 parent_c2 = c2->in(1);
2689 }
2690
2691 if (parent_c1 == parent_c2) {
2692 _igvn._worklist.push(n1);
2693 _igvn._worklist.push(n2);
2694 continue;
2695 }
2696
2697 // Look for identical expensive node up the dominator chain.
2698 if (is_dominator(c1, c2)) {
2699 c2 = c1;
2700 } else if (is_dominator(c2, c1)) {
2701 c1 = c2;
2702 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2703 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2704 // Both branches have the same expensive node so move it up
2705 // before the if.
2706 c1 = c2 = idom(parent_c1->in(0));
2707 }
2708 // Do the actual moves
2709 if (n1->in(0) != c1) {
2710 _igvn.hash_delete(n1);
2711 n1->set_req(0, c1);
2712 _igvn.hash_insert(n1);
2713 _igvn._worklist.push(n1);
2714 progress = true;
2715 }
2716 if (n2->in(0) != c2) {
2717 _igvn.hash_delete(n2);
2718 n2->set_req(0, c2);
2719 _igvn.hash_insert(n2);
2720 _igvn._worklist.push(n2);
2721 progress = true;
2722 }
2723 }
2724 }
2725 }
2726
2727 return progress;
2728 }
2729
2730
2731 //=============================================================================
2732 //----------------------------build_and_optimize-------------------------------
2733 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2734 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2735 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
2736 bool do_split_ifs = (mode == LoopOptsDefault || mode == LoopOptsLastRound);
2737 bool skip_loop_opts = (mode == LoopOptsNone);
2738 bool shenandoah_opts = (mode == LoopOptsShenandoahExpand ||
2739 mode == LoopOptsShenandoahPostExpand);
2740
2741 ResourceMark rm;
2742
2743 int old_progress = C->major_progress();
2744 uint orig_worklist_size = _igvn._worklist.size();
2745
2746 // Reset major-progress flag for the driver's heuristics
2747 C->clear_major_progress();
2748
2749 #ifndef PRODUCT
2750 // Capture for later assert
2751 uint unique = C->unique();
2752 _loop_invokes++;
2753 _loop_work += unique;
2754 #endif
2755
2756 // True if the method has at least 1 irreducible loop
2757 _has_irreducible_loops = false;
2758
2759 _created_loop_node = false;
2760
2761 Arena *a = Thread::current()->resource_area();
2762 VectorSet visited(a);
2763 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2764 _nodes.map(C->unique(), NULL);
2765 memset(_nodes.adr(), 0, wordSize * C->unique());
2766
2767 // Pre-build the top-level outermost loop tree entry
2768 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2769 // Do not need a safepoint at the top level
2770 _ltree_root->_has_sfpt = 1;
2771
2772 // Initialize Dominators.
2773 // Checked in clone_loop_predicate() during beautify_loops().
2774 _idom_size = 0;
2775 _idom = NULL;
2776 _dom_depth = NULL;
2777 _dom_stk = NULL;
2778
2779 // Empty pre-order array
2780 allocate_preorders();
2781
2782 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2783 // IdealLoopTree entries. Data nodes are NOT walked.
2784 build_loop_tree();
2785 // Check for bailout, and return
2786 if (C->failing()) {
2787 return;
2788 }
2789
2790 // No loops after all
2791 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2792
2793 // There should always be an outer loop containing the Root and Return nodes.
2794 // If not, we have a degenerate empty program. Bail out in this case.
2795 if (!has_node(C->root())) {
2796 if (!_verify_only) {
2797 C->clear_major_progress();
2798 C->record_method_not_compilable("empty program detected during loop optimization");
2799 }
2800 return;
2801 }
2802
2803 // Nothing to do, so get out
2804 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only && !shenandoah_opts;
2805 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2806 if (stop_early && !do_expensive_nodes) {
2807 _igvn.optimize(); // Cleanup NeverBranches
2808 return;
2809 }
2810
2811 // Set loop nesting depth
2812 _ltree_root->set_nest( 0 );
2813
2814 // Split shared headers and insert loop landing pads.
2815 // Do not bother doing this on the Root loop of course.
2816 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2817 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2818 if( _ltree_root->_child->beautify_loops( this ) ) {
2819 // Re-build loop tree!
2820 _ltree_root->_child = NULL;
2821 _nodes.clear();
2822 reallocate_preorders();
2823 build_loop_tree();
2824 // Check for bailout, and return
2825 if (C->failing()) {
2826 return;
2827 }
2828 // Reset loop nesting depth
2829 _ltree_root->set_nest( 0 );
2830
2831 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2832 }
2833 }
2834
2835 // Build Dominators for elision of NULL checks & loop finding.
2836 // Since nodes do not have a slot for immediate dominator, make
2837 // a persistent side array for that info indexed on node->_idx.
2838 _idom_size = C->unique();
2839 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2840 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2841 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2842 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2843
2844 Dominators();
2845
2846 if (!_verify_only) {
2847 // As a side effect, Dominators removed any unreachable CFG paths
2848 // into RegionNodes. It doesn't do this test against Root, so
2849 // we do it here.
2850 for( uint i = 1; i < C->root()->req(); i++ ) {
2851 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2852 _igvn.delete_input_of(C->root(), i);
2853 i--; // Rerun same iteration on compressed edges
2854 }
2855 }
2856
2857 // Given dominators, try to find inner loops with calls that must
2858 // always be executed (call dominates loop tail). These loops do
2859 // not need a separate safepoint.
2860 Node_List cisstack(a);
2861 _ltree_root->check_safepts(visited, cisstack);
2862 }
2863
2864 // Walk the DATA nodes and place into loops. Find earliest control
2865 // node. For CFG nodes, the _nodes array starts out and remains
2866 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2867 // _nodes array holds the earliest legal controlling CFG node.
2868
2869 // Allocate stack with enough space to avoid frequent realloc
2870 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2871 Node_Stack nstack( a, stack_size );
2872
2873 visited.Clear();
2874 Node_List worklist(a);
2875 // Don't need C->root() on worklist since
2876 // it will be processed among C->top() inputs
2877 worklist.push( C->top() );
2878 visited.set( C->top()->_idx ); // Set C->top() as visited now
2879 build_loop_early( visited, worklist, nstack );
2880
2881 // Given early legal placement, try finding counted loops. This placement
2882 // is good enough to discover most loop invariants.
2883 if (!_verify_me && !_verify_only && !shenandoah_opts) {
2884 _ltree_root->counted_loop(this);
2885 }
2886
2887 // Find latest loop placement. Find ideal loop placement.
2888 visited.Clear();
2889 init_dom_lca_tags();
2890 // Need C->root() on worklist when processing outs
2891 worklist.push( C->root() );
2892 NOT_PRODUCT( C->verify_graph_edges(); )
2893 worklist.push( C->top() );
2894 build_loop_late(visited, worklist, nstack, !shenandoah_opts);
2895
2896 if (_verify_only) {
2897 // restore major progress flag
2898 for (int i = 0; i < old_progress; i++)
2899 C->set_major_progress();
2900 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2901 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2902 return;
2903 }
2904
2905 // clear out the dead code after build_loop_late
2906 while (_deadlist.size()) {
2907 _igvn.remove_globally_dead_node(_deadlist.pop());
2908 }
2909
2910 if (stop_early) {
2911 assert(do_expensive_nodes, "why are we here?");
2912 if (process_expensive_nodes()) {
2913 // If we made some progress when processing expensive nodes then
2914 // the IGVN may modify the graph in a way that will allow us to
2915 // make some more progress: we need to try processing expensive
2916 // nodes again.
2917 C->set_major_progress();
2918 }
2919 _igvn.optimize();
2920 return;
2921 }
2922
2923 // Some parser-inserted loop predicates could never be used by loop
2924 // predication or they were moved away from loop during some optimizations.
2925 // For example, peeling. Eliminate them before next loop optimizations.
2926 eliminate_useless_predicates();
2927
2928 #ifndef PRODUCT
2929 C->verify_graph_edges();
2930 if (_verify_me) { // Nested verify pass?
2931 // Check to see if the verify mode is broken
2932 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2933 return;
2934 }
2935 if(VerifyLoopOptimizations) verify();
2936 if(TraceLoopOpts && C->has_loops()) {
2937 _ltree_root->dump();
2938 }
2939 #endif
2940
2941 #if INCLUDE_SHENANDOAHGC
2942 if (mode == LoopOptsShenandoahExpand) {
2943 assert(UseShenandoahGC, "only for shenandoah");
2944 ShenandoahWriteBarrierNode::pin_and_expand(this);
2945 } else if (mode == LoopOptsShenandoahPostExpand) {
2946 assert(UseShenandoahGC, "only for shenandoah");
2947 visited.Clear();
2948 ShenandoahWriteBarrierNode::optimize_after_expansion(visited, nstack, worklist, this);
2949 }
2950
2951 if (shenandoah_opts) {
2952 _igvn.optimize();
2953 if (C->log() != NULL) {
2954 log_loop_tree(_ltree_root, _ltree_root, C->log());
2955 }
2956 return;
2957 }
2958 #endif
2959
2960 if (skip_loop_opts) {
2961 // restore major progress flag
2962 for (int i = 0; i < old_progress; i++) {
2963 C->set_major_progress();
2964 }
2965
2966 // Cleanup any modified bits
2967 _igvn.optimize();
2968
2969 if (C->log() != NULL) {
2970 log_loop_tree(_ltree_root, _ltree_root, C->log());
2971 }
2972 return;
2973 }
2974
2975 #if INCLUDE_SHENANDOAHGC
2976 if (UseShenandoahGC) {
2977 GrowableArray<MemoryGraphFixer*> memory_graph_fixers;
2978 ShenandoahWriteBarrierNode::optimize_before_expansion(this, memory_graph_fixers, false);
2979 }
2980 #endif
2981
2982 if (ReassociateInvariants) {
2983 // Reassociate invariants and prep for split_thru_phi
2984 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2985 IdealLoopTree* lpt = iter.current();
2986 bool is_counted = lpt->is_counted();
2987 if (!is_counted || !lpt->is_inner()) continue;
2988
2989 // check for vectorized loops, any reassociation of invariants was already done
2990 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2991
2992 lpt->reassociate_invariants(this);
2993
2994 // Because RCE opportunities can be masked by split_thru_phi,
2995 // look for RCE candidates and inhibit split_thru_phi
2996 // on just their loop-phi's for this pass of loop opts
2997 if (SplitIfBlocks && do_split_ifs) {
2998 if (lpt->policy_range_check(this)) {
2999 lpt->_rce_candidate = 1; // = true
3000 }
3001 }
3002 }
3003 }
3004
3005 // Check for aggressive application of split-if and other transforms
3006 // that require basic-block info (like cloning through Phi's)
3007 if( SplitIfBlocks && do_split_ifs ) {
3008 visited.Clear();
3009 split_if_with_blocks( visited, nstack, mode == LoopOptsLastRound );
3010 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
3011 if (mode == LoopOptsLastRound) {
3012 C->set_major_progress();
3013 }
3014 }
3015
3016 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
3017 C->set_major_progress();
3018 }
3019
3020 // Perform loop predication before iteration splitting
3021 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
3022 _ltree_root->_child->loop_predication(this);
3023 }
3024
3025 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
3026 if (do_intrinsify_fill()) {
3027 C->set_major_progress();
3028 }
3029 }
3030
3031 // Perform iteration-splitting on inner loops. Split iterations to avoid
3032 // range checks or one-shot null checks.
3033
3034 // If split-if's didn't hack the graph too bad (no CFG changes)
3035 // then do loop opts.
3036 if (C->has_loops() && !C->major_progress()) {
3037 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
3038 _ltree_root->_child->iteration_split( this, worklist );
3039 // No verify after peeling! GCM has hoisted code out of the loop.
3040 // After peeling, the hoisted code could sink inside the peeled area.
3041 // The peeling code does not try to recompute the best location for
3042 // all the code before the peeled area, so the verify pass will always
3043 // complain about it.
3044 }
3045 // Do verify graph edges in any case
3046 NOT_PRODUCT( C->verify_graph_edges(); );
3047
3048 if (!do_split_ifs) {
3049 // We saw major progress in Split-If to get here. We forced a
3050 // pass with unrolling and not split-if, however more split-if's
3051 // might make progress. If the unrolling didn't make progress
3052 // then the major-progress flag got cleared and we won't try
3053 // another round of Split-If. In particular the ever-common
3054 // instance-of/check-cast pattern requires at least 2 rounds of
3055 // Split-If to clear out.
3056 C->set_major_progress();
3057 }
3058
3059 // Repeat loop optimizations if new loops were seen
3060 if (created_loop_node()) {
3061 C->set_major_progress();
3062 }
3063
3064 // Keep loop predicates and perform optimizations with them
3065 // until no more loop optimizations could be done.
3066 // After that switch predicates off and do more loop optimizations.
3067 if (!C->major_progress() && (C->predicate_count() > 0)) {
3068 C->cleanup_loop_predicates(_igvn);
3069 if (TraceLoopOpts) {
3070 tty->print_cr("PredicatesOff");
3071 }
3072 C->set_major_progress();
3073 }
3074
3075 // Convert scalar to superword operations at the end of all loop opts.
3076 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
3077 // SuperWord transform
3078 SuperWord sw(this);
3079 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3080 IdealLoopTree* lpt = iter.current();
3081 if (lpt->is_counted()) {
3082 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
3083
3084 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
3085 // Check that the rce'd post loop is encountered first, multiversion after all
3086 // major main loop optimization are concluded
3087 if (!C->major_progress()) {
3088 IdealLoopTree *lpt_next = lpt->_next;
3089 if (lpt_next && lpt_next->is_counted()) {
3090 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
3091 has_range_checks(lpt_next);
3092 if (cl->is_post_loop() && cl->range_checks_present()) {
3093 if (!cl->is_multiversioned()) {
3094 if (multi_version_post_loops(lpt, lpt_next) == false) {
3095 // Cause the rce loop to be optimized away if we fail
3096 cl->mark_is_multiversioned();
3097 cl->set_slp_max_unroll(0);
3098 poison_rce_post_loop(lpt);
3099 }
3100 }
3101 }
3102 }
3103 sw.transform_loop(lpt, true);
3104 }
3105 } else if (cl->is_main_loop()) {
3106 sw.transform_loop(lpt, true);
3107 }
3108 }
3109 }
3110 }
3111
3112 // Cleanup any modified bits
3113 _igvn.optimize();
3114
3115 // disable assert until issue with split_flow_path is resolved (6742111)
3116 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
3117 // "shouldn't introduce irreducible loops");
3118
3119 if (C->log() != NULL) {
3120 log_loop_tree(_ltree_root, _ltree_root, C->log());
3121 }
3122 }
3123
3124 #ifndef PRODUCT
3125 //------------------------------print_statistics-------------------------------
3126 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
3127 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
3128 void PhaseIdealLoop::print_statistics() {
3129 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
3130 }
3131
3132 //------------------------------verify-----------------------------------------
3133 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
3134 static int fail; // debug only, so its multi-thread dont care
3135 void PhaseIdealLoop::verify() const {
3136 int old_progress = C->major_progress();
3137 ResourceMark rm;
3138 PhaseIdealLoop loop_verify( _igvn, this );
3139 VectorSet visited(Thread::current()->resource_area());
3140
3141 fail = 0;
3142 verify_compare( C->root(), &loop_verify, visited );
3143 assert( fail == 0, "verify loops failed" );
3144 // Verify loop structure is the same
3145 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
3146 // Reset major-progress. It was cleared by creating a verify version of
3147 // PhaseIdealLoop.
3148 for( int i=0; i<old_progress; i++ )
3149 C->set_major_progress();
3150 }
3151
3152 //------------------------------verify_compare---------------------------------
3153 // Make sure me and the given PhaseIdealLoop agree on key data structures
3154 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
3155 if( !n ) return;
3156 if( visited.test_set( n->_idx ) ) return;
3157 if( !_nodes[n->_idx] ) { // Unreachable
3158 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
3159 return;
3160 }
3161
3162 uint i;
3163 for( i = 0; i < n->req(); i++ )
3164 verify_compare( n->in(i), loop_verify, visited );
3165
3166 // Check the '_nodes' block/loop structure
3167 i = n->_idx;
3168 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
3169 if( _nodes[i] != loop_verify->_nodes[i] &&
3170 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
3171 tty->print("Mismatched control setting for: ");
3172 n->dump();
3173 if( fail++ > 10 ) return;
3174 Node *c = get_ctrl_no_update(n);
3175 tty->print("We have it as: ");
3176 if( c->in(0) ) c->dump();
3177 else tty->print_cr("N%d",c->_idx);
3178 tty->print("Verify thinks: ");
3179 if( loop_verify->has_ctrl(n) )
3180 loop_verify->get_ctrl_no_update(n)->dump();
3181 else
3182 loop_verify->get_loop_idx(n)->dump();
3183 tty->cr();
3184 }
3185 } else { // We have a loop
3186 IdealLoopTree *us = get_loop_idx(n);
3187 if( loop_verify->has_ctrl(n) ) {
3188 tty->print("Mismatched loop setting for: ");
3189 n->dump();
3190 if( fail++ > 10 ) return;
3191 tty->print("We have it as: ");
3192 us->dump();
3193 tty->print("Verify thinks: ");
3194 loop_verify->get_ctrl_no_update(n)->dump();
3195 tty->cr();
3196 } else if (!C->major_progress()) {
3197 // Loop selection can be messed up if we did a major progress
3198 // operation, like split-if. Do not verify in that case.
3199 IdealLoopTree *them = loop_verify->get_loop_idx(n);
3200 if( us->_head != them->_head || us->_tail != them->_tail ) {
3201 tty->print("Unequals loops for: ");
3202 n->dump();
3203 if( fail++ > 10 ) return;
3204 tty->print("We have it as: ");
3205 us->dump();
3206 tty->print("Verify thinks: ");
3207 them->dump();
3208 tty->cr();
3209 }
3210 }
3211 }
3212
3213 // Check for immediate dominators being equal
3214 if( i >= _idom_size ) {
3215 if( !n->is_CFG() ) return;
3216 tty->print("CFG Node with no idom: ");
3217 n->dump();
3218 return;
3219 }
3220 if( !n->is_CFG() ) return;
3221 if( n == C->root() ) return; // No IDOM here
3222
3223 assert(n->_idx == i, "sanity");
3224 Node *id = idom_no_update(n);
3225 if( id != loop_verify->idom_no_update(n) ) {
3226 tty->print("Unequals idoms for: ");
3227 n->dump();
3228 if( fail++ > 10 ) return;
3229 tty->print("We have it as: ");
3230 id->dump();
3231 tty->print("Verify thinks: ");
3232 loop_verify->idom_no_update(n)->dump();
3233 tty->cr();
3234 }
3235
3236 }
3237
3238 //------------------------------verify_tree------------------------------------
3239 // Verify that tree structures match. Because the CFG can change, siblings
3240 // within the loop tree can be reordered. We attempt to deal with that by
3241 // reordering the verify's loop tree if possible.
3242 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
3243 assert( _parent == parent, "Badly formed loop tree" );
3244
3245 // Siblings not in same order? Attempt to re-order.
3246 if( _head != loop->_head ) {
3247 // Find _next pointer to update
3248 IdealLoopTree **pp = &loop->_parent->_child;
3249 while( *pp != loop )
3250 pp = &((*pp)->_next);
3251 // Find proper sibling to be next
3252 IdealLoopTree **nn = &loop->_next;
3253 while( (*nn) && (*nn)->_head != _head )
3254 nn = &((*nn)->_next);
3255
3256 // Check for no match.
3257 if( !(*nn) ) {
3258 // Annoyingly, irreducible loops can pick different headers
3259 // after a major_progress operation, so the rest of the loop
3260 // tree cannot be matched.
3261 if (_irreducible && Compile::current()->major_progress()) return;
3262 assert( 0, "failed to match loop tree" );
3263 }
3264
3265 // Move (*nn) to (*pp)
3266 IdealLoopTree *hit = *nn;
3267 *nn = hit->_next;
3268 hit->_next = loop;
3269 *pp = loop;
3270 loop = hit;
3271 // Now try again to verify
3272 }
3273
3274 assert( _head == loop->_head , "mismatched loop head" );
3275 Node *tail = _tail; // Inline a non-updating version of
3276 while( !tail->in(0) ) // the 'tail()' call.
3277 tail = tail->in(1);
3278 assert( tail == loop->_tail, "mismatched loop tail" );
3279
3280 // Counted loops that are guarded should be able to find their guards
3281 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
3282 CountedLoopNode *cl = _head->as_CountedLoop();
3283 Node *init = cl->init_trip();
3284 Node *ctrl = cl->in(LoopNode::EntryControl);
3285 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
3286 Node *iff = ctrl->in(0);
3287 assert( iff->Opcode() == Op_If, "" );
3288 Node *bol = iff->in(1);
3289 assert( bol->Opcode() == Op_Bool, "" );
3290 Node *cmp = bol->in(1);
3291 assert( cmp->Opcode() == Op_CmpI, "" );
3292 Node *add = cmp->in(1);
3293 Node *opaq;
3294 if( add->Opcode() == Op_Opaque1 ) {
3295 opaq = add;
3296 } else {
3297 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
3298 assert( add == init, "" );
3299 opaq = cmp->in(2);
3300 }
3301 assert( opaq->Opcode() == Op_Opaque1, "" );
3302
3303 }
3304
3305 if (_child != NULL) _child->verify_tree(loop->_child, this);
3306 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
3307 // Innermost loops need to verify loop bodies,
3308 // but only if no 'major_progress'
3309 int fail = 0;
3310 if (!Compile::current()->major_progress() && _child == NULL) {
3311 for( uint i = 0; i < _body.size(); i++ ) {
3312 Node *n = _body.at(i);
3313 if (n->outcnt() == 0) continue; // Ignore dead
3314 uint j;
3315 for( j = 0; j < loop->_body.size(); j++ )
3316 if( loop->_body.at(j) == n )
3317 break;
3318 if( j == loop->_body.size() ) { // Not found in loop body
3319 // Last ditch effort to avoid assertion: Its possible that we
3320 // have some users (so outcnt not zero) but are still dead.
3321 // Try to find from root.
3322 if (Compile::current()->root()->find(n->_idx)) {
3323 fail++;
3324 tty->print("We have that verify does not: ");
3325 n->dump();
3326 }
3327 }
3328 }
3329 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
3330 Node *n = loop->_body.at(i2);
3331 if (n->outcnt() == 0) continue; // Ignore dead
3332 uint j;
3333 for( j = 0; j < _body.size(); j++ )
3334 if( _body.at(j) == n )
3335 break;
3336 if( j == _body.size() ) { // Not found in loop body
3337 // Last ditch effort to avoid assertion: Its possible that we
3338 // have some users (so outcnt not zero) but are still dead.
3339 // Try to find from root.
3340 if (Compile::current()->root()->find(n->_idx)) {
3341 fail++;
3342 tty->print("Verify has that we do not: ");
3343 n->dump();
3344 }
3345 }
3346 }
3347 assert( !fail, "loop body mismatch" );
3348 }
3349 }
3350
3351 #endif
3352
3353 //------------------------------set_idom---------------------------------------
3354 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
3355 uint idx = d->_idx;
3356 if (idx >= _idom_size) {
3357 uint newsize = _idom_size<<1;
3358 while( idx >= newsize ) {
3359 newsize <<= 1;
3360 }
3361 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
3362 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
3363 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
3364 _idom_size = newsize;
3365 }
3366 _idom[idx] = n;
3367 _dom_depth[idx] = dom_depth;
3368 }
3369
3370 //------------------------------recompute_dom_depth---------------------------------------
3371 // The dominator tree is constructed with only parent pointers.
3372 // This recomputes the depth in the tree by first tagging all
3373 // nodes as "no depth yet" marker. The next pass then runs up
3374 // the dom tree from each node marked "no depth yet", and computes
3375 // the depth on the way back down.
3376 void PhaseIdealLoop::recompute_dom_depth() {
3377 uint no_depth_marker = C->unique();
3378 uint i;
3379 // Initialize depth to "no depth yet" and realize all lazy updates
3380 for (i = 0; i < _idom_size; i++) {
3381 // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized).
3382 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
3383 _dom_depth[i] = no_depth_marker;
3384
3385 // heal _idom if it has a fwd mapping in _nodes
3386 if (_idom[i]->in(0) == NULL) {
3387 idom(i);
3388 }
3389 }
3390 }
3391 if (_dom_stk == NULL) {
3392 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
3393 if (init_size < 10) init_size = 10;
3394 _dom_stk = new GrowableArray<uint>(init_size);
3395 }
3396 // Compute new depth for each node.
3397 for (i = 0; i < _idom_size; i++) {
3398 uint j = i;
3399 // Run up the dom tree to find a node with a depth
3400 while (_dom_depth[j] == no_depth_marker) {
3401 _dom_stk->push(j);
3402 j = _idom[j]->_idx;
3403 }
3404 // Compute the depth on the way back down this tree branch
3405 uint dd = _dom_depth[j] + 1;
3406 while (_dom_stk->length() > 0) {
3407 uint j = _dom_stk->pop();
3408 _dom_depth[j] = dd;
3409 dd++;
3410 }
3411 }
3412 }
3413
3414 //------------------------------sort-------------------------------------------
3415 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
3416 // loop tree, not the root.
3417 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
3418 if( !innermost ) return loop; // New innermost loop
3419
3420 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
3421 assert( loop_preorder, "not yet post-walked loop" );
3422 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
3423 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
3424
3425 // Insert at start of list
3426 while( l ) { // Insertion sort based on pre-order
3427 if( l == loop ) return innermost; // Already on list!
3428 int l_preorder = get_preorder(l->_head); // Cache pre-order number
3429 assert( l_preorder, "not yet post-walked l" );
3430 // Check header pre-order number to figure proper nesting
3431 if( loop_preorder > l_preorder )
3432 break; // End of insertion
3433 // If headers tie (e.g., shared headers) check tail pre-order numbers.
3434 // Since I split shared headers, you'd think this could not happen.
3435 // BUT: I must first do the preorder numbering before I can discover I
3436 // have shared headers, so the split headers all get the same preorder
3437 // number as the RegionNode they split from.
3438 if( loop_preorder == l_preorder &&
3439 get_preorder(loop->_tail) < get_preorder(l->_tail) )
3440 break; // Also check for shared headers (same pre#)
3441 pp = &l->_parent; // Chain up list
3442 l = *pp;
3443 }
3444 // Link into list
3445 // Point predecessor to me
3446 *pp = loop;
3447 // Point me to successor
3448 IdealLoopTree *p = loop->_parent;
3449 loop->_parent = l; // Point me to successor
3450 if( p ) sort( p, innermost ); // Insert my parents into list as well
3451 return innermost;
3452 }
3453
3454 //------------------------------build_loop_tree--------------------------------
3455 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
3456 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
3457 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
3458 // tightest enclosing IdealLoopTree for post-walked.
3459 //
3460 // During my forward walk I do a short 1-layer lookahead to see if I can find
3461 // a loop backedge with that doesn't have any work on the backedge. This
3462 // helps me construct nested loops with shared headers better.
3463 //
3464 // Once I've done the forward recursion, I do the post-work. For each child
3465 // I check to see if there is a backedge. Backedges define a loop! I
3466 // insert an IdealLoopTree at the target of the backedge.
3467 //
3468 // During the post-work I also check to see if I have several children
3469 // belonging to different loops. If so, then this Node is a decision point
3470 // where control flow can choose to change loop nests. It is at this
3471 // decision point where I can figure out how loops are nested. At this
3472 // time I can properly order the different loop nests from my children.
3473 // Note that there may not be any backedges at the decision point!
3474 //
3475 // Since the decision point can be far removed from the backedges, I can't
3476 // order my loops at the time I discover them. Thus at the decision point
3477 // I need to inspect loop header pre-order numbers to properly nest my
3478 // loops. This means I need to sort my childrens' loops by pre-order.
3479 // The sort is of size number-of-control-children, which generally limits
3480 // it to size 2 (i.e., I just choose between my 2 target loops).
3481 void PhaseIdealLoop::build_loop_tree() {
3482 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3483 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
3484 Node *n = C->root();
3485 bltstack.push(n);
3486 int pre_order = 1;
3487 int stack_size;
3488
3489 while ( ( stack_size = bltstack.length() ) != 0 ) {
3490 n = bltstack.top(); // Leave node on stack
3491 if ( !is_visited(n) ) {
3492 // ---- Pre-pass Work ----
3493 // Pre-walked but not post-walked nodes need a pre_order number.
3494
3495 set_preorder_visited( n, pre_order ); // set as visited
3496
3497 // ---- Scan over children ----
3498 // Scan first over control projections that lead to loop headers.
3499 // This helps us find inner-to-outer loops with shared headers better.
3500
3501 // Scan children's children for loop headers.
3502 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
3503 Node* m = n->raw_out(i); // Child
3504 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
3505 // Scan over children's children to find loop
3506 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3507 Node* l = m->fast_out(j);
3508 if( is_visited(l) && // Been visited?
3509 !is_postvisited(l) && // But not post-visited
3510 get_preorder(l) < pre_order ) { // And smaller pre-order
3511 // Found! Scan the DFS down this path before doing other paths
3512 bltstack.push(m);
3513 break;
3514 }
3515 }
3516 }
3517 }
3518 pre_order++;
3519 }
3520 else if ( !is_postvisited(n) ) {
3521 // Note: build_loop_tree_impl() adds out edges on rare occasions,
3522 // such as com.sun.rsasign.am::a.
3523 // For non-recursive version, first, process current children.
3524 // On next iteration, check if additional children were added.
3525 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
3526 Node* u = n->raw_out(k);
3527 if ( u->is_CFG() && !is_visited(u) ) {
3528 bltstack.push(u);
3529 }
3530 }
3531 if ( bltstack.length() == stack_size ) {
3532 // There were no additional children, post visit node now
3533 (void)bltstack.pop(); // Remove node from stack
3534 pre_order = build_loop_tree_impl( n, pre_order );
3535 // Check for bailout
3536 if (C->failing()) {
3537 return;
3538 }
3539 // Check to grow _preorders[] array for the case when
3540 // build_loop_tree_impl() adds new nodes.
3541 check_grow_preorders();
3542 }
3543 }
3544 else {
3545 (void)bltstack.pop(); // Remove post-visited node from stack
3546 }
3547 }
3548 }
3549
3550 //------------------------------build_loop_tree_impl---------------------------
3551 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
3552 // ---- Post-pass Work ----
3553 // Pre-walked but not post-walked nodes need a pre_order number.
3554
3555 // Tightest enclosing loop for this Node
3556 IdealLoopTree *innermost = NULL;
3557
3558 // For all children, see if any edge is a backedge. If so, make a loop
3559 // for it. Then find the tightest enclosing loop for the self Node.
3560 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3561 Node* m = n->fast_out(i); // Child
3562 if( n == m ) continue; // Ignore control self-cycles
3563 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
3564
3565 IdealLoopTree *l; // Child's loop
3566 if( !is_postvisited(m) ) { // Child visited but not post-visited?
3567 // Found a backedge
3568 assert( get_preorder(m) < pre_order, "should be backedge" );
3569 // Check for the RootNode, which is already a LoopNode and is allowed
3570 // to have multiple "backedges".
3571 if( m == C->root()) { // Found the root?
3572 l = _ltree_root; // Root is the outermost LoopNode
3573 } else { // Else found a nested loop
3574 // Insert a LoopNode to mark this loop.
3575 l = new IdealLoopTree(this, m, n);
3576 } // End of Else found a nested loop
3577 if( !has_loop(m) ) // If 'm' does not already have a loop set
3578 set_loop(m, l); // Set loop header to loop now
3579
3580 } else { // Else not a nested loop
3581 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
3582 l = get_loop(m); // Get previously determined loop
3583 // If successor is header of a loop (nest), move up-loop till it
3584 // is a member of some outer enclosing loop. Since there are no
3585 // shared headers (I've split them already) I only need to go up
3586 // at most 1 level.
3587 while( l && l->_head == m ) // Successor heads loop?
3588 l = l->_parent; // Move up 1 for me
3589 // If this loop is not properly parented, then this loop
3590 // has no exit path out, i.e. its an infinite loop.
3591 if( !l ) {
3592 // Make loop "reachable" from root so the CFG is reachable. Basically
3593 // insert a bogus loop exit that is never taken. 'm', the loop head,
3594 // points to 'n', one (of possibly many) fall-in paths. There may be
3595 // many backedges as well.
3596
3597 // Here I set the loop to be the root loop. I could have, after
3598 // inserting a bogus loop exit, restarted the recursion and found my
3599 // new loop exit. This would make the infinite loop a first-class
3600 // loop and it would then get properly optimized. What's the use of
3601 // optimizing an infinite loop?
3602 l = _ltree_root; // Oops, found infinite loop
3603
3604 if (!_verify_only) {
3605 // Insert the NeverBranch between 'm' and it's control user.
3606 NeverBranchNode *iff = new NeverBranchNode( m );
3607 _igvn.register_new_node_with_optimizer(iff);
3608 set_loop(iff, l);
3609 Node *if_t = new CProjNode( iff, 0 );
3610 _igvn.register_new_node_with_optimizer(if_t);
3611 set_loop(if_t, l);
3612
3613 Node* cfg = NULL; // Find the One True Control User of m
3614 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3615 Node* x = m->fast_out(j);
3616 if (x->is_CFG() && x != m && x != iff)
3617 { cfg = x; break; }
3618 }
3619 assert(cfg != NULL, "must find the control user of m");
3620 uint k = 0; // Probably cfg->in(0)
3621 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
3622 cfg->set_req( k, if_t ); // Now point to NeverBranch
3623 _igvn._worklist.push(cfg);
3624
3625 // Now create the never-taken loop exit
3626 Node *if_f = new CProjNode( iff, 1 );
3627 _igvn.register_new_node_with_optimizer(if_f);
3628 set_loop(if_f, l);
3629 // Find frame ptr for Halt. Relies on the optimizer
3630 // V-N'ing. Easier and quicker than searching through
3631 // the program structure.
3632 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3633 _igvn.register_new_node_with_optimizer(frame);
3634 // Halt & Catch Fire
3635 Node *halt = new HaltNode( if_f, frame );
3636 _igvn.register_new_node_with_optimizer(halt);
3637 set_loop(halt, l);
3638 C->root()->add_req(halt);
3639 }
3640 set_loop(C->root(), _ltree_root);
3641 }
3642 }
3643 // Weeny check for irreducible. This child was already visited (this
3644 // IS the post-work phase). Is this child's loop header post-visited
3645 // as well? If so, then I found another entry into the loop.
3646 if (!_verify_only) {
3647 while( is_postvisited(l->_head) ) {
3648 // found irreducible
3649 l->_irreducible = 1; // = true
3650 l = l->_parent;
3651 _has_irreducible_loops = true;
3652 // Check for bad CFG here to prevent crash, and bailout of compile
3653 if (l == NULL) {
3654 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3655 return pre_order;
3656 }
3657 }
3658 C->set_has_irreducible_loop(_has_irreducible_loops);
3659 }
3660
3661 // This Node might be a decision point for loops. It is only if
3662 // it's children belong to several different loops. The sort call
3663 // does a trivial amount of work if there is only 1 child or all
3664 // children belong to the same loop. If however, the children
3665 // belong to different loops, the sort call will properly set the
3666 // _parent pointers to show how the loops nest.
3667 //
3668 // In any case, it returns the tightest enclosing loop.
3669 innermost = sort( l, innermost );
3670 }
3671
3672 // Def-use info will have some dead stuff; dead stuff will have no
3673 // loop decided on.
3674
3675 // Am I a loop header? If so fix up my parent's child and next ptrs.
3676 if( innermost && innermost->_head == n ) {
3677 assert( get_loop(n) == innermost, "" );
3678 IdealLoopTree *p = innermost->_parent;
3679 IdealLoopTree *l = innermost;
3680 while( p && l->_head == n ) {
3681 l->_next = p->_child; // Put self on parents 'next child'
3682 p->_child = l; // Make self as first child of parent
3683 l = p; // Now walk up the parent chain
3684 p = l->_parent;
3685 }
3686 } else {
3687 // Note that it is possible for a LoopNode to reach here, if the
3688 // backedge has been made unreachable (hence the LoopNode no longer
3689 // denotes a Loop, and will eventually be removed).
3690
3691 // Record tightest enclosing loop for self. Mark as post-visited.
3692 set_loop(n, innermost);
3693 // Also record has_call flag early on
3694 if( innermost ) {
3695 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3696 // Do not count uncommon calls
3697 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3698 Node *iff = n->in(0)->in(0);
3699 // No any calls for vectorized loops.
3700 if( UseSuperWord || !iff->is_If() ||
3701 (n->in(0)->Opcode() == Op_IfFalse &&
3702 (1.0 - iff->as_If()->_prob) >= 0.01) ||
3703 (iff->as_If()->_prob >= 0.01) )
3704 innermost->_has_call = 1;
3705 }
3706 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3707 // Disable loop optimizations if the loop has a scalar replaceable
3708 // allocation. This disabling may cause a potential performance lost
3709 // if the allocation is not eliminated for some reason.
3710 innermost->_allow_optimizations = false;
3711 innermost->_has_call = 1; // = true
3712 } else if (n->Opcode() == Op_SafePoint) {
3713 // Record all safepoints in this loop.
3714 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3715 innermost->_safepts->push(n);
3716 }
3717 }
3718 }
3719
3720 // Flag as post-visited now
3721 set_postvisited(n);
3722 return pre_order;
3723 }
3724
3725
3726 //------------------------------build_loop_early-------------------------------
3727 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3728 // First pass computes the earliest controlling node possible. This is the
3729 // controlling input with the deepest dominating depth.
3730 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3731 while (worklist.size() != 0) {
3732 // Use local variables nstack_top_n & nstack_top_i to cache values
3733 // on nstack's top.
3734 Node *nstack_top_n = worklist.pop();
3735 uint nstack_top_i = 0;
3736 //while_nstack_nonempty:
3737 while (true) {
3738 // Get parent node and next input's index from stack's top.
3739 Node *n = nstack_top_n;
3740 uint i = nstack_top_i;
3741 uint cnt = n->req(); // Count of inputs
3742 if (i == 0) { // Pre-process the node.
3743 if( has_node(n) && // Have either loop or control already?
3744 !has_ctrl(n) ) { // Have loop picked out already?
3745 // During "merge_many_backedges" we fold up several nested loops
3746 // into a single loop. This makes the members of the original
3747 // loop bodies pointing to dead loops; they need to move up
3748 // to the new UNION'd larger loop. I set the _head field of these
3749 // dead loops to NULL and the _parent field points to the owning
3750 // loop. Shades of UNION-FIND algorithm.
3751 IdealLoopTree *ilt;
3752 while( !(ilt = get_loop(n))->_head ) {
3753 // Normally I would use a set_loop here. But in this one special
3754 // case, it is legal (and expected) to change what loop a Node
3755 // belongs to.
3756 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3757 }
3758 // Remove safepoints ONLY if I've already seen I don't need one.
3759 // (the old code here would yank a 2nd safepoint after seeing a
3760 // first one, even though the 1st did not dominate in the loop body
3761 // and thus could be avoided indefinitely)
3762 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3763 is_deleteable_safept(n)) {
3764 Node *in = n->in(TypeFunc::Control);
3765 lazy_replace(n,in); // Pull safepoint now
3766 if (ilt->_safepts != NULL) {
3767 ilt->_safepts->yank(n);
3768 }
3769 // Carry on with the recursion "as if" we are walking
3770 // only the control input
3771 if( !visited.test_set( in->_idx ) ) {
3772 worklist.push(in); // Visit this guy later, using worklist
3773 }
3774 // Get next node from nstack:
3775 // - skip n's inputs processing by setting i > cnt;
3776 // - we also will not call set_early_ctrl(n) since
3777 // has_node(n) == true (see the condition above).
3778 i = cnt + 1;
3779 }
3780 }
3781 } // if (i == 0)
3782
3783 // Visit all inputs
3784 bool done = true; // Assume all n's inputs will be processed
3785 while (i < cnt) {
3786 Node *in = n->in(i);
3787 ++i;
3788 if (in == NULL) continue;
3789 if (in->pinned() && !in->is_CFG())
3790 set_ctrl(in, in->in(0));
3791 int is_visited = visited.test_set( in->_idx );
3792 if (!has_node(in)) { // No controlling input yet?
3793 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3794 assert( !is_visited, "visit only once" );
3795 nstack.push(n, i); // Save parent node and next input's index.
3796 nstack_top_n = in; // Process current input now.
3797 nstack_top_i = 0;
3798 done = false; // Not all n's inputs processed.
3799 break; // continue while_nstack_nonempty;
3800 } else if (!is_visited) {
3801 // This guy has a location picked out for him, but has not yet
3802 // been visited. Happens to all CFG nodes, for instance.
3803 // Visit him using the worklist instead of recursion, to break
3804 // cycles. Since he has a location already we do not need to
3805 // find his location before proceeding with the current Node.
3806 worklist.push(in); // Visit this guy later, using worklist
3807 }
3808 }
3809 if (done) {
3810 // All of n's inputs have been processed, complete post-processing.
3811
3812 // Compute earliest point this Node can go.
3813 // CFG, Phi, pinned nodes already know their controlling input.
3814 if (!has_node(n)) {
3815 // Record earliest legal location
3816 set_early_ctrl( n );
3817 }
3818 if (nstack.is_empty()) {
3819 // Finished all nodes on stack.
3820 // Process next node on the worklist.
3821 break;
3822 }
3823 // Get saved parent node and next input's index.
3824 nstack_top_n = nstack.node();
3825 nstack_top_i = nstack.index();
3826 nstack.pop();
3827 }
3828 } // while (true)
3829 }
3830 }
3831
3832 //------------------------------dom_lca_internal--------------------------------
3833 // Pair-wise LCA
3834 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3835 if( !n1 ) return n2; // Handle NULL original LCA
3836 assert( n1->is_CFG(), "" );
3837 assert( n2->is_CFG(), "" );
3838 // find LCA of all uses
3839 uint d1 = dom_depth(n1);
3840 uint d2 = dom_depth(n2);
3841 while (n1 != n2) {
3842 if (d1 > d2) {
3843 n1 = idom(n1);
3844 d1 = dom_depth(n1);
3845 } else if (d1 < d2) {
3846 n2 = idom(n2);
3847 d2 = dom_depth(n2);
3848 } else {
3849 // Here d1 == d2. Due to edits of the dominator-tree, sections
3850 // of the tree might have the same depth. These sections have
3851 // to be searched more carefully.
3852
3853 // Scan up all the n1's with equal depth, looking for n2.
3854 Node *t1 = idom(n1);
3855 while (dom_depth(t1) == d1) {
3856 if (t1 == n2) return n2;
3857 t1 = idom(t1);
3858 }
3859 // Scan up all the n2's with equal depth, looking for n1.
3860 Node *t2 = idom(n2);
3861 while (dom_depth(t2) == d2) {
3862 if (t2 == n1) return n1;
3863 t2 = idom(t2);
3864 }
3865 // Move up to a new dominator-depth value as well as up the dom-tree.
3866 n1 = t1;
3867 n2 = t2;
3868 d1 = dom_depth(n1);
3869 d2 = dom_depth(n2);
3870 }
3871 }
3872 return n1;
3873 }
3874
3875 //------------------------------compute_idom-----------------------------------
3876 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3877 // IDOMs are correct.
3878 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3879 assert( region->is_Region(), "" );
3880 Node *LCA = NULL;
3881 for( uint i = 1; i < region->req(); i++ ) {
3882 if( region->in(i) != C->top() )
3883 LCA = dom_lca( LCA, region->in(i) );
3884 }
3885 return LCA;
3886 }
3887
3888 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3889 bool had_error = false;
3890 #ifdef ASSERT
3891 if (early != C->root()) {
3892 // Make sure that there's a dominance path from LCA to early
3893 Node* d = LCA;
3894 while (d != early) {
3895 if (d == C->root()) {
3896 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3897 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3898 had_error = true;
3899 break;
3900 }
3901 d = idom(d);
3902 }
3903 }
3904 #endif
3905 return had_error;
3906 }
3907
3908
3909 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3910 // Compute LCA over list of uses
3911 bool had_error = false;
3912 Node *LCA = NULL;
3913 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3914 Node* c = n->fast_out(i);
3915 if (_nodes[c->_idx] == NULL)
3916 continue; // Skip the occasional dead node
3917 if( c->is_Phi() ) { // For Phis, we must land above on the path
3918 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3919 if( c->in(j) == n ) { // Found matching input?
3920 Node *use = c->in(0)->in(j);
3921 if (_verify_only && use->is_top()) continue;
3922 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3923 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3924 }
3925 }
3926 } else {
3927 // For CFG data-users, use is in the block just prior
3928 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3929 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3930 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3931 }
3932 }
3933 assert(!had_error, "bad dominance");
3934 return LCA;
3935 }
3936
3937 // Check the shape of the graph at the loop entry. In some cases,
3938 // the shape of the graph does not match the shape outlined below.
3939 // That is caused by the Opaque1 node "protecting" the shape of
3940 // the graph being removed by, for example, the IGVN performed
3941 // in PhaseIdealLoop::build_and_optimize().
3942 //
3943 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3944 // loop unswitching, and IGVN, or a combination of them) can freely change
3945 // the graph's shape. As a result, the graph shape outlined below cannot
3946 // be guaranteed anymore.
3947 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3948 if (!cl->is_main_loop() && !cl->is_post_loop()) {
3949 return false;
3950 }
3951 Node* ctrl = cl->skip_predicates();
3952
3953 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3954 return false;
3955 }
3956 Node* iffm = ctrl->in(0);
3957 if (iffm == NULL || !iffm->is_If()) {
3958 return false;
3959 }
3960 Node* bolzm = iffm->in(1);
3961 if (bolzm == NULL || !bolzm->is_Bool()) {
3962 return false;
3963 }
3964 Node* cmpzm = bolzm->in(1);
3965 if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3966 return false;
3967 }
3968 // compares can get conditionally flipped
3969 bool found_opaque = false;
3970 for (uint i = 1; i < cmpzm->req(); i++) {
3971 Node* opnd = cmpzm->in(i);
3972 if (opnd && opnd->Opcode() == Op_Opaque1) {
3973 found_opaque = true;
3974 break;
3975 }
3976 }
3977 if (!found_opaque) {
3978 return false;
3979 }
3980 return true;
3981 }
3982
3983 //------------------------------get_late_ctrl----------------------------------
3984 // Compute latest legal control.
3985 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3986 assert(early != NULL, "early control should not be NULL");
3987
3988 Node* LCA = compute_lca_of_uses(n, early);
3989 #ifdef ASSERT
3990 if (LCA == C->root() && LCA != early) {
3991 // def doesn't dominate uses so print some useful debugging output
3992 compute_lca_of_uses(n, early, true);
3993 }
3994 #endif
3995
3996 // if this is a load, check for anti-dependent stores
3997 // We use a conservative algorithm to identify potential interfering
3998 // instructions and for rescheduling the load. The users of the memory
3999 // input of this load are examined. Any use which is not a load and is
4000 // dominated by early is considered a potentially interfering store.
4001 // This can produce false positives.
4002 if (n->is_Load() && LCA != early) {
4003 Node_List worklist;
4004
4005 Node *mem = n->in(MemNode::Memory);
4006 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
4007 Node* s = mem->fast_out(i);
4008 worklist.push(s);
4009 }
4010 while(worklist.size() != 0 && LCA != early) {
4011 Node* s = worklist.pop();
4012 if (s->is_Load() || s->is_ShenandoahBarrier() || s->Opcode() == Op_SafePoint ||
4013 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) {
4014 continue;
4015 } else if (s->is_MergeMem()) {
4016 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
4017 Node* s1 = s->fast_out(i);
4018 worklist.push(s1);
4019 }
4020 } else {
4021 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
4022 assert(sctrl != NULL || s->outcnt() == 0 || s->is_ShenandoahBarrier(), "must have control");
4023 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
4024 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
4025 }
4026 }
4027 }
4028 }
4029
4030 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
4031 return LCA;
4032 }
4033
4034 // true if CFG node d dominates CFG node n
4035 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
4036 if (d == n)
4037 return true;
4038 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
4039 uint dd = dom_depth(d);
4040 while (dom_depth(n) >= dd) {
4041 if (n == d)
4042 return true;
4043 n = idom(n);
4044 }
4045 return false;
4046 }
4047
4048 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
4049 // Pair-wise LCA with tags.
4050 // Tag each index with the node 'tag' currently being processed
4051 // before advancing up the dominator chain using idom().
4052 // Later calls that find a match to 'tag' know that this path has already
4053 // been considered in the current LCA (which is input 'n1' by convention).
4054 // Since get_late_ctrl() is only called once for each node, the tag array
4055 // does not need to be cleared between calls to get_late_ctrl().
4056 // Algorithm trades a larger constant factor for better asymptotic behavior
4057 //
4058 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
4059 uint d1 = dom_depth(n1);
4060 uint d2 = dom_depth(n2);
4061
4062 do {
4063 if (d1 > d2) {
4064 // current lca is deeper than n2
4065 _dom_lca_tags.map(n1->_idx, tag);
4066 n1 = idom(n1);
4067 d1 = dom_depth(n1);
4068 } else if (d1 < d2) {
4069 // n2 is deeper than current lca
4070 Node *memo = _dom_lca_tags[n2->_idx];
4071 if( memo == tag ) {
4072 return n1; // Return the current LCA
4073 }
4074 _dom_lca_tags.map(n2->_idx, tag);
4075 n2 = idom(n2);
4076 d2 = dom_depth(n2);
4077 } else {
4078 // Here d1 == d2. Due to edits of the dominator-tree, sections
4079 // of the tree might have the same depth. These sections have
4080 // to be searched more carefully.
4081
4082 // Scan up all the n1's with equal depth, looking for n2.
4083 _dom_lca_tags.map(n1->_idx, tag);
4084 Node *t1 = idom(n1);
4085 while (dom_depth(t1) == d1) {
4086 if (t1 == n2) return n2;
4087 _dom_lca_tags.map(t1->_idx, tag);
4088 t1 = idom(t1);
4089 }
4090 // Scan up all the n2's with equal depth, looking for n1.
4091 _dom_lca_tags.map(n2->_idx, tag);
4092 Node *t2 = idom(n2);
4093 while (dom_depth(t2) == d2) {
4094 if (t2 == n1) return n1;
4095 _dom_lca_tags.map(t2->_idx, tag);
4096 t2 = idom(t2);
4097 }
4098 // Move up to a new dominator-depth value as well as up the dom-tree.
4099 n1 = t1;
4100 n2 = t2;
4101 d1 = dom_depth(n1);
4102 d2 = dom_depth(n2);
4103 }
4104 } while (n1 != n2);
4105 return n1;
4106 }
4107
4108 //------------------------------init_dom_lca_tags------------------------------
4109 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4110 // Intended use does not involve any growth for the array, so it could
4111 // be of fixed size.
4112 void PhaseIdealLoop::init_dom_lca_tags() {
4113 uint limit = C->unique() + 1;
4114 _dom_lca_tags.map( limit, NULL );
4115 #ifdef ASSERT
4116 for( uint i = 0; i < limit; ++i ) {
4117 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4118 }
4119 #endif // ASSERT
4120 }
4121
4122 //------------------------------clear_dom_lca_tags------------------------------
4123 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4124 // Intended use does not involve any growth for the array, so it could
4125 // be of fixed size.
4126 void PhaseIdealLoop::clear_dom_lca_tags() {
4127 uint limit = C->unique() + 1;
4128 _dom_lca_tags.map( limit, NULL );
4129 _dom_lca_tags.clear();
4130 #ifdef ASSERT
4131 for( uint i = 0; i < limit; ++i ) {
4132 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4133 }
4134 #endif // ASSERT
4135 }
4136
4137 //------------------------------build_loop_late--------------------------------
4138 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4139 // Second pass finds latest legal placement, and ideal loop placement.
4140 void PhaseIdealLoop::build_loop_late(VectorSet &visited, Node_List &worklist, Node_Stack &nstack, bool verify_strip_mined) {
4141 while (worklist.size() != 0) {
4142 Node *n = worklist.pop();
4143 // Only visit once
4144 if (visited.test_set(n->_idx)) continue;
4145 uint cnt = n->outcnt();
4146 uint i = 0;
4147 while (true) {
4148 assert( _nodes[n->_idx], "no dead nodes" );
4149 // Visit all children
4150 if (i < cnt) {
4151 Node* use = n->raw_out(i);
4152 ++i;
4153 // Check for dead uses. Aggressively prune such junk. It might be
4154 // dead in the global sense, but still have local uses so I cannot
4155 // easily call 'remove_dead_node'.
4156 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
4157 // Due to cycles, we might not hit the same fixed point in the verify
4158 // pass as we do in the regular pass. Instead, visit such phis as
4159 // simple uses of the loop head.
4160 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
4161 if( !visited.test(use->_idx) )
4162 worklist.push(use);
4163 } else if( !visited.test_set(use->_idx) ) {
4164 nstack.push(n, i); // Save parent and next use's index.
4165 n = use; // Process all children of current use.
4166 cnt = use->outcnt();
4167 i = 0;
4168 }
4169 } else {
4170 // Do not visit around the backedge of loops via data edges.
4171 // push dead code onto a worklist
4172 _deadlist.push(use);
4173 }
4174 } else {
4175 // All of n's children have been processed, complete post-processing.
4176 build_loop_late_post(n, verify_strip_mined);
4177 if (nstack.is_empty()) {
4178 // Finished all nodes on stack.
4179 // Process next node on the worklist.
4180 break;
4181 }
4182 // Get saved parent node and next use's index. Visit the rest of uses.
4183 n = nstack.node();
4184 cnt = n->outcnt();
4185 i = nstack.index();
4186 nstack.pop();
4187 }
4188 }
4189 }
4190 }
4191
4192 // Verify that no data node is scheduled in the outer loop of a strip
4193 // mined loop.
4194 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
4195 #ifdef ASSERT
4196 if (get_loop(least)->_nest == 0) {
4197 return;
4198 }
4199 IdealLoopTree* loop = get_loop(least);
4200 Node* head = loop->_head;
4201 if (head->is_OuterStripMinedLoop()) {
4202 Node* sfpt = head->as_Loop()->outer_safepoint();
4203 ResourceMark rm;
4204 Unique_Node_List wq;
4205 wq.push(sfpt);
4206 for (uint i = 0; i < wq.size(); i++) {
4207 Node *m = wq.at(i);
4208 for (uint i = 1; i < m->req(); i++) {
4209 Node* nn = m->in(i);
4210 if (nn == n) {
4211 return;
4212 }
4213 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
4214 wq.push(nn);
4215 }
4216 }
4217 }
4218 ShouldNotReachHere();
4219 }
4220 #endif
4221 }
4222
4223
4224 //------------------------------build_loop_late_post---------------------------
4225 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4226 // Second pass finds latest legal placement, and ideal loop placement.
4227 void PhaseIdealLoop::build_loop_late_post(Node *n, bool verify_strip_mined) {
4228
4229 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
4230 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
4231 }
4232
4233 #ifdef ASSERT
4234 if (_verify_only && !n->is_CFG()) {
4235 // Check def-use domination.
4236 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
4237 }
4238 #endif
4239
4240 // CFG and pinned nodes already handled
4241 if( n->in(0) ) {
4242 if( n->in(0)->is_top() ) return; // Dead?
4243
4244 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
4245 // _must_ be pinned (they have to observe their control edge of course).
4246 // Unlike Stores (which modify an unallocable resource, the memory
4247 // state), Mods/Loads can float around. So free them up.
4248 bool pinned = true;
4249 switch( n->Opcode() ) {
4250 case Op_DivI:
4251 case Op_DivF:
4252 case Op_DivD:
4253 case Op_ModI:
4254 case Op_ModF:
4255 case Op_ModD:
4256 case Op_LoadB: // Same with Loads; they can sink
4257 case Op_LoadUB: // during loop optimizations.
4258 case Op_LoadUS:
4259 case Op_LoadD:
4260 case Op_LoadF:
4261 case Op_LoadI:
4262 case Op_LoadKlass:
4263 case Op_LoadNKlass:
4264 case Op_LoadL:
4265 case Op_LoadS:
4266 case Op_LoadP:
4267 case Op_LoadBarrierSlowReg:
4268 case Op_LoadBarrierWeakSlowReg:
4269 case Op_LoadN:
4270 case Op_LoadRange:
4271 case Op_LoadD_unaligned:
4272 case Op_LoadL_unaligned:
4273 case Op_StrComp: // Does a bunch of load-like effects
4274 case Op_StrEquals:
4275 case Op_StrIndexOf:
4276 case Op_StrIndexOfChar:
4277 case Op_AryEq:
4278 #if INCLUDE_SHENANDOAHGC
4279 case Op_ShenandoahReadBarrier:
4280 case Op_ShenandoahWriteBarrier:
4281 case Op_ShenandoahWBMemProj:
4282 #endif
4283 case Op_HasNegatives:
4284 pinned = false;
4285 }
4286
4287 if( pinned ) {
4288 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
4289 if( !chosen_loop->_child ) // Inner loop?
4290 chosen_loop->_body.push(n); // Collect inner loops
4291 return;
4292 }
4293 } else { // No slot zero
4294 if( n->is_CFG() ) { // CFG with no slot 0 is dead
4295 _nodes.map(n->_idx,0); // No block setting, it's globally dead
4296 return;
4297 }
4298 assert(!n->is_CFG() || n->outcnt() == 0, "");
4299 }
4300
4301 // Do I have a "safe range" I can select over?
4302 Node *early = get_ctrl(n);// Early location already computed
4303
4304 // Compute latest point this Node can go
4305 Node *LCA = get_late_ctrl( n, early );
4306 // LCA is NULL due to uses being dead
4307 if( LCA == NULL ) {
4308 #ifdef ASSERT
4309 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
4310 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
4311 }
4312 #endif
4313 _nodes.map(n->_idx, 0); // This node is useless
4314 _deadlist.push(n);
4315 return;
4316 }
4317 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
4318
4319 Node *legal = LCA; // Walk 'legal' up the IDOM chain
4320 Node *least = legal; // Best legal position so far
4321 while( early != legal ) { // While not at earliest legal
4322 #ifdef ASSERT
4323 if (legal->is_Start() && !early->is_Root()) {
4324 // Bad graph. Print idom path and fail.
4325 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
4326 assert(false, "Bad graph detected in build_loop_late");
4327 }
4328 #endif
4329 // Find least loop nesting depth
4330 legal = idom(legal); // Bump up the IDOM tree
4331 // Check for lower nesting depth
4332 if( get_loop(legal)->_nest < get_loop(least)->_nest )
4333 least = legal;
4334 }
4335 assert(early == legal || legal != C->root(), "bad dominance of inputs");
4336
4337 // Try not to place code on a loop entry projection
4338 // which can inhibit range check elimination.
4339 if (least != early) {
4340 Node* ctrl_out = least->unique_ctrl_out();
4341 if (ctrl_out && ctrl_out->is_Loop() &&
4342 least == ctrl_out->in(LoopNode::EntryControl)) {
4343 // Move the node above predicates as far up as possible so a
4344 // following pass of loop predication doesn't hoist a predicate
4345 // that depends on it above that node.
4346 Node* new_ctrl = least;
4347 for (;;) {
4348 if (!new_ctrl->is_Proj()) {
4349 break;
4350 }
4351 CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
4352 if (call == NULL) {
4353 break;
4354 }
4355 int req = call->uncommon_trap_request();
4356 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
4357 if (trap_reason != Deoptimization::Reason_loop_limit_check &&
4358 trap_reason != Deoptimization::Reason_predicate &&
4359 trap_reason != Deoptimization::Reason_profile_predicate) {
4360 break;
4361 }
4362 Node* c = new_ctrl->in(0)->in(0);
4363 if (is_dominator(c, early) && c != early) {
4364 break;
4365 }
4366 new_ctrl = c;
4367 }
4368 least = new_ctrl;
4369 }
4370 }
4371
4372 #ifdef ASSERT
4373 // If verifying, verify that 'verify_me' has a legal location
4374 // and choose it as our location.
4375 if( _verify_me ) {
4376 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
4377 Node *legal = LCA;
4378 while( early != legal ) { // While not at earliest legal
4379 if( legal == v_ctrl ) break; // Check for prior good location
4380 legal = idom(legal) ;// Bump up the IDOM tree
4381 }
4382 // Check for prior good location
4383 if( legal == v_ctrl ) least = legal; // Keep prior if found
4384 }
4385 #endif
4386
4387 // Assign discovered "here or above" point
4388 least = find_non_split_ctrl(least);
4389 if (verify_strip_mined && !_verify_only) {
4390 verify_strip_mined_scheduling(n, least);
4391 }
4392 set_ctrl(n, least);
4393
4394 // Collect inner loop bodies
4395 IdealLoopTree *chosen_loop = get_loop(least);
4396 if( !chosen_loop->_child ) // Inner loop?
4397 chosen_loop->_body.push(n);// Collect inner loops
4398
4399 if (n->Opcode() == Op_ShenandoahWriteBarrier) {
4400 // The write barrier and its memory proj must have the same
4401 // control otherwise some loop opts could put nodes (Phis) between
4402 // them
4403 Node* proj = n->find_out_with(Op_ShenandoahWBMemProj);
4404 if (proj != NULL) {
4405 set_ctrl_and_loop(proj, least);
4406 }
4407 }
4408 }
4409
4410 #ifdef ASSERT
4411 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
4412 tty->print_cr("%s", msg);
4413 tty->print("n: "); n->dump();
4414 tty->print("early(n): "); early->dump();
4415 if (n->in(0) != NULL && !n->in(0)->is_top() &&
4416 n->in(0) != early && !n->in(0)->is_Root()) {
4417 tty->print("n->in(0): "); n->in(0)->dump();
4418 }
4419 for (uint i = 1; i < n->req(); i++) {
4420 Node* in1 = n->in(i);
4421 if (in1 != NULL && in1 != n && !in1->is_top()) {
4422 tty->print("n->in(%d): ", i); in1->dump();
4423 Node* in1_early = get_ctrl(in1);
4424 tty->print("early(n->in(%d)): ", i); in1_early->dump();
4425 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
4426 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
4427 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
4428 }
4429 for (uint j = 1; j < in1->req(); j++) {
4430 Node* in2 = in1->in(j);
4431 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
4432 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
4433 Node* in2_early = get_ctrl(in2);
4434 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
4435 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
4436 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
4437 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
4438 }
4439 }
4440 }
4441 }
4442 }
4443 tty->cr();
4444 tty->print("LCA(n): "); LCA->dump();
4445 for (uint i = 0; i < n->outcnt(); i++) {
4446 Node* u1 = n->raw_out(i);
4447 if (u1 == n)
4448 continue;
4449 tty->print("n->out(%d): ", i); u1->dump();
4450 if (u1->is_CFG()) {
4451 for (uint j = 0; j < u1->outcnt(); j++) {
4452 Node* u2 = u1->raw_out(j);
4453 if (u2 != u1 && u2 != n && u2->is_CFG()) {
4454 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4455 }
4456 }
4457 } else {
4458 Node* u1_later = get_ctrl(u1);
4459 tty->print("later(n->out(%d)): ", i); u1_later->dump();
4460 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
4461 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
4462 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
4463 }
4464 for (uint j = 0; j < u1->outcnt(); j++) {
4465 Node* u2 = u1->raw_out(j);
4466 if (u2 == n || u2 == u1)
4467 continue;
4468 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4469 if (!u2->is_CFG()) {
4470 Node* u2_later = get_ctrl(u2);
4471 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
4472 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
4473 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
4474 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
4475 }
4476 }
4477 }
4478 }
4479 }
4480 tty->cr();
4481 int ct = 0;
4482 Node *dbg_legal = LCA;
4483 while(!dbg_legal->is_Start() && ct < 100) {
4484 tty->print("idom[%d] ",ct); dbg_legal->dump();
4485 ct++;
4486 dbg_legal = idom(dbg_legal);
4487 }
4488 tty->cr();
4489 }
4490 #endif
4491
4492 #ifndef PRODUCT
4493 //------------------------------dump-------------------------------------------
4494 void PhaseIdealLoop::dump( ) const {
4495 ResourceMark rm;
4496 Arena* arena = Thread::current()->resource_area();
4497 Node_Stack stack(arena, C->live_nodes() >> 2);
4498 Node_List rpo_list;
4499 VectorSet visited(arena);
4500 visited.set(C->top()->_idx);
4501 rpo( C->root(), stack, visited, rpo_list );
4502 // Dump root loop indexed by last element in PO order
4503 dump( _ltree_root, rpo_list.size(), rpo_list );
4504 }
4505
4506 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
4507 loop->dump_head();
4508
4509 // Now scan for CFG nodes in the same loop
4510 for( uint j=idx; j > 0; j-- ) {
4511 Node *n = rpo_list[j-1];
4512 if( !_nodes[n->_idx] ) // Skip dead nodes
4513 continue;
4514 if( get_loop(n) != loop ) { // Wrong loop nest
4515 if( get_loop(n)->_head == n && // Found nested loop?
4516 get_loop(n)->_parent == loop )
4517 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
4518 continue;
4519 }
4520
4521 // Dump controlling node
4522 for( uint x = 0; x < loop->_nest; x++ )
4523 tty->print(" ");
4524 tty->print("C");
4525 if( n == C->root() ) {
4526 n->dump();
4527 } else {
4528 Node* cached_idom = idom_no_update(n);
4529 Node *computed_idom = n->in(0);
4530 if( n->is_Region() ) {
4531 computed_idom = compute_idom(n);
4532 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
4533 // any MultiBranch ctrl node), so apply a similar transform to
4534 // the cached idom returned from idom_no_update.
4535 cached_idom = find_non_split_ctrl(cached_idom);
4536 }
4537 tty->print(" ID:%d",computed_idom->_idx);
4538 n->dump();
4539 if( cached_idom != computed_idom ) {
4540 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
4541 computed_idom->_idx, cached_idom->_idx);
4542 }
4543 }
4544 // Dump nodes it controls
4545 for( uint k = 0; k < _nodes.Size(); k++ ) {
4546 // (k < C->unique() && get_ctrl(find(k)) == n)
4547 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
4548 Node *m = C->root()->find(k);
4549 if( m && m->outcnt() > 0 ) {
4550 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
4551 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
4552 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
4553 }
4554 for( uint j = 0; j < loop->_nest; j++ )
4555 tty->print(" ");
4556 tty->print(" ");
4557 m->dump();
4558 }
4559 }
4560 }
4561 }
4562 }
4563 #endif
4564
4565 // Collect a R-P-O for the whole CFG.
4566 // Result list is in post-order (scan backwards for RPO)
4567 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
4568 stk.push(start, 0);
4569 visited.set(start->_idx);
4570
4571 while (stk.is_nonempty()) {
4572 Node* m = stk.node();
4573 uint idx = stk.index();
4574 if (idx < m->outcnt()) {
4575 stk.set_index(idx + 1);
4576 Node* n = m->raw_out(idx);
4577 if (n->is_CFG() && !visited.test_set(n->_idx)) {
4578 stk.push(n, 0);
4579 }
4580 } else {
4581 rpo_list.push(m);
4582 stk.pop();
4583 }
4584 }
4585 }
4586
4587
4588 //=============================================================================
4589 //------------------------------LoopTreeIterator-----------------------------------
4590
4591 // Advance to next loop tree using a preorder, left-to-right traversal.
4592 void LoopTreeIterator::next() {
4593 assert(!done(), "must not be done.");
4594 if (_curnt->_child != NULL) {
4595 _curnt = _curnt->_child;
4596 } else if (_curnt->_next != NULL) {
4597 _curnt = _curnt->_next;
4598 } else {
4599 while (_curnt != _root && _curnt->_next == NULL) {
4600 _curnt = _curnt->_parent;
4601 }
4602 if (_curnt == _root) {
4603 _curnt = NULL;
4604 assert(done(), "must be done.");
4605 } else {
4606 assert(_curnt->_next != NULL, "must be more to do");
4607 _curnt = _curnt->_next;
4608 }
4609 }
4610 }