1 /*
2 * Copyright (c) 1998, 2019, 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
44 //=============================================================================
45 //--------------------------is_cloop_ind_var-----------------------------------
46 // Determine if a node is a counted loop induction variable.
47 // NOTE: The method is declared in "node.hpp".
48 bool Node::is_cloop_ind_var() const {
49 return (is_Phi() && !as_Phi()->is_copy() &&
50 as_Phi()->region()->is_CountedLoop() &&
51 as_Phi()->region()->as_CountedLoop()->phi() == this);
52 }
53
54 //=============================================================================
55 //------------------------------dump_spec--------------------------------------
56 // Dump special per-node info
57 #ifndef PRODUCT
58 void LoopNode::dump_spec(outputStream *st) const {
59 if (is_inner_loop()) st->print( "inner " );
60 if (is_partial_peel_loop()) st->print( "partial_peel " );
61 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
62 }
63 #endif
64
65 //------------------------------is_valid_counted_loop-------------------------
66 bool LoopNode::is_valid_counted_loop() const {
67 if (is_CountedLoop()) {
68 CountedLoopNode* l = as_CountedLoop();
69 CountedLoopEndNode* le = l->loopexit_or_null();
70 if (le != NULL &&
71 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
72 Node* phi = l->phi();
73 Node* exit = le->proj_out_or_null(0 /* false */);
74 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
75 phi != NULL && phi->is_Phi() &&
76 phi->in(LoopNode::LoopBackControl) == l->incr() &&
77 le->loopnode() == l && le->stride_is_con()) {
78 return true;
79 }
80 }
81 }
82 return false;
83 }
84
85 //------------------------------get_early_ctrl---------------------------------
86 // Compute earliest legal control
87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
88 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
89 uint i;
90 Node *early;
91 if (n->in(0) && !n->is_expensive()) {
92 early = n->in(0);
93 if (!early->is_CFG()) // Might be a non-CFG multi-def
94 early = get_ctrl(early); // So treat input as a straight data input
95 i = 1;
96 } else {
97 early = get_ctrl(n->in(1));
98 i = 2;
99 }
100 uint e_d = dom_depth(early);
101 assert( early, "" );
102 for (; i < n->req(); i++) {
103 Node *cin = get_ctrl(n->in(i));
104 assert( cin, "" );
105 // Keep deepest dominator depth
106 uint c_d = dom_depth(cin);
107 if (c_d > e_d) { // Deeper guy?
108 early = cin; // Keep deepest found so far
109 e_d = c_d;
110 } else if (c_d == e_d && // Same depth?
111 early != cin) { // If not equal, must use slower algorithm
112 // If same depth but not equal, one _must_ dominate the other
113 // and we want the deeper (i.e., dominated) guy.
114 Node *n1 = early;
115 Node *n2 = cin;
116 while (1) {
117 n1 = idom(n1); // Walk up until break cycle
118 n2 = idom(n2);
119 if (n1 == cin || // Walked early up to cin
120 dom_depth(n2) < c_d)
121 break; // early is deeper; keep him
122 if (n2 == early || // Walked cin up to early
123 dom_depth(n1) < c_d) {
124 early = cin; // cin is deeper; keep him
125 break;
126 }
127 }
128 e_d = dom_depth(early); // Reset depth register cache
129 }
130 }
131
132 // Return earliest legal location
133 assert(early == find_non_split_ctrl(early), "unexpected early control");
134
135 if (n->is_expensive() && !_verify_only && !_verify_me) {
136 assert(n->in(0), "should have control input");
137 early = get_early_ctrl_for_expensive(n, early);
138 }
139
140 return early;
141 }
142
143 //------------------------------get_early_ctrl_for_expensive---------------------------------
144 // Move node up the dominator tree as high as legal while still beneficial
145 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
146 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
147 assert(OptimizeExpensiveOps, "optimization off?");
148
149 Node* ctl = n->in(0);
150 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
151 uint min_dom_depth = dom_depth(earliest);
152 #ifdef ASSERT
153 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
154 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
155 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
156 }
157 #endif
158 if (dom_depth(ctl) < min_dom_depth) {
159 return earliest;
160 }
161
162 while (1) {
163 Node *next = ctl;
164 // Moving the node out of a loop on the projection of a If
165 // confuses loop predication. So once we hit a Loop in a If branch
166 // that doesn't branch to an UNC, we stop. The code that process
167 // expensive nodes will notice the loop and skip over it to try to
168 // move the node further up.
169 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
170 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
171 break;
172 }
173 next = idom(ctl->in(1)->in(0));
174 } else if (ctl->is_Proj()) {
175 // We only move it up along a projection if the projection is
176 // the single control projection for its parent: same code path,
177 // if it's a If with UNC or fallthrough of a call.
178 Node* parent_ctl = ctl->in(0);
179 if (parent_ctl == NULL) {
180 break;
181 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
182 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
183 } else if (parent_ctl->is_If()) {
184 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
185 break;
186 }
187 assert(idom(ctl) == parent_ctl, "strange");
188 next = idom(parent_ctl);
189 } else if (ctl->is_CatchProj()) {
190 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
191 break;
192 }
193 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
194 next = parent_ctl->in(0)->in(0)->in(0);
195 } else {
196 // Check if parent control has a single projection (this
197 // control is the only possible successor of the parent
198 // control). If so, we can try to move the node above the
199 // parent control.
200 int nb_ctl_proj = 0;
201 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
202 Node *p = parent_ctl->fast_out(i);
203 if (p->is_Proj() && p->is_CFG()) {
204 nb_ctl_proj++;
205 if (nb_ctl_proj > 1) {
206 break;
207 }
208 }
209 }
210
211 if (nb_ctl_proj > 1) {
212 break;
213 }
214 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
215 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
216 assert(idom(ctl) == parent_ctl, "strange");
217 next = idom(parent_ctl);
218 }
219 } else {
220 next = idom(ctl);
221 }
222 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
223 break;
224 }
225 ctl = next;
226 }
227
228 if (ctl != n->in(0)) {
229 _igvn.replace_input_of(n, 0, ctl);
230 _igvn.hash_insert(n);
231 }
232
233 return ctl;
234 }
235
236
237 //------------------------------set_early_ctrl---------------------------------
238 // Set earliest legal control
239 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
240 Node *early = get_early_ctrl(n);
241
242 // Record earliest legal location
243 set_ctrl(n, early);
244 }
245
246 //------------------------------set_subtree_ctrl-------------------------------
247 // set missing _ctrl entries on new nodes
248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
249 // Already set? Get out.
250 if( _nodes[n->_idx] ) return;
251 // Recursively set _nodes array to indicate where the Node goes
252 uint i;
253 for( i = 0; i < n->req(); ++i ) {
254 Node *m = n->in(i);
255 if( m && m != C->root() )
256 set_subtree_ctrl( m );
257 }
258
259 // Fixup self
260 set_early_ctrl( n );
261 }
262
263 // Create a skeleton strip mined outer loop: a Loop head before the
264 // inner strip mined loop, a safepoint and an exit condition guarded
265 // by an opaque node after the inner strip mined loop with a backedge
266 // to the loop head. The inner strip mined loop is left as it is. Only
267 // once loop optimizations are over, do we adjust the inner loop exit
268 // condition to limit its number of iterations, set the outer loop
269 // exit condition and add Phis to the outer loop head. Some loop
270 // optimizations that operate on the inner strip mined loop need to be
271 // aware of the outer strip mined loop: loop unswitching needs to
272 // clone the outer loop as well as the inner, unrolling needs to only
273 // clone the inner loop etc. No optimizations need to change the outer
274 // strip mined loop as it is only a skeleton.
275 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
276 IdealLoopTree* loop, float cl_prob, float le_fcnt,
277 Node*& entry_control, Node*& iffalse) {
278 Node* outer_test = _igvn.intcon(0);
279 set_ctrl(outer_test, C->root());
280 Node *orig = iffalse;
281 iffalse = iffalse->clone();
282 _igvn.register_new_node_with_optimizer(iffalse);
283 set_idom(iffalse, idom(orig), dom_depth(orig));
284
285 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
286 Node *outer_ift = new IfTrueNode (outer_le);
287 Node* outer_iff = orig;
288 _igvn.replace_input_of(outer_iff, 0, outer_le);
289
290 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
291 entry_control = outer_l;
292
293 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
294 IdealLoopTree* parent = loop->_parent;
295 IdealLoopTree* sibling = parent->_child;
296 if (sibling == loop) {
297 parent->_child = outer_ilt;
298 } else {
299 while (sibling->_next != loop) {
300 sibling = sibling->_next;
301 }
302 sibling->_next = outer_ilt;
303 }
304 outer_ilt->_next = loop->_next;
305 outer_ilt->_parent = parent;
306 outer_ilt->_child = loop;
307 outer_ilt->_nest = loop->_nest;
308 loop->_parent = outer_ilt;
309 loop->_next = NULL;
310 loop->_nest++;
311
312 set_loop(iffalse, outer_ilt);
313 register_control(outer_le, outer_ilt, iffalse);
314 register_control(outer_ift, outer_ilt, outer_le);
315 set_idom(outer_iff, outer_le, dom_depth(outer_le));
316 _igvn.register_new_node_with_optimizer(outer_l);
317 set_loop(outer_l, outer_ilt);
318 set_idom(outer_l, init_control, dom_depth(init_control)+1);
319
320 return outer_ilt;
321 }
322
323 void PhaseIdealLoop::insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol) {
324 Node* new_predicate_proj = create_new_if_for_predicate(limit_check_proj, NULL,
325 Deoptimization::Reason_loop_limit_check,
326 Op_If);
327 Node* iff = new_predicate_proj->in(0);
328 assert(iff->Opcode() == Op_If, "bad graph shape");
329 Node* conv = iff->in(1);
330 assert(conv->Opcode() == Op_Conv2B, "bad graph shape");
331 Node* opaq = conv->in(1);
332 assert(opaq->Opcode() == Op_Opaque1, "bad graph shape");
333 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
334 bol = _igvn.register_new_node_with_optimizer(bol);
335 set_subtree_ctrl(bol);
336 _igvn.replace_input_of(iff, 1, bol);
337
338 #ifndef PRODUCT
339 // report that the loop predication has been actually performed
340 // for this loop
341 if (TraceLoopLimitCheck) {
342 tty->print_cr("Counted Loop Limit Check generated:");
343 debug_only( bol->dump(2); )
344 }
345 #endif
346 }
347
348 //------------------------------is_counted_loop--------------------------------
349 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) {
350 PhaseGVN *gvn = &_igvn;
351
352 // Counted loop head must be a good RegionNode with only 3 not NULL
353 // control input edges: Self, Entry, LoopBack.
354 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
355 return false;
356 }
357 Node *init_control = x->in(LoopNode::EntryControl);
358 Node *back_control = x->in(LoopNode::LoopBackControl);
359 if (init_control == NULL || back_control == NULL) // Partially dead
360 return false;
361 // Must also check for TOP when looking for a dead loop
362 if (init_control->is_top() || back_control->is_top())
363 return false;
364
365 // Allow funny placement of Safepoint
366 if (back_control->Opcode() == Op_SafePoint) {
367 if (LoopStripMiningIter != 0) {
368 // Leaving the safepoint on the backedge and creating a
369 // CountedLoop will confuse optimizations. We can't move the
370 // safepoint around because its jvm state wouldn't match a new
371 // location. Give up on that loop.
372 return false;
373 }
374 back_control = back_control->in(TypeFunc::Control);
375 }
376
377 // Controlling test for loop
378 Node *iftrue = back_control;
379 uint iftrue_op = iftrue->Opcode();
380 if (iftrue_op != Op_IfTrue &&
381 iftrue_op != Op_IfFalse)
382 // I have a weird back-control. Probably the loop-exit test is in
383 // the middle of the loop and I am looking at some trailing control-flow
384 // merge point. To fix this I would have to partially peel the loop.
385 return false; // Obscure back-control
386
387 // Get boolean guarding loop-back test
388 Node *iff = iftrue->in(0);
389 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
390 return false;
391 BoolNode *test = iff->in(1)->as_Bool();
392 BoolTest::mask bt = test->_test._test;
393 float cl_prob = iff->as_If()->_prob;
394 if (iftrue_op == Op_IfFalse) {
395 bt = BoolTest(bt).negate();
396 cl_prob = 1.0 - cl_prob;
397 }
398 // Get backedge compare
399 Node *cmp = test->in(1);
400 int cmp_op = cmp->Opcode();
401 if (cmp_op != Op_CmpI)
402 return false; // Avoid pointer & float compares
403
404 // Find the trip-counter increment & limit. Limit must be loop invariant.
405 Node *incr = cmp->in(1);
406 Node *limit = cmp->in(2);
407
408 // ---------
409 // need 'loop()' test to tell if limit is loop invariant
410 // ---------
411
412 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
413 Node *tmp = incr; // Then reverse order into the CmpI
414 incr = limit;
415 limit = tmp;
416 bt = BoolTest(bt).commute(); // And commute the exit test
417 }
418 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
419 return false;
420 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
421 return false;
422
423 Node* phi_incr = NULL;
424 // Trip-counter increment must be commutative & associative.
425 if (incr->Opcode() == Op_CastII) {
426 incr = incr->in(1);
427 }
428 if (incr->is_Phi()) {
429 if (incr->as_Phi()->region() != x || incr->req() != 3)
430 return false; // Not simple trip counter expression
431 phi_incr = incr;
432 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
433 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
434 return false;
435 }
436
437 Node* trunc1 = NULL;
438 Node* trunc2 = NULL;
439 const TypeInt* iv_trunc_t = NULL;
440 Node* orig_incr = incr;
441 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
442 return false; // Funny increment opcode
443 }
444 assert(incr->Opcode() == Op_AddI, "wrong increment code");
445
446 const TypeInt* limit_t = gvn->type(limit)->is_int();
447 if (trunc1 != NULL) {
448 // When there is a truncation, we must be sure that after the truncation
449 // the trip counter will end up higher than the limit, otherwise we are looking
450 // at an endless loop. Can happen with range checks.
451
452 // Example:
453 // int i = 0;
454 // while (true)
455 // sum + = array[i];
456 // i++;
457 // i = i && 0x7fff;
458 // }
459 //
460 // If the array is shorter than 0x8000 this exits through a AIOOB
461 // - Counted loop transformation is ok
462 // If the array is longer then this is an endless loop
463 // - No transformation can be done.
464
465 const TypeInt* incr_t = gvn->type(orig_incr)->is_int();
466 if (limit_t->_hi > incr_t->_hi) {
467 // if the limit can have a higher value than the increment (before the phi)
468 return false;
469 }
470 }
471
472 // Get merge point
473 Node *xphi = incr->in(1);
474 Node *stride = incr->in(2);
475 if (!stride->is_Con()) { // Oops, swap these
476 if (!xphi->is_Con()) // Is the other guy a constant?
477 return false; // Nope, unknown stride, bail out
478 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
479 xphi = stride;
480 stride = tmp;
481 }
482 if (xphi->Opcode() == Op_CastII) {
483 xphi = xphi->in(1);
484 }
485 // Stride must be constant
486 int stride_con = stride->get_int();
487 if (stride_con == 0)
488 return false; // missed some peephole opt
489
490 if (!xphi->is_Phi())
491 return false; // Too much math on the trip counter
492 if (phi_incr != NULL && phi_incr != xphi)
493 return false;
494 PhiNode *phi = xphi->as_Phi();
495
496 // Phi must be of loop header; backedge must wrap to increment
497 if (phi->region() != x)
498 return false;
499 if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
500 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
501 return false;
502 }
503 Node *init_trip = phi->in(LoopNode::EntryControl);
504
505 // If iv trunc type is smaller than int, check for possible wrap.
506 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
507 assert(trunc1 != NULL, "must have found some truncation");
508
509 // Get a better type for the phi (filtered thru if's)
510 const TypeInt* phi_ft = filtered_type(phi);
511
512 // Can iv take on a value that will wrap?
513 //
514 // Ensure iv's limit is not within "stride" of the wrap value.
515 //
516 // Example for "short" type
517 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
518 // If the stride is +10, then the last value of the induction
519 // variable before the increment (phi_ft->_hi) must be
520 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
521 // ensure no truncation occurs after the increment.
522
523 if (stride_con > 0) {
524 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
525 iv_trunc_t->_lo > phi_ft->_lo) {
526 return false; // truncation may occur
527 }
528 } else if (stride_con < 0) {
529 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
530 iv_trunc_t->_hi < phi_ft->_hi) {
531 return false; // truncation may occur
532 }
533 }
534 // No possibility of wrap so truncation can be discarded
535 // Promote iv type to Int
536 } else {
537 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
538 }
539
540 // If the condition is inverted and we will be rolling
541 // through MININT to MAXINT, then bail out.
542 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
543 // Odd stride
544 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
545 // Count down loop rolls through MAXINT
546 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
547 // Count up loop rolls through MININT
548 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
549 return false; // Bail out
550 }
551
552 const TypeInt* init_t = gvn->type(init_trip)->is_int();
553
554 if (stride_con > 0) {
555 jlong init_p = (jlong)init_t->_lo + stride_con;
556 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
557 return false; // cyclic loop or this loop trips only once
558 } else {
559 jlong init_p = (jlong)init_t->_hi + stride_con;
560 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
561 return false; // cyclic loop or this loop trips only once
562 }
563
564 if (phi_incr != NULL && bt != BoolTest::ne) {
565 // check if there is a possiblity of IV overflowing after the first increment
566 if (stride_con > 0) {
567 if (init_t->_hi > max_jint - stride_con) {
568 return false;
569 }
570 } else {
571 if (init_t->_lo < min_jint - stride_con) {
572 return false;
573 }
574 }
575 }
576
577 // =================================================
578 // ---- SUCCESS! Found A Trip-Counted Loop! -----
579 //
580 assert(x->Opcode() == Op_Loop, "regular loops only");
581 C->print_method(PHASE_BEFORE_CLOOPS, 3);
582
583 Node *hook = new Node(6);
584
585 // ===================================================
586 // Generate loop limit check to avoid integer overflow
587 // in cases like next (cyclic loops):
588 //
589 // for (i=0; i <= max_jint; i++) {}
590 // for (i=0; i < max_jint; i+=2) {}
591 //
592 //
593 // Limit check predicate depends on the loop test:
594 //
595 // for(;i != limit; i++) --> limit <= (max_jint)
596 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
597 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
598 //
599
600 // Check if limit is excluded to do more precise int overflow check.
601 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
602 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
603
604 // If compare points directly to the phi we need to adjust
605 // the compare so that it points to the incr. Limit have
606 // to be adjusted to keep trip count the same and the
607 // adjusted limit should be checked for int overflow.
608 if (phi_incr != NULL) {
609 stride_m += stride_con;
610 }
611
612 if (limit->is_Con()) {
613 int limit_con = limit->get_int();
614 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
615 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
616 // Bailout: it could be integer overflow.
617 return false;
618 }
619 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
620 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
621 // Limit's type may satisfy the condition, for example,
622 // when it is an array length.
623 } else {
624 // Generate loop's limit check.
625 // Loop limit check predicate should be near the loop.
626 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
627 if (!limit_check_proj) {
628 // The limit check predicate is not generated if this method trapped here before.
629 #ifdef ASSERT
630 if (TraceLoopLimitCheck) {
631 tty->print("missing loop limit check:");
632 loop->dump_head();
633 x->dump(1);
634 }
635 #endif
636 return false;
637 }
638
639 IfNode* check_iff = limit_check_proj->in(0)->as_If();
640
641 if (!is_dominator(get_ctrl(limit), check_iff->in(0))) {
642 return false;
643 }
644
645 Node* cmp_limit;
646 Node* bol;
647
648 if (stride_con > 0) {
649 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
650 bol = new BoolNode(cmp_limit, BoolTest::le);
651 } else {
652 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
653 bol = new BoolNode(cmp_limit, BoolTest::ge);
654 }
655
656 insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
657 }
658
659 // Now we need to canonicalize loop condition.
660 if (bt == BoolTest::ne) {
661 assert(stride_con == 1 || stride_con == -1, "simple increment only");
662 if (stride_con > 0 && init_t->_hi < limit_t->_lo) {
663 // 'ne' can be replaced with 'lt' only when init < limit.
664 bt = BoolTest::lt;
665 } else if (stride_con < 0 && init_t->_lo > limit_t->_hi) {
666 // 'ne' can be replaced with 'gt' only when init > limit.
667 bt = BoolTest::gt;
668 } else {
669 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
670 if (!limit_check_proj) {
671 // The limit check predicate is not generated if this method trapped here before.
672 #ifdef ASSERT
673 if (TraceLoopLimitCheck) {
674 tty->print("missing loop limit check:");
675 loop->dump_head();
676 x->dump(1);
677 }
678 #endif
679 return false;
680 }
681 IfNode* check_iff = limit_check_proj->in(0)->as_If();
682
683 if (!is_dominator(get_ctrl(limit), check_iff->in(0)) ||
684 !is_dominator(get_ctrl(init_trip), check_iff->in(0))) {
685 return false;
686 }
687
688 Node* cmp_limit;
689 Node* bol;
690
691 if (stride_con > 0) {
692 cmp_limit = new CmpINode(init_trip, limit);
693 bol = new BoolNode(cmp_limit, BoolTest::lt);
694 } else {
695 cmp_limit = new CmpINode(init_trip, limit);
696 bol = new BoolNode(cmp_limit, BoolTest::gt);
697 }
698
699 insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
700
701 if (stride_con > 0) {
702 // 'ne' can be replaced with 'lt' only when init < limit.
703 bt = BoolTest::lt;
704 } else if (stride_con < 0) {
705 // 'ne' can be replaced with 'gt' only when init > limit.
706 bt = BoolTest::gt;
707 }
708 }
709 }
710
711 if (phi_incr != NULL) {
712 // If compare points directly to the phi we need to adjust
713 // the compare so that it points to the incr. Limit have
714 // to be adjusted to keep trip count the same and we
715 // should avoid int overflow.
716 //
717 // i = init; do {} while(i++ < limit);
718 // is converted to
719 // i = init; do {} while(++i < limit+1);
720 //
721 limit = gvn->transform(new AddINode(limit, stride));
722 }
723
724 if (incl_limit) {
725 // The limit check guaranties that 'limit <= (max_jint - stride)' so
726 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
727 //
728 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
729 limit = gvn->transform(new AddINode(limit, one));
730 if (bt == BoolTest::le)
731 bt = BoolTest::lt;
732 else if (bt == BoolTest::ge)
733 bt = BoolTest::gt;
734 else
735 ShouldNotReachHere();
736 }
737 set_subtree_ctrl( limit );
738
739 if (LoopStripMiningIter == 0) {
740 // Check for SafePoint on backedge and remove
741 Node *sfpt = x->in(LoopNode::LoopBackControl);
742 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
743 lazy_replace( sfpt, iftrue );
744 if (loop->_safepts != NULL) {
745 loop->_safepts->yank(sfpt);
746 }
747 loop->_tail = iftrue;
748 }
749 }
750
751 // Build a canonical trip test.
752 // Clone code, as old values may be in use.
753 incr = incr->clone();
754 incr->set_req(1,phi);
755 incr->set_req(2,stride);
756 incr = _igvn.register_new_node_with_optimizer(incr);
757 set_early_ctrl( incr );
758 _igvn.rehash_node_delayed(phi);
759 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
760
761 // If phi type is more restrictive than Int, raise to
762 // Int to prevent (almost) infinite recursion in igvn
763 // which can only handle integer types for constants or minint..maxint.
764 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
765 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
766 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
767 nphi = _igvn.register_new_node_with_optimizer(nphi);
768 set_ctrl(nphi, get_ctrl(phi));
769 _igvn.replace_node(phi, nphi);
770 phi = nphi->as_Phi();
771 }
772 cmp = cmp->clone();
773 cmp->set_req(1,incr);
774 cmp->set_req(2,limit);
775 cmp = _igvn.register_new_node_with_optimizer(cmp);
776 set_ctrl(cmp, iff->in(0));
777
778 test = test->clone()->as_Bool();
779 (*(BoolTest*)&test->_test)._test = bt;
780 test->set_req(1,cmp);
781 _igvn.register_new_node_with_optimizer(test);
782 set_ctrl(test, iff->in(0));
783
784 // Replace the old IfNode with a new LoopEndNode
785 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
786 IfNode *le = lex->as_If();
787 uint dd = dom_depth(iff);
788 set_idom(le, le->in(0), dd); // Update dominance for loop exit
789 set_loop(le, loop);
790
791 // Get the loop-exit control
792 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
793
794 // Need to swap loop-exit and loop-back control?
795 if (iftrue_op == Op_IfFalse) {
796 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
797 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
798
799 loop->_tail = back_control = ift2;
800 set_loop(ift2, loop);
801 set_loop(iff2, get_loop(iffalse));
802
803 // Lazy update of 'get_ctrl' mechanism.
804 lazy_replace(iffalse, iff2);
805 lazy_replace(iftrue, ift2);
806
807 // Swap names
808 iffalse = iff2;
809 iftrue = ift2;
810 } else {
811 _igvn.rehash_node_delayed(iffalse);
812 _igvn.rehash_node_delayed(iftrue);
813 iffalse->set_req_X( 0, le, &_igvn );
814 iftrue ->set_req_X( 0, le, &_igvn );
815 }
816
817 set_idom(iftrue, le, dd+1);
818 set_idom(iffalse, le, dd+1);
819 assert(iff->outcnt() == 0, "should be dead now");
820 lazy_replace( iff, le ); // fix 'get_ctrl'
821
822 Node *sfpt2 = le->in(0);
823
824 Node* entry_control = init_control;
825 bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL &&
826 sfpt2->Opcode() == Op_SafePoint && !loop->_has_call;
827 IdealLoopTree* outer_ilt = NULL;
828 if (strip_mine_loop) {
829 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
830 cl_prob, le->_fcnt, entry_control,
831 iffalse);
832 }
833
834 // Now setup a new CountedLoopNode to replace the existing LoopNode
835 CountedLoopNode *l = new CountedLoopNode(entry_control, back_control);
836 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
837 // The following assert is approximately true, and defines the intention
838 // of can_be_counted_loop. It fails, however, because phase->type
839 // is not yet initialized for this loop and its parts.
840 //assert(l->can_be_counted_loop(this), "sanity");
841 _igvn.register_new_node_with_optimizer(l);
842 set_loop(l, loop);
843 loop->_head = l;
844 // Fix all data nodes placed at the old loop head.
845 // Uses the lazy-update mechanism of 'get_ctrl'.
846 lazy_replace( x, l );
847 set_idom(l, entry_control, dom_depth(entry_control) + 1);
848
849 if (LoopStripMiningIter == 0 || strip_mine_loop) {
850 // Check for immediately preceding SafePoint and remove
851 if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) {
852 if (strip_mine_loop) {
853 Node* outer_le = outer_ilt->_tail->in(0);
854 Node* sfpt = sfpt2->clone();
855 sfpt->set_req(0, iffalse);
856 outer_le->set_req(0, sfpt);
857 register_control(sfpt, outer_ilt, iffalse);
858 set_idom(outer_le, sfpt, dom_depth(sfpt));
859 }
860 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
861 if (loop->_safepts != NULL) {
862 loop->_safepts->yank(sfpt2);
863 }
864 }
865 }
866
867 // Free up intermediate goo
868 _igvn.remove_dead_node(hook);
869
870 #ifdef ASSERT
871 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
872 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
873 #endif
874 #ifndef PRODUCT
875 if (TraceLoopOpts) {
876 tty->print("Counted ");
877 loop->dump_head();
878 }
879 #endif
880
881 C->print_method(PHASE_AFTER_CLOOPS, 3);
882
883 // Capture bounds of the loop in the induction variable Phi before
884 // subsequent transformation (iteration splitting) obscures the
885 // bounds
886 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
887
888 if (strip_mine_loop) {
889 l->mark_strip_mined();
890 l->verify_strip_mined(1);
891 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
892 loop = outer_ilt;
893 }
894
895 return true;
896 }
897
898 //----------------------exact_limit-------------------------------------------
899 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
900 assert(loop->_head->is_CountedLoop(), "");
901 CountedLoopNode *cl = loop->_head->as_CountedLoop();
902 assert(cl->is_valid_counted_loop(), "");
903
904 if (ABS(cl->stride_con()) == 1 ||
905 cl->limit()->Opcode() == Op_LoopLimit) {
906 // Old code has exact limit (it could be incorrect in case of int overflow).
907 // Loop limit is exact with stride == 1. And loop may already have exact limit.
908 return cl->limit();
909 }
910 Node *limit = NULL;
911 #ifdef ASSERT
912 BoolTest::mask bt = cl->loopexit()->test_trip();
913 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
914 #endif
915 if (cl->has_exact_trip_count()) {
916 // Simple case: loop has constant boundaries.
917 // Use jlongs to avoid integer overflow.
918 int stride_con = cl->stride_con();
919 jlong init_con = cl->init_trip()->get_int();
920 jlong limit_con = cl->limit()->get_int();
921 julong trip_cnt = cl->trip_count();
922 jlong final_con = init_con + trip_cnt*stride_con;
923 int final_int = (int)final_con;
924 // The final value should be in integer range since the loop
925 // is counted and the limit was checked for overflow.
926 assert(final_con == (jlong)final_int, "final value should be integer");
927 limit = _igvn.intcon(final_int);
928 } else {
929 // Create new LoopLimit node to get exact limit (final iv value).
930 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
931 register_new_node(limit, cl->in(LoopNode::EntryControl));
932 }
933 assert(limit != NULL, "sanity");
934 return limit;
935 }
936
937 //------------------------------Ideal------------------------------------------
938 // Return a node which is more "ideal" than the current node.
939 // Attempt to convert into a counted-loop.
940 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
941 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
942 phase->C->set_major_progress();
943 }
944 return RegionNode::Ideal(phase, can_reshape);
945 }
946
947 #ifdef ASSERT
948 void LoopNode::verify_strip_mined(int expect_skeleton) const {
949 if (!is_valid_counted_loop()) {
950 return; // Skip malformed counted loop
951 }
952 const OuterStripMinedLoopNode* outer = NULL;
953 const CountedLoopNode* inner = NULL;
954 if (is_strip_mined()) {
955 assert(is_CountedLoop(), "no Loop should be marked strip mined");
956 inner = as_CountedLoop();
957 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
958 } else if (is_OuterStripMinedLoop()) {
959 outer = this->as_OuterStripMinedLoop();
960 inner = outer->unique_ctrl_out()->as_CountedLoop();
961 assert(inner->is_valid_counted_loop(), "OuterStripMinedLoop should have been removed");
962 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
963 }
964 if (inner != NULL || outer != NULL) {
965 assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
966 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
967 Node* outer_le = outer_tail->in(0);
968 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
969 Node* sfpt = outer_le->in(0);
970 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
971 Node* inner_out = sfpt->in(0);
972 if (inner_out->outcnt() != 1) {
973 ResourceMark rm;
974 Unique_Node_List wq;
975
976 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
977 Node* u = inner_out->fast_out(i);
978 if (u == sfpt) {
979 continue;
980 }
981 wq.clear();
982 wq.push(u);
983 bool found_sfpt = false;
984 for (uint next = 0; next < wq.size() && !found_sfpt; next++) {
985 Node* n = wq.at(next);
986 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) {
987 Node* u = n->fast_out(i);
988 if (u == sfpt) {
989 found_sfpt = true;
990 }
991 if (!u->is_CFG()) {
992 wq.push(u);
993 }
994 }
995 }
996 assert(found_sfpt, "no node in loop that's not input to safepoint");
997 }
998 }
999
1000 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
1001 assert(cle == inner->loopexit_or_null(), "mismatch");
1002 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
1003 if (has_skeleton) {
1004 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
1005 assert(outer->outcnt() == 2, "only phis");
1006 } else {
1007 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
1008 uint phis = 0;
1009 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
1010 Node* u = inner->fast_out(i);
1011 if (u->is_Phi()) {
1012 phis++;
1013 }
1014 }
1015 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
1016 Node* u = outer->fast_out(i);
1017 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
1018 }
1019 uint stores = 0;
1020 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
1021 Node* u = inner_out->fast_out(i);
1022 if (u->is_Store()) {
1023 stores++;
1024 }
1025 }
1026 assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
1027 }
1028 assert(sfpt->outcnt() == 1, "no data node");
1029 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
1030 }
1031 }
1032 #endif
1033
1034 //=============================================================================
1035 //------------------------------Ideal------------------------------------------
1036 // Return a node which is more "ideal" than the current node.
1037 // Attempt to convert into a counted-loop.
1038 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1039 return RegionNode::Ideal(phase, can_reshape);
1040 }
1041
1042 //------------------------------dump_spec--------------------------------------
1043 // Dump special per-node info
1044 #ifndef PRODUCT
1045 void CountedLoopNode::dump_spec(outputStream *st) const {
1046 LoopNode::dump_spec(st);
1047 if (stride_is_con()) {
1048 st->print("stride: %d ",stride_con());
1049 }
1050 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
1051 if (is_main_loop()) st->print("main of N%d", _idx);
1052 if (is_post_loop()) st->print("post of N%d", _main_idx);
1053 if (is_strip_mined()) st->print(" strip mined");
1054 }
1055 #endif
1056
1057 //=============================================================================
1058 int CountedLoopEndNode::stride_con() const {
1059 return stride()->bottom_type()->is_int()->get_con();
1060 }
1061
1062 //=============================================================================
1063 //------------------------------Value-----------------------------------------
1064 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
1065 const Type* init_t = phase->type(in(Init));
1066 const Type* limit_t = phase->type(in(Limit));
1067 const Type* stride_t = phase->type(in(Stride));
1068 // Either input is TOP ==> the result is TOP
1069 if (init_t == Type::TOP) return Type::TOP;
1070 if (limit_t == Type::TOP) return Type::TOP;
1071 if (stride_t == Type::TOP) return Type::TOP;
1072
1073 int stride_con = stride_t->is_int()->get_con();
1074 if (stride_con == 1)
1075 return NULL; // Identity
1076
1077 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
1078 // Use jlongs to avoid integer overflow.
1079 jlong init_con = init_t->is_int()->get_con();
1080 jlong limit_con = limit_t->is_int()->get_con();
1081 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
1082 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
1083 jlong final_con = init_con + stride_con*trip_count;
1084 int final_int = (int)final_con;
1085 // The final value should be in integer range since the loop
1086 // is counted and the limit was checked for overflow.
1087 assert(final_con == (jlong)final_int, "final value should be integer");
1088 return TypeInt::make(final_int);
1089 }
1090
1091 return bottom_type(); // TypeInt::INT
1092 }
1093
1094 //------------------------------Ideal------------------------------------------
1095 // Return a node which is more "ideal" than the current node.
1096 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1097 if (phase->type(in(Init)) == Type::TOP ||
1098 phase->type(in(Limit)) == Type::TOP ||
1099 phase->type(in(Stride)) == Type::TOP)
1100 return NULL; // Dead
1101
1102 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1103 if (stride_con == 1)
1104 return NULL; // Identity
1105
1106 if (in(Init)->is_Con() && in(Limit)->is_Con())
1107 return NULL; // Value
1108
1109 // Delay following optimizations until all loop optimizations
1110 // done to keep Ideal graph simple.
1111 if (!can_reshape || phase->C->major_progress())
1112 return NULL;
1113
1114 const TypeInt* init_t = phase->type(in(Init) )->is_int();
1115 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
1116 int stride_p;
1117 jlong lim, ini;
1118 julong max;
1119 if (stride_con > 0) {
1120 stride_p = stride_con;
1121 lim = limit_t->_hi;
1122 ini = init_t->_lo;
1123 max = (julong)max_jint;
1124 } else {
1125 stride_p = -stride_con;
1126 lim = init_t->_hi;
1127 ini = limit_t->_lo;
1128 max = (julong)min_jint;
1129 }
1130 julong range = lim - ini + stride_p;
1131 if (range <= max) {
1132 // Convert to integer expression if it is not overflow.
1133 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
1134 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
1135 Node *bias = phase->transform(new AddINode(range, stride_m));
1136 Node *trip = phase->transform(new DivINode(0, bias, in(Stride)));
1137 Node *span = phase->transform(new MulINode(trip, in(Stride)));
1138 return new AddINode(span, in(Init)); // exact limit
1139 }
1140
1141 if (is_power_of_2(stride_p) || // divisor is 2^n
1142 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
1143 // Convert to long expression to avoid integer overflow
1144 // and let igvn optimizer convert this division.
1145 //
1146 Node* init = phase->transform( new ConvI2LNode(in(Init)));
1147 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
1148 Node* stride = phase->longcon(stride_con);
1149 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
1150
1151 Node *range = phase->transform(new SubLNode(limit, init));
1152 Node *bias = phase->transform(new AddLNode(range, stride_m));
1153 Node *span;
1154 if (stride_con > 0 && is_power_of_2(stride_p)) {
1155 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
1156 // and avoid generating rounding for division. Zero trip guard should
1157 // guarantee that init < limit but sometimes the guard is missing and
1158 // we can get situation when init > limit. Note, for the empty loop
1159 // optimization zero trip guard is generated explicitly which leaves
1160 // only RCE predicate where exact limit is used and the predicate
1161 // will simply fail forcing recompilation.
1162 Node* neg_stride = phase->longcon(-stride_con);
1163 span = phase->transform(new AndLNode(bias, neg_stride));
1164 } else {
1165 Node *trip = phase->transform(new DivLNode(0, bias, stride));
1166 span = phase->transform(new MulLNode(trip, stride));
1167 }
1168 // Convert back to int
1169 Node *span_int = phase->transform(new ConvL2INode(span));
1170 return new AddINode(span_int, in(Init)); // exact limit
1171 }
1172
1173 return NULL; // No progress
1174 }
1175
1176 //------------------------------Identity---------------------------------------
1177 // If stride == 1 return limit node.
1178 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1179 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1180 if (stride_con == 1 || stride_con == -1)
1181 return in(Limit);
1182 return this;
1183 }
1184
1185 //=============================================================================
1186 //----------------------match_incr_with_optional_truncation--------------------
1187 // Match increment with optional truncation:
1188 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1189 // Return NULL for failure. Success returns the increment node.
1190 Node* CountedLoopNode::match_incr_with_optional_truncation(
1191 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1192 // Quick cutouts:
1193 if (expr == NULL || expr->req() != 3) return NULL;
1194
1195 Node *t1 = NULL;
1196 Node *t2 = NULL;
1197 const TypeInt* trunc_t = TypeInt::INT;
1198 Node* n1 = expr;
1199 int n1op = n1->Opcode();
1200
1201 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1202 if (n1op == Op_AndI &&
1203 n1->in(2)->is_Con() &&
1204 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1205 // %%% This check should match any mask of 2**K-1.
1206 t1 = n1;
1207 n1 = t1->in(1);
1208 n1op = n1->Opcode();
1209 trunc_t = TypeInt::CHAR;
1210 } else if (n1op == Op_RShiftI &&
1211 n1->in(1) != NULL &&
1212 n1->in(1)->Opcode() == Op_LShiftI &&
1213 n1->in(2) == n1->in(1)->in(2) &&
1214 n1->in(2)->is_Con()) {
1215 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1216 // %%% This check should match any shift in [1..31].
1217 if (shift == 16 || shift == 8) {
1218 t1 = n1;
1219 t2 = t1->in(1);
1220 n1 = t2->in(1);
1221 n1op = n1->Opcode();
1222 if (shift == 16) {
1223 trunc_t = TypeInt::SHORT;
1224 } else if (shift == 8) {
1225 trunc_t = TypeInt::BYTE;
1226 }
1227 }
1228 }
1229
1230 // If (maybe after stripping) it is an AddI, we won:
1231 if (n1op == Op_AddI) {
1232 *trunc1 = t1;
1233 *trunc2 = t2;
1234 *trunc_type = trunc_t;
1235 return n1;
1236 }
1237
1238 // failed
1239 return NULL;
1240 }
1241
1242 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
1243 if (is_strip_mined()) {
1244 verify_strip_mined(expect_skeleton);
1245 return in(EntryControl)->as_Loop();
1246 }
1247 return this;
1248 }
1249
1250 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
1251 assert(is_strip_mined(), "not a strip mined loop");
1252 Node* c = in(EntryControl);
1253 if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) {
1254 return NULL;
1255 }
1256 return c->as_OuterStripMinedLoop();
1257 }
1258
1259 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
1260 Node* c = in(LoopBackControl);
1261 if (c == NULL || c->is_top()) {
1262 return NULL;
1263 }
1264 return c->as_IfTrue();
1265 }
1266
1267 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
1268 LoopNode* l = outer_loop();
1269 if (l == NULL) {
1270 return NULL;
1271 }
1272 return l->outer_loop_tail();
1273 }
1274
1275 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
1276 IfTrueNode* proj = outer_loop_tail();
1277 if (proj == NULL) {
1278 return NULL;
1279 }
1280 Node* c = proj->in(0);
1281 if (c == NULL || c->is_top() || c->outcnt() != 2) {
1282 return NULL;
1283 }
1284 return c->as_OuterStripMinedLoopEnd();
1285 }
1286
1287 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
1288 LoopNode* l = outer_loop();
1289 if (l == NULL) {
1290 return NULL;
1291 }
1292 return l->outer_loop_end();
1293 }
1294
1295 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
1296 IfNode* le = outer_loop_end();
1297 if (le == NULL) {
1298 return NULL;
1299 }
1300 Node* c = le->proj_out_or_null(false);
1301 if (c == NULL) {
1302 return NULL;
1303 }
1304 return c->as_IfFalse();
1305 }
1306
1307 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
1308 LoopNode* l = outer_loop();
1309 if (l == NULL) {
1310 return NULL;
1311 }
1312 return l->outer_loop_exit();
1313 }
1314
1315 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
1316 IfNode* le = outer_loop_end();
1317 if (le == NULL) {
1318 return NULL;
1319 }
1320 Node* c = le->in(0);
1321 if (c == NULL || c->is_top()) {
1322 return NULL;
1323 }
1324 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
1325 return c->as_SafePoint();
1326 }
1327
1328 SafePointNode* CountedLoopNode::outer_safepoint() const {
1329 LoopNode* l = outer_loop();
1330 if (l == NULL) {
1331 return NULL;
1332 }
1333 return l->outer_safepoint();
1334 }
1335
1336 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
1337 while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
1338 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
1339 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
1340 ctrl = ctrl->in(0)->in(0);
1341 }
1342
1343 return ctrl;
1344 }
1345
1346 Node* CountedLoopNode::skip_predicates() {
1347 if (is_main_loop()) {
1348 Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
1349
1350 return skip_predicates_from_entry(ctrl);
1351 }
1352 return in(LoopNode::EntryControl);
1353 }
1354
1355 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
1356 // Look for the outer & inner strip mined loop, reduce number of
1357 // iterations of the inner loop, set exit condition of outer loop,
1358 // construct required phi nodes for outer loop.
1359 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
1360 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
1361 Node* inner_iv_phi = inner_cl->phi();
1362 if (inner_iv_phi == NULL) {
1363 IfNode* outer_le = outer_loop_end();
1364 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1365 igvn->replace_node(outer_le, iff);
1366 inner_cl->clear_strip_mined();
1367 return;
1368 }
1369 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
1370
1371 int stride = inner_cl->stride_con();
1372 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride);
1373 int scaled_iters = (int)scaled_iters_long;
1374 int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride);
1375 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
1376 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
1377 assert(iter_estimate > 0, "broken");
1378 if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) {
1379 // Remove outer loop and safepoint (too few iterations)
1380 Node* outer_sfpt = outer_safepoint();
1381 Node* outer_out = outer_loop_exit();
1382 igvn->replace_node(outer_out, outer_sfpt->in(0));
1383 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
1384 inner_cl->clear_strip_mined();
1385 return;
1386 }
1387 if (iter_estimate <= scaled_iters_long) {
1388 // We would only go through one iteration of
1389 // the outer loop: drop the outer loop but
1390 // keep the safepoint so we don't run for
1391 // too long without a safepoint
1392 IfNode* outer_le = outer_loop_end();
1393 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1394 igvn->replace_node(outer_le, iff);
1395 inner_cl->clear_strip_mined();
1396 return;
1397 }
1398
1399 Node* cle_tail = inner_cle->proj_out(true);
1400 ResourceMark rm;
1401 Node_List old_new;
1402 if (cle_tail->outcnt() > 1) {
1403 // Look for nodes on backedge of inner loop and clone them
1404 Unique_Node_List backedge_nodes;
1405 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
1406 Node* u = cle_tail->fast_out(i);
1407 if (u != inner_cl) {
1408 assert(!u->is_CFG(), "control flow on the backedge?");
1409 backedge_nodes.push(u);
1410 }
1411 }
1412 uint last = igvn->C->unique();
1413 for (uint next = 0; next < backedge_nodes.size(); next++) {
1414 Node* n = backedge_nodes.at(next);
1415 old_new.map(n->_idx, n->clone());
1416 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1417 Node* u = n->fast_out(i);
1418 assert(!u->is_CFG(), "broken");
1419 if (u->_idx >= last) {
1420 continue;
1421 }
1422 if (!u->is_Phi()) {
1423 backedge_nodes.push(u);
1424 } else {
1425 assert(u->in(0) == inner_cl, "strange phi on the backedge");
1426 }
1427 }
1428 }
1429 // Put the clones on the outer loop backedge
1430 Node* le_tail = outer_loop_tail();
1431 for (uint next = 0; next < backedge_nodes.size(); next++) {
1432 Node *n = old_new[backedge_nodes.at(next)->_idx];
1433 for (uint i = 1; i < n->req(); i++) {
1434 if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
1435 n->set_req(i, old_new[n->in(i)->_idx]);
1436 }
1437 }
1438 if (n->in(0) != NULL && n->in(0) == cle_tail) {
1439 n->set_req(0, le_tail);
1440 }
1441 igvn->register_new_node_with_optimizer(n);
1442 }
1443 }
1444
1445 Node* iv_phi = NULL;
1446 // Make a clone of each phi in the inner loop
1447 // for the outer loop
1448 for (uint i = 0; i < inner_cl->outcnt(); i++) {
1449 Node* u = inner_cl->raw_out(i);
1450 if (u->is_Phi()) {
1451 assert(u->in(0) == inner_cl, "inconsistent");
1452 Node* phi = u->clone();
1453 phi->set_req(0, this);
1454 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
1455 if (be != NULL) {
1456 phi->set_req(LoopNode::LoopBackControl, be);
1457 }
1458 phi = igvn->transform(phi);
1459 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
1460 if (u == inner_iv_phi) {
1461 iv_phi = phi;
1462 }
1463 }
1464 }
1465 Node* cle_out = inner_cle->proj_out(false);
1466 if (cle_out->outcnt() > 1) {
1467 // Look for chains of stores that were sunk
1468 // out of the inner loop and are in the outer loop
1469 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
1470 Node* u = cle_out->fast_out(i);
1471 if (u->is_Store()) {
1472 Node* first = u;
1473 for(;;) {
1474 Node* next = first->in(MemNode::Memory);
1475 if (!next->is_Store() || next->in(0) != cle_out) {
1476 break;
1477 }
1478 first = next;
1479 }
1480 Node* last = u;
1481 for(;;) {
1482 Node* next = NULL;
1483 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
1484 Node* uu = last->fast_out(j);
1485 if (uu->is_Store() && uu->in(0) == cle_out) {
1486 assert(next == NULL, "only one in the outer loop");
1487 next = uu;
1488 }
1489 }
1490 if (next == NULL) {
1491 break;
1492 }
1493 last = next;
1494 }
1495 Node* phi = NULL;
1496 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1497 Node* uu = fast_out(j);
1498 if (uu->is_Phi()) {
1499 Node* be = uu->in(LoopNode::LoopBackControl);
1500 if (be->is_Store() && old_new[be->_idx] != NULL) {
1501 assert(false, "store on the backedge + sunk stores: unsupported");
1502 // drop outer loop
1503 IfNode* outer_le = outer_loop_end();
1504 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1505 igvn->replace_node(outer_le, iff);
1506 inner_cl->clear_strip_mined();
1507 return;
1508 }
1509 if (be == last || be == first->in(MemNode::Memory)) {
1510 assert(phi == NULL, "only one phi");
1511 phi = uu;
1512 }
1513 }
1514 }
1515 #ifdef ASSERT
1516 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1517 Node* uu = fast_out(j);
1518 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) {
1519 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
1520 assert(phi == uu, "what's that phi?");
1521 } else if (uu->adr_type() == TypePtr::BOTTOM) {
1522 Node* n = uu->in(LoopNode::LoopBackControl);
1523 uint limit = igvn->C->live_nodes();
1524 uint i = 0;
1525 while (n != uu) {
1526 i++;
1527 assert(i < limit, "infinite loop");
1528 if (n->is_Proj()) {
1529 n = n->in(0);
1530 } else if (n->is_SafePoint() || n->is_MemBar()) {
1531 n = n->in(TypeFunc::Memory);
1532 } else if (n->is_Phi()) {
1533 n = n->in(1);
1534 } else if (n->is_MergeMem()) {
1535 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
1536 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
1537 n = n->in(MemNode::Memory);
1538 } else {
1539 n->dump();
1540 ShouldNotReachHere();
1541 }
1542 }
1543 }
1544 }
1545 }
1546 #endif
1547 if (phi == NULL) {
1548 // If the an entire chains was sunk, the
1549 // inner loop has no phi for that memory
1550 // slice, create one for the outer loop
1551 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY,
1552 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
1553 phi->set_req(LoopNode::LoopBackControl, last);
1554 phi = igvn->transform(phi);
1555 igvn->replace_input_of(first, MemNode::Memory, phi);
1556 } else {
1557 // Or fix the outer loop fix to include
1558 // that chain of stores.
1559 Node* be = phi->in(LoopNode::LoopBackControl);
1560 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported");
1561 if (be == first->in(MemNode::Memory)) {
1562 if (be == phi->in(LoopNode::LoopBackControl)) {
1563 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
1564 } else {
1565 igvn->replace_input_of(be, MemNode::Memory, last);
1566 }
1567 } else {
1568 #ifdef ASSERT
1569 if (be == phi->in(LoopNode::LoopBackControl)) {
1570 assert(phi->in(LoopNode::LoopBackControl) == last, "");
1571 } else {
1572 assert(be->in(MemNode::Memory) == last, "");
1573 }
1574 #endif
1575 }
1576 }
1577 }
1578 }
1579 }
1580
1581 if (iv_phi != NULL) {
1582 // Now adjust the inner loop's exit condition
1583 Node* limit = inner_cl->limit();
1584 Node* sub = NULL;
1585 if (stride > 0) {
1586 sub = igvn->transform(new SubINode(limit, iv_phi));
1587 } else {
1588 sub = igvn->transform(new SubINode(iv_phi, limit));
1589 }
1590 Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters)));
1591 Node* new_limit = NULL;
1592 if (stride > 0) {
1593 new_limit = igvn->transform(new AddINode(min, iv_phi));
1594 } else {
1595 new_limit = igvn->transform(new SubINode(iv_phi, min));
1596 }
1597 Node* inner_cmp = inner_cle->cmp_node();
1598 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
1599 Node* outer_bol = inner_bol;
1600 // cmp node for inner loop may be shared
1601 inner_cmp = inner_cmp->clone();
1602 inner_cmp->set_req(2, new_limit);
1603 inner_bol = inner_bol->clone();
1604 inner_bol->set_req(1, igvn->transform(inner_cmp));
1605 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
1606 // Set the outer loop's exit condition too
1607 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
1608 } else {
1609 assert(false, "should be able to adjust outer loop");
1610 IfNode* outer_le = outer_loop_end();
1611 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1612 igvn->replace_node(outer_le, iff);
1613 inner_cl->clear_strip_mined();
1614 }
1615 }
1616
1617 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
1618 if (!in(0)) return Type::TOP;
1619 if (phase->type(in(0)) == Type::TOP)
1620 return Type::TOP;
1621
1622 return TypeTuple::IFBOTH;
1623 }
1624
1625 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1626 if (remove_dead_region(phase, can_reshape)) return this;
1627
1628 return NULL;
1629 }
1630
1631 //------------------------------filtered_type--------------------------------
1632 // Return a type based on condition control flow
1633 // A successful return will be a type that is restricted due
1634 // to a series of dominating if-tests, such as:
1635 // if (i < 10) {
1636 // if (i > 0) {
1637 // here: "i" type is [1..10)
1638 // }
1639 // }
1640 // or a control flow merge
1641 // if (i < 10) {
1642 // do {
1643 // phi( , ) -- at top of loop type is [min_int..10)
1644 // i = ?
1645 // } while ( i < 10)
1646 //
1647 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1648 assert(n && n->bottom_type()->is_int(), "must be int");
1649 const TypeInt* filtered_t = NULL;
1650 if (!n->is_Phi()) {
1651 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1652 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1653
1654 } else {
1655 Node* phi = n->as_Phi();
1656 Node* region = phi->in(0);
1657 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1658 if (region && region != C->top()) {
1659 for (uint i = 1; i < phi->req(); i++) {
1660 Node* val = phi->in(i);
1661 Node* use_c = region->in(i);
1662 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1663 if (val_t != NULL) {
1664 if (filtered_t == NULL) {
1665 filtered_t = val_t;
1666 } else {
1667 filtered_t = filtered_t->meet(val_t)->is_int();
1668 }
1669 }
1670 }
1671 }
1672 }
1673 const TypeInt* n_t = _igvn.type(n)->is_int();
1674 if (filtered_t != NULL) {
1675 n_t = n_t->join(filtered_t)->is_int();
1676 }
1677 return n_t;
1678 }
1679
1680
1681 //------------------------------filtered_type_from_dominators--------------------------------
1682 // Return a possibly more restrictive type for val based on condition control flow of dominators
1683 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1684 if (val->is_Con()) {
1685 return val->bottom_type()->is_int();
1686 }
1687 uint if_limit = 10; // Max number of dominating if's visited
1688 const TypeInt* rtn_t = NULL;
1689
1690 if (use_ctrl && use_ctrl != C->top()) {
1691 Node* val_ctrl = get_ctrl(val);
1692 uint val_dom_depth = dom_depth(val_ctrl);
1693 Node* pred = use_ctrl;
1694 uint if_cnt = 0;
1695 while (if_cnt < if_limit) {
1696 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1697 if_cnt++;
1698 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1699 if (if_t != NULL) {
1700 if (rtn_t == NULL) {
1701 rtn_t = if_t;
1702 } else {
1703 rtn_t = rtn_t->join(if_t)->is_int();
1704 }
1705 }
1706 }
1707 pred = idom(pred);
1708 if (pred == NULL || pred == C->top()) {
1709 break;
1710 }
1711 // Stop if going beyond definition block of val
1712 if (dom_depth(pred) < val_dom_depth) {
1713 break;
1714 }
1715 }
1716 }
1717 return rtn_t;
1718 }
1719
1720
1721 //------------------------------dump_spec--------------------------------------
1722 // Dump special per-node info
1723 #ifndef PRODUCT
1724 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1725 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1726 BoolTest bt( test_trip()); // Added this for g++.
1727
1728 st->print("[");
1729 bt.dump_on(st);
1730 st->print("]");
1731 }
1732 st->print(" ");
1733 IfNode::dump_spec(st);
1734 }
1735 #endif
1736
1737 //=============================================================================
1738 //------------------------------is_member--------------------------------------
1739 // Is 'l' a member of 'this'?
1740 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1741 while( l->_nest > _nest ) l = l->_parent;
1742 return l == this;
1743 }
1744
1745 //------------------------------set_nest---------------------------------------
1746 // Set loop tree nesting depth. Accumulate _has_call bits.
1747 int IdealLoopTree::set_nest( uint depth ) {
1748 _nest = depth;
1749 int bits = _has_call;
1750 if( _child ) bits |= _child->set_nest(depth+1);
1751 if( bits ) _has_call = 1;
1752 if( _next ) bits |= _next ->set_nest(depth );
1753 return bits;
1754 }
1755
1756 //------------------------------split_fall_in----------------------------------
1757 // Split out multiple fall-in edges from the loop header. Move them to a
1758 // private RegionNode before the loop. This becomes the loop landing pad.
1759 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1760 PhaseIterGVN &igvn = phase->_igvn;
1761 uint i;
1762
1763 // Make a new RegionNode to be the landing pad.
1764 Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1765 phase->set_loop(landing_pad,_parent);
1766 // Gather all the fall-in control paths into the landing pad
1767 uint icnt = fall_in_cnt;
1768 uint oreq = _head->req();
1769 for( i = oreq-1; i>0; i-- )
1770 if( !phase->is_member( this, _head->in(i) ) )
1771 landing_pad->set_req(icnt--,_head->in(i));
1772
1773 // Peel off PhiNode edges as well
1774 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1775 Node *oj = _head->fast_out(j);
1776 if( oj->is_Phi() ) {
1777 PhiNode* old_phi = oj->as_Phi();
1778 assert( old_phi->region() == _head, "" );
1779 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1780 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1781 uint icnt = fall_in_cnt;
1782 for( i = oreq-1; i>0; i-- ) {
1783 if( !phase->is_member( this, _head->in(i) ) ) {
1784 p->init_req(icnt--, old_phi->in(i));
1785 // Go ahead and clean out old edges from old phi
1786 old_phi->del_req(i);
1787 }
1788 }
1789 // Search for CSE's here, because ZKM.jar does a lot of
1790 // loop hackery and we need to be a little incremental
1791 // with the CSE to avoid O(N^2) node blow-up.
1792 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1793 if( p2 ) { // Found CSE
1794 p->destruct(); // Recover useless new node
1795 p = p2; // Use old node
1796 } else {
1797 igvn.register_new_node_with_optimizer(p, old_phi);
1798 }
1799 // Make old Phi refer to new Phi.
1800 old_phi->add_req(p);
1801 // Check for the special case of making the old phi useless and
1802 // disappear it. In JavaGrande I have a case where this useless
1803 // Phi is the loop limit and prevents recognizing a CountedLoop
1804 // which in turn prevents removing an empty loop.
1805 Node *id_old_phi = igvn.apply_identity(old_phi);
1806 if( id_old_phi != old_phi ) { // Found a simple identity?
1807 // Note that I cannot call 'replace_node' here, because
1808 // that will yank the edge from old_phi to the Region and
1809 // I'm mid-iteration over the Region's uses.
1810 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1811 Node* use = old_phi->last_out(i);
1812 igvn.rehash_node_delayed(use);
1813 uint uses_found = 0;
1814 for (uint j = 0; j < use->len(); j++) {
1815 if (use->in(j) == old_phi) {
1816 if (j < use->req()) use->set_req (j, id_old_phi);
1817 else use->set_prec(j, id_old_phi);
1818 uses_found++;
1819 }
1820 }
1821 i -= uses_found; // we deleted 1 or more copies of this edge
1822 }
1823 }
1824 igvn._worklist.push(old_phi);
1825 }
1826 }
1827 // Finally clean out the fall-in edges from the RegionNode
1828 for( i = oreq-1; i>0; i-- ) {
1829 if( !phase->is_member( this, _head->in(i) ) ) {
1830 _head->del_req(i);
1831 }
1832 }
1833 igvn.rehash_node_delayed(_head);
1834 // Transform landing pad
1835 igvn.register_new_node_with_optimizer(landing_pad, _head);
1836 // Insert landing pad into the header
1837 _head->add_req(landing_pad);
1838 }
1839
1840 //------------------------------split_outer_loop-------------------------------
1841 // Split out the outermost loop from this shared header.
1842 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1843 PhaseIterGVN &igvn = phase->_igvn;
1844
1845 // Find index of outermost loop; it should also be my tail.
1846 uint outer_idx = 1;
1847 while( _head->in(outer_idx) != _tail ) outer_idx++;
1848
1849 // Make a LoopNode for the outermost loop.
1850 Node *ctl = _head->in(LoopNode::EntryControl);
1851 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1852 outer = igvn.register_new_node_with_optimizer(outer, _head);
1853 phase->set_created_loop_node();
1854
1855 // Outermost loop falls into '_head' loop
1856 _head->set_req(LoopNode::EntryControl, outer);
1857 _head->del_req(outer_idx);
1858 // Split all the Phis up between '_head' loop and 'outer' loop.
1859 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1860 Node *out = _head->fast_out(j);
1861 if( out->is_Phi() ) {
1862 PhiNode *old_phi = out->as_Phi();
1863 assert( old_phi->region() == _head, "" );
1864 Node *phi = PhiNode::make_blank(outer, old_phi);
1865 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1866 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1867 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1868 // Make old Phi point to new Phi on the fall-in path
1869 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1870 old_phi->del_req(outer_idx);
1871 }
1872 }
1873
1874 // Use the new loop head instead of the old shared one
1875 _head = outer;
1876 phase->set_loop(_head, this);
1877 }
1878
1879 //------------------------------fix_parent-------------------------------------
1880 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1881 loop->_parent = parent;
1882 if( loop->_child ) fix_parent( loop->_child, loop );
1883 if( loop->_next ) fix_parent( loop->_next , parent );
1884 }
1885
1886 //------------------------------estimate_path_freq-----------------------------
1887 static float estimate_path_freq( Node *n ) {
1888 // Try to extract some path frequency info
1889 IfNode *iff;
1890 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1891 uint nop = n->Opcode();
1892 if( nop == Op_SafePoint ) { // Skip any safepoint
1893 n = n->in(0);
1894 continue;
1895 }
1896 if( nop == Op_CatchProj ) { // Get count from a prior call
1897 // Assume call does not always throw exceptions: means the call-site
1898 // count is also the frequency of the fall-through path.
1899 assert( n->is_CatchProj(), "" );
1900 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1901 return 0.0f; // Assume call exception path is rare
1902 Node *call = n->in(0)->in(0)->in(0);
1903 assert( call->is_Call(), "expect a call here" );
1904 const JVMState *jvms = ((CallNode*)call)->jvms();
1905 ciMethodData* methodData = jvms->method()->method_data();
1906 if (!methodData->is_mature()) return 0.0f; // No call-site data
1907 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1908 if ((data == NULL) || !data->is_CounterData()) {
1909 // no call profile available, try call's control input
1910 n = n->in(0);
1911 continue;
1912 }
1913 return data->as_CounterData()->count()/FreqCountInvocations;
1914 }
1915 // See if there's a gating IF test
1916 Node *n_c = n->in(0);
1917 if( !n_c->is_If() ) break; // No estimate available
1918 iff = n_c->as_If();
1919 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1920 // Compute how much count comes on this path
1921 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1922 // Have no count info. Skip dull uncommon-trap like branches.
1923 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1924 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1925 break;
1926 // Skip through never-taken branch; look for a real loop exit.
1927 n = iff->in(0);
1928 }
1929 return 0.0f; // No estimate available
1930 }
1931
1932 //------------------------------merge_many_backedges---------------------------
1933 // Merge all the backedges from the shared header into a private Region.
1934 // Feed that region as the one backedge to this loop.
1935 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1936 uint i;
1937
1938 // Scan for the top 2 hottest backedges
1939 float hotcnt = 0.0f;
1940 float warmcnt = 0.0f;
1941 uint hot_idx = 0;
1942 // Loop starts at 2 because slot 1 is the fall-in path
1943 for( i = 2; i < _head->req(); i++ ) {
1944 float cnt = estimate_path_freq(_head->in(i));
1945 if( cnt > hotcnt ) { // Grab hottest path
1946 warmcnt = hotcnt;
1947 hotcnt = cnt;
1948 hot_idx = i;
1949 } else if( cnt > warmcnt ) { // And 2nd hottest path
1950 warmcnt = cnt;
1951 }
1952 }
1953
1954 // See if the hottest backedge is worthy of being an inner loop
1955 // by being much hotter than the next hottest backedge.
1956 if( hotcnt <= 0.0001 ||
1957 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1958
1959 // Peel out the backedges into a private merge point; peel
1960 // them all except optionally hot_idx.
1961 PhaseIterGVN &igvn = phase->_igvn;
1962
1963 Node *hot_tail = NULL;
1964 // Make a Region for the merge point
1965 Node *r = new RegionNode(1);
1966 for( i = 2; i < _head->req(); i++ ) {
1967 if( i != hot_idx )
1968 r->add_req( _head->in(i) );
1969 else hot_tail = _head->in(i);
1970 }
1971 igvn.register_new_node_with_optimizer(r, _head);
1972 // Plug region into end of loop _head, followed by hot_tail
1973 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1974 igvn.replace_input_of(_head, 2, r);
1975 if( hot_idx ) _head->add_req(hot_tail);
1976
1977 // Split all the Phis up between '_head' loop and the Region 'r'
1978 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1979 Node *out = _head->fast_out(j);
1980 if( out->is_Phi() ) {
1981 PhiNode* n = out->as_Phi();
1982 igvn.hash_delete(n); // Delete from hash before hacking edges
1983 Node *hot_phi = NULL;
1984 Node *phi = new PhiNode(r, n->type(), n->adr_type());
1985 // Check all inputs for the ones to peel out
1986 uint j = 1;
1987 for( uint i = 2; i < n->req(); i++ ) {
1988 if( i != hot_idx )
1989 phi->set_req( j++, n->in(i) );
1990 else hot_phi = n->in(i);
1991 }
1992 // Register the phi but do not transform until whole place transforms
1993 igvn.register_new_node_with_optimizer(phi, n);
1994 // Add the merge phi to the old Phi
1995 while( n->req() > 3 ) n->del_req( n->req()-1 );
1996 igvn.replace_input_of(n, 2, phi);
1997 if( hot_idx ) n->add_req(hot_phi);
1998 }
1999 }
2000
2001
2002 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
2003 // of self loop tree. Turn self into a loop headed by _head and with
2004 // tail being the new merge point.
2005 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
2006 phase->set_loop(_tail,ilt); // Adjust tail
2007 _tail = r; // Self's tail is new merge point
2008 phase->set_loop(r,this);
2009 ilt->_child = _child; // New guy has my children
2010 _child = ilt; // Self has new guy as only child
2011 ilt->_parent = this; // new guy has self for parent
2012 ilt->_nest = _nest; // Same nesting depth (for now)
2013
2014 // Starting with 'ilt', look for child loop trees using the same shared
2015 // header. Flatten these out; they will no longer be loops in the end.
2016 IdealLoopTree **pilt = &_child;
2017 while( ilt ) {
2018 if( ilt->_head == _head ) {
2019 uint i;
2020 for( i = 2; i < _head->req(); i++ )
2021 if( _head->in(i) == ilt->_tail )
2022 break; // Still a loop
2023 if( i == _head->req() ) { // No longer a loop
2024 // Flatten ilt. Hang ilt's "_next" list from the end of
2025 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
2026 IdealLoopTree **cp = &ilt->_child;
2027 while( *cp ) cp = &(*cp)->_next; // Find end of child list
2028 *cp = ilt->_next; // Hang next list at end of child list
2029 *pilt = ilt->_child; // Move child up to replace ilt
2030 ilt->_head = NULL; // Flag as a loop UNIONED into parent
2031 ilt = ilt->_child; // Repeat using new ilt
2032 continue; // do not advance over ilt->_child
2033 }
2034 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
2035 phase->set_loop(_head,ilt);
2036 }
2037 pilt = &ilt->_child; // Advance to next
2038 ilt = *pilt;
2039 }
2040
2041 if( _child ) fix_parent( _child, this );
2042 }
2043
2044 //------------------------------beautify_loops---------------------------------
2045 // Split shared headers and insert loop landing pads.
2046 // Insert a LoopNode to replace the RegionNode.
2047 // Return TRUE if loop tree is structurally changed.
2048 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
2049 bool result = false;
2050 // Cache parts in locals for easy
2051 PhaseIterGVN &igvn = phase->_igvn;
2052
2053 igvn.hash_delete(_head); // Yank from hash before hacking edges
2054
2055 // Check for multiple fall-in paths. Peel off a landing pad if need be.
2056 int fall_in_cnt = 0;
2057 for( uint i = 1; i < _head->req(); i++ )
2058 if( !phase->is_member( this, _head->in(i) ) )
2059 fall_in_cnt++;
2060 assert( fall_in_cnt, "at least 1 fall-in path" );
2061 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
2062 split_fall_in( phase, fall_in_cnt );
2063
2064 // Swap inputs to the _head and all Phis to move the fall-in edge to
2065 // the left.
2066 fall_in_cnt = 1;
2067 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
2068 fall_in_cnt++;
2069 if( fall_in_cnt > 1 ) {
2070 // Since I am just swapping inputs I do not need to update def-use info
2071 Node *tmp = _head->in(1);
2072 igvn.rehash_node_delayed(_head);
2073 _head->set_req( 1, _head->in(fall_in_cnt) );
2074 _head->set_req( fall_in_cnt, tmp );
2075 // Swap also all Phis
2076 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
2077 Node* phi = _head->fast_out(i);
2078 if( phi->is_Phi() ) {
2079 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
2080 tmp = phi->in(1);
2081 phi->set_req( 1, phi->in(fall_in_cnt) );
2082 phi->set_req( fall_in_cnt, tmp );
2083 }
2084 }
2085 }
2086 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
2087 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
2088
2089 // If I am a shared header (multiple backedges), peel off the many
2090 // backedges into a private merge point and use the merge point as
2091 // the one true backedge.
2092 if( _head->req() > 3 ) {
2093 // Merge the many backedges into a single backedge but leave
2094 // the hottest backedge as separate edge for the following peel.
2095 merge_many_backedges( phase );
2096 result = true;
2097 }
2098
2099 // If I have one hot backedge, peel off myself loop.
2100 // I better be the outermost loop.
2101 if (_head->req() > 3 && !_irreducible) {
2102 split_outer_loop( phase );
2103 result = true;
2104
2105 } else if (!_head->is_Loop() && !_irreducible) {
2106 // Make a new LoopNode to replace the old loop head
2107 Node *l = new LoopNode( _head->in(1), _head->in(2) );
2108 l = igvn.register_new_node_with_optimizer(l, _head);
2109 phase->set_created_loop_node();
2110 // Go ahead and replace _head
2111 phase->_igvn.replace_node( _head, l );
2112 _head = l;
2113 phase->set_loop(_head, this);
2114 }
2115
2116 // Now recursively beautify nested loops
2117 if( _child ) result |= _child->beautify_loops( phase );
2118 if( _next ) result |= _next ->beautify_loops( phase );
2119 return result;
2120 }
2121
2122 //------------------------------allpaths_check_safepts----------------------------
2123 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
2124 // encountered. Helper for check_safepts.
2125 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
2126 assert(stack.size() == 0, "empty stack");
2127 stack.push(_tail);
2128 visited.clear();
2129 visited.set(_tail->_idx);
2130 while (stack.size() > 0) {
2131 Node* n = stack.pop();
2132 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2133 // Terminate this path
2134 } else if (n->Opcode() == Op_SafePoint) {
2135 if (_phase->get_loop(n) != this) {
2136 if (_required_safept == NULL) _required_safept = new Node_List();
2137 _required_safept->push(n); // save the one closest to the tail
2138 }
2139 // Terminate this path
2140 } else {
2141 uint start = n->is_Region() ? 1 : 0;
2142 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
2143 for (uint i = start; i < end; i++) {
2144 Node* in = n->in(i);
2145 assert(in->is_CFG(), "must be");
2146 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
2147 stack.push(in);
2148 }
2149 }
2150 }
2151 }
2152 }
2153
2154 //------------------------------check_safepts----------------------------
2155 // Given dominators, try to find loops with calls that must always be
2156 // executed (call dominates loop tail). These loops do not need non-call
2157 // safepoints (ncsfpt).
2158 //
2159 // A complication is that a safepoint in a inner loop may be needed
2160 // by an outer loop. In the following, the inner loop sees it has a
2161 // call (block 3) on every path from the head (block 2) to the
2162 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
2163 // in block 2, _but_ this leaves the outer loop without a safepoint.
2164 //
2165 // entry 0
2166 // |
2167 // v
2168 // outer 1,2 +->1
2169 // | |
2170 // | v
2171 // | 2<---+ ncsfpt in 2
2172 // |_/|\ |
2173 // | v |
2174 // inner 2,3 / 3 | call in 3
2175 // / | |
2176 // v +--+
2177 // exit 4
2178 //
2179 //
2180 // This method creates a list (_required_safept) of ncsfpt nodes that must
2181 // be protected is created for each loop. When a ncsfpt maybe deleted, it
2182 // is first looked for in the lists for the outer loops of the current loop.
2183 //
2184 // The insights into the problem:
2185 // A) counted loops are okay
2186 // B) innermost loops are okay (only an inner loop can delete
2187 // a ncsfpt needed by an outer loop)
2188 // C) a loop is immune from an inner loop deleting a safepoint
2189 // if the loop has a call on the idom-path
2190 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
2191 // idom-path that is not in a nested loop
2192 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
2193 // loop needs to be prevented from deletion by an inner loop
2194 //
2195 // There are two analyses:
2196 // 1) The first, and cheaper one, scans the loop body from
2197 // tail to head following the idom (immediate dominator)
2198 // chain, looking for the cases (C,D,E) above.
2199 // Since inner loops are scanned before outer loops, there is summary
2200 // information about inner loops. Inner loops can be skipped over
2201 // when the tail of an inner loop is encountered.
2202 //
2203 // 2) The second, invoked if the first fails to find a call or ncsfpt on
2204 // the idom path (which is rare), scans all predecessor control paths
2205 // from the tail to the head, terminating a path when a call or sfpt
2206 // is encountered, to find the ncsfpt's that are closest to the tail.
2207 //
2208 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
2209 // Bottom up traversal
2210 IdealLoopTree* ch = _child;
2211 if (_child) _child->check_safepts(visited, stack);
2212 if (_next) _next ->check_safepts(visited, stack);
2213
2214 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
2215 bool has_call = false; // call on dom-path
2216 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
2217 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
2218 // Scan the dom-path nodes from tail to head
2219 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
2220 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2221 has_call = true;
2222 _has_sfpt = 1; // Then no need for a safept!
2223 break;
2224 } else if (n->Opcode() == Op_SafePoint) {
2225 if (_phase->get_loop(n) == this) {
2226 has_local_ncsfpt = true;
2227 break;
2228 }
2229 if (nonlocal_ncsfpt == NULL) {
2230 nonlocal_ncsfpt = n; // save the one closest to the tail
2231 }
2232 } else {
2233 IdealLoopTree* nlpt = _phase->get_loop(n);
2234 if (this != nlpt) {
2235 // If at an inner loop tail, see if the inner loop has already
2236 // recorded seeing a call on the dom-path (and stop.) If not,
2237 // jump to the head of the inner loop.
2238 assert(is_member(nlpt), "nested loop");
2239 Node* tail = nlpt->_tail;
2240 if (tail->in(0)->is_If()) tail = tail->in(0);
2241 if (n == tail) {
2242 // If inner loop has call on dom-path, so does outer loop
2243 if (nlpt->_has_sfpt) {
2244 has_call = true;
2245 _has_sfpt = 1;
2246 break;
2247 }
2248 // Skip to head of inner loop
2249 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
2250 n = nlpt->_head;
2251 }
2252 }
2253 }
2254 }
2255 // Record safept's that this loop needs preserved when an
2256 // inner loop attempts to delete it's safepoints.
2257 if (_child != NULL && !has_call && !has_local_ncsfpt) {
2258 if (nonlocal_ncsfpt != NULL) {
2259 if (_required_safept == NULL) _required_safept = new Node_List();
2260 _required_safept->push(nonlocal_ncsfpt);
2261 } else {
2262 // Failed to find a suitable safept on the dom-path. Now use
2263 // an all paths walk from tail to head, looking for safepoints to preserve.
2264 allpaths_check_safepts(visited, stack);
2265 }
2266 }
2267 }
2268 }
2269
2270 //---------------------------is_deleteable_safept----------------------------
2271 // Is safept not required by an outer loop?
2272 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
2273 assert(sfpt->Opcode() == Op_SafePoint, "");
2274 IdealLoopTree* lp = get_loop(sfpt)->_parent;
2275 while (lp != NULL) {
2276 Node_List* sfpts = lp->_required_safept;
2277 if (sfpts != NULL) {
2278 for (uint i = 0; i < sfpts->size(); i++) {
2279 if (sfpt == sfpts->at(i))
2280 return false;
2281 }
2282 }
2283 lp = lp->_parent;
2284 }
2285 return true;
2286 }
2287
2288 //---------------------------replace_parallel_iv-------------------------------
2289 // Replace parallel induction variable (parallel to trip counter)
2290 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
2291 assert(loop->_head->is_CountedLoop(), "");
2292 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2293 if (!cl->is_valid_counted_loop())
2294 return; // skip malformed counted loop
2295 Node *incr = cl->incr();
2296 if (incr == NULL)
2297 return; // Dead loop?
2298 Node *init = cl->init_trip();
2299 Node *phi = cl->phi();
2300 int stride_con = cl->stride_con();
2301
2302 // Visit all children, looking for Phis
2303 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
2304 Node *out = cl->out(i);
2305 // Look for other phis (secondary IVs). Skip dead ones
2306 if (!out->is_Phi() || out == phi || !has_node(out))
2307 continue;
2308 PhiNode* phi2 = out->as_Phi();
2309 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
2310 // Look for induction variables of the form: X += constant
2311 if (phi2->region() != loop->_head ||
2312 incr2->req() != 3 ||
2313 incr2->in(1) != phi2 ||
2314 incr2 == incr ||
2315 incr2->Opcode() != Op_AddI ||
2316 !incr2->in(2)->is_Con())
2317 continue;
2318
2319 // Check for parallel induction variable (parallel to trip counter)
2320 // via an affine function. In particular, count-down loops with
2321 // count-up array indices are common. We only RCE references off
2322 // the trip-counter, so we need to convert all these to trip-counter
2323 // expressions.
2324 Node *init2 = phi2->in( LoopNode::EntryControl );
2325 int stride_con2 = incr2->in(2)->get_int();
2326
2327 // The ratio of the two strides cannot be represented as an int
2328 // if stride_con2 is min_int and stride_con is -1.
2329 if (stride_con2 == min_jint && stride_con == -1) {
2330 continue;
2331 }
2332
2333 // The general case here gets a little tricky. We want to find the
2334 // GCD of all possible parallel IV's and make a new IV using this
2335 // GCD for the loop. Then all possible IVs are simple multiples of
2336 // the GCD. In practice, this will cover very few extra loops.
2337 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
2338 // where +/-1 is the common case, but other integer multiples are
2339 // also easy to handle.
2340 int ratio_con = stride_con2/stride_con;
2341
2342 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
2343 #ifndef PRODUCT
2344 if (TraceLoopOpts) {
2345 tty->print("Parallel IV: %d ", phi2->_idx);
2346 loop->dump_head();
2347 }
2348 #endif
2349 // Convert to using the trip counter. The parallel induction
2350 // variable differs from the trip counter by a loop-invariant
2351 // amount, the difference between their respective initial values.
2352 // It is scaled by the 'ratio_con'.
2353 Node* ratio = _igvn.intcon(ratio_con);
2354 set_ctrl(ratio, C->root());
2355 Node* ratio_init = new MulINode(init, ratio);
2356 _igvn.register_new_node_with_optimizer(ratio_init, init);
2357 set_early_ctrl(ratio_init);
2358 Node* diff = new SubINode(init2, ratio_init);
2359 _igvn.register_new_node_with_optimizer(diff, init2);
2360 set_early_ctrl(diff);
2361 Node* ratio_idx = new MulINode(phi, ratio);
2362 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
2363 set_ctrl(ratio_idx, cl);
2364 Node* add = new AddINode(ratio_idx, diff);
2365 _igvn.register_new_node_with_optimizer(add);
2366 set_ctrl(add, cl);
2367 _igvn.replace_node( phi2, add );
2368 // Sometimes an induction variable is unused
2369 if (add->outcnt() == 0) {
2370 _igvn.remove_dead_node(add);
2371 }
2372 --i; // deleted this phi; rescan starting with next position
2373 continue;
2374 }
2375 }
2376 }
2377
2378 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
2379 Node* keep = NULL;
2380 if (keep_one) {
2381 // Look for a safepoint on the idom-path.
2382 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
2383 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
2384 keep = i;
2385 break; // Found one
2386 }
2387 }
2388 }
2389
2390 // Don't remove any safepoints if it is requested to keep a single safepoint and
2391 // no safepoint was found on idom-path. It is not safe to remove any safepoint
2392 // in this case since there's no safepoint dominating all paths in the loop body.
2393 bool prune = !keep_one || keep != NULL;
2394
2395 // Delete other safepoints in this loop.
2396 Node_List* sfpts = _safepts;
2397 if (prune && sfpts != NULL) {
2398 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
2399 for (uint i = 0; i < sfpts->size(); i++) {
2400 Node* n = sfpts->at(i);
2401 assert(phase->get_loop(n) == this, "");
2402 if (n != keep && phase->is_deleteable_safept(n)) {
2403 phase->lazy_replace(n, n->in(TypeFunc::Control));
2404 }
2405 }
2406 }
2407 }
2408
2409 //------------------------------counted_loop-----------------------------------
2410 // Convert to counted loops where possible
2411 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
2412
2413 // For grins, set the inner-loop flag here
2414 if (!_child) {
2415 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
2416 }
2417
2418 IdealLoopTree* loop = this;
2419 if (_head->is_CountedLoop() ||
2420 phase->is_counted_loop(_head, loop)) {
2421
2422 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) {
2423 // Indicate we do not need a safepoint here
2424 _has_sfpt = 1;
2425 }
2426
2427 // Remove safepoints
2428 bool keep_one_sfpt = !(_has_call || _has_sfpt);
2429 remove_safepoints(phase, keep_one_sfpt);
2430
2431 // Look for induction variables
2432 phase->replace_parallel_iv(this);
2433
2434 } else if (_parent != NULL && !_irreducible) {
2435 // Not a counted loop. Keep one safepoint.
2436 bool keep_one_sfpt = true;
2437 remove_safepoints(phase, keep_one_sfpt);
2438 }
2439
2440 // Recursively
2441 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
2442 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops");
2443 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
2444 if (loop->_next) loop->_next ->counted_loop(phase);
2445 }
2446
2447
2448 // The Estimated Loop Clone Size:
2449 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
2450 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
2451 // respectively, used to ensure that the node usage estimates made are on the
2452 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
2453 // possible additional/excessive nodes generated due to data and control flow
2454 // merging, for edges reaching outside the loop.
2455 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
2456
2457 precond(0 < factor && factor < 16);
2458
2459 uint const bc = 13;
2460 uint const cc = 17;
2461 uint const sz = _body.size() + (_body.size() + 7) / 8;
2462 uint estimate = factor * (sz + bc) + cc;
2463
2464 assert((estimate - cc) / factor == sz + bc, "overflow");
2465
2466 return estimate + est_loop_flow_merge_sz();
2467 }
2468
2469 // The Estimated Loop (full-) Unroll Size:
2470 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
2471 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
2472 // node usage estimates made are on the safe side, for the most part. This is
2473 // a "light" version of the loop clone size calculation (above), based on the
2474 // assumption that most of the loop-construct overhead will be unraveled when
2475 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
2476 // including an overflow check and returning UINT_MAX in case of an overflow.
2477 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
2478
2479 precond(factor > 0);
2480
2481 // Take into account that after unroll conjoined heads and tails will fold.
2482 uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
2483 uint const cc = 7;
2484 uint const sz = b0 + (b0 + 15) / 16;
2485 uint estimate = factor * sz + cc;
2486
2487 if ((estimate - cc) / factor != sz) {
2488 return UINT_MAX;
2489 }
2490
2491 return estimate + est_loop_flow_merge_sz();
2492 }
2493
2494 // Estimate the growth effect (in nodes) of merging control and data flow when
2495 // cloning a loop body, based on the amount of control and data flow reaching
2496 // outside of the (current) loop body.
2497 uint IdealLoopTree::est_loop_flow_merge_sz() const {
2498
2499 uint ctrl_edge_out_cnt = 0;
2500 uint data_edge_out_cnt = 0;
2501
2502 for (uint i = 0; i < _body.size(); i++) {
2503 Node* node = _body.at(i);
2504 uint outcnt = node->outcnt();
2505
2506 for (uint k = 0; k < outcnt; k++) {
2507 Node* out = node->raw_out(k);
2508
2509 if (out->is_CFG()) {
2510 if (!is_member(_phase->get_loop(out))) {
2511 ctrl_edge_out_cnt++;
2512 }
2513 } else {
2514 Node* ctrl = _phase->get_ctrl(out);
2515 assert(ctrl->is_CFG(), "must be");
2516 if (!is_member(_phase->get_loop(ctrl))) {
2517 data_edge_out_cnt++;
2518 }
2519 }
2520 }
2521 }
2522 // Use data and control count (x2.0) in estimate iff both are > 0. This is
2523 // a rather pessimistic estimate for the most part, in particular for some
2524 // complex loops, but still not enough to capture all loops.
2525 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
2526 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
2527 }
2528 return 0;
2529 }
2530
2531 #ifndef PRODUCT
2532 //------------------------------dump_head--------------------------------------
2533 // Dump 1 liner for loop header info
2534 void IdealLoopTree::dump_head() const {
2535 tty->sp(2 * _nest);
2536 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);
2537 if (_irreducible) tty->print(" IRREDUCIBLE");
2538 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
2539 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
2540 if (predicate != NULL ) {
2541 tty->print(" limit_check");
2542 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2543 }
2544 if (UseLoopPredicate) {
2545 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2546 if (entry != NULL) {
2547 tty->print(" predicated");
2548 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2549 }
2550 }
2551 if (UseProfiledLoopPredicate) {
2552 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
2553 if (entry != NULL) {
2554 tty->print(" profile_predicated");
2555 }
2556 }
2557 if (_head->is_CountedLoop()) {
2558 CountedLoopNode *cl = _head->as_CountedLoop();
2559 tty->print(" counted");
2560
2561 Node* init_n = cl->init_trip();
2562 if (init_n != NULL && init_n->is_Con())
2563 tty->print(" [%d,", cl->init_trip()->get_int());
2564 else
2565 tty->print(" [int,");
2566 Node* limit_n = cl->limit();
2567 if (limit_n != NULL && limit_n->is_Con())
2568 tty->print("%d),", cl->limit()->get_int());
2569 else
2570 tty->print("int),");
2571 int stride_con = cl->stride_con();
2572 if (stride_con > 0) tty->print("+");
2573 tty->print("%d", stride_con);
2574
2575 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
2576
2577 if (cl->is_pre_loop ()) tty->print(" pre" );
2578 if (cl->is_main_loop()) tty->print(" main");
2579 if (cl->is_post_loop()) tty->print(" post");
2580 if (cl->is_vectorized_loop()) tty->print(" vector");
2581 if (cl->range_checks_present()) tty->print(" rc ");
2582 if (cl->is_multiversioned()) tty->print(" multi ");
2583 }
2584 if (_has_call) tty->print(" has_call");
2585 if (_has_sfpt) tty->print(" has_sfpt");
2586 if (_rce_candidate) tty->print(" rce");
2587 if (_safepts != NULL && _safepts->size() > 0) {
2588 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
2589 }
2590 if (_required_safept != NULL && _required_safept->size() > 0) {
2591 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
2592 }
2593 if (Verbose) {
2594 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
2595 }
2596 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
2597 tty->print(" strip_mined");
2598 }
2599 tty->cr();
2600 }
2601
2602 //------------------------------dump-------------------------------------------
2603 // Dump loops by loop tree
2604 void IdealLoopTree::dump() const {
2605 dump_head();
2606 if (_child) _child->dump();
2607 if (_next) _next ->dump();
2608 }
2609
2610 #endif
2611
2612 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
2613 if (loop == root) {
2614 if (loop->_child != NULL) {
2615 log->begin_head("loop_tree");
2616 log->end_head();
2617 if( loop->_child ) log_loop_tree(root, loop->_child, log);
2618 log->tail("loop_tree");
2619 assert(loop->_next == NULL, "what?");
2620 }
2621 } else {
2622 Node* head = loop->_head;
2623 log->begin_head("loop");
2624 log->print(" idx='%d' ", head->_idx);
2625 if (loop->_irreducible) log->print("irreducible='1' ");
2626 if (head->is_Loop()) {
2627 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
2628 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
2629 }
2630 if (head->is_CountedLoop()) {
2631 CountedLoopNode* cl = head->as_CountedLoop();
2632 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
2633 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
2634 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
2635 }
2636 log->end_head();
2637 if( loop->_child ) log_loop_tree(root, loop->_child, log);
2638 log->tail("loop");
2639 if( loop->_next ) log_loop_tree(root, loop->_next, log);
2640 }
2641 }
2642
2643 //---------------------collect_potentially_useful_predicates-----------------------
2644 // Helper function to collect potentially useful predicates to prevent them from
2645 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2646 void PhaseIdealLoop::collect_potentially_useful_predicates(
2647 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2648 if (loop->_child) { // child
2649 collect_potentially_useful_predicates(loop->_child, useful_predicates);
2650 }
2651
2652 // self (only loops that we can apply loop predication may use their predicates)
2653 if (loop->_head->is_Loop() &&
2654 !loop->_irreducible &&
2655 !loop->tail()->is_top()) {
2656 LoopNode* lpn = loop->_head->as_Loop();
2657 Node* entry = lpn->in(LoopNode::EntryControl);
2658 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2659 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2660 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2661 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2662 entry = skip_loop_predicates(entry);
2663 }
2664 predicate_proj = find_predicate(entry); // Predicate
2665 if (predicate_proj != NULL ) {
2666 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2667 entry = skip_loop_predicates(entry);
2668 }
2669 if (UseProfiledLoopPredicate) {
2670 predicate_proj = find_predicate(entry); // Predicate
2671 if (predicate_proj != NULL ) {
2672 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2673 }
2674 }
2675 }
2676
2677 if (loop->_next) { // sibling
2678 collect_potentially_useful_predicates(loop->_next, useful_predicates);
2679 }
2680 }
2681
2682 //------------------------eliminate_useless_predicates-----------------------------
2683 // Eliminate all inserted predicates if they could not be used by loop predication.
2684 // Note: it will also eliminates loop limits check predicate since it also uses
2685 // Opaque1 node (see Parse::add_predicate()).
2686 void PhaseIdealLoop::eliminate_useless_predicates() {
2687 if (C->predicate_count() == 0)
2688 return; // no predicate left
2689
2690 Unique_Node_List useful_predicates; // to store useful predicates
2691 if (C->has_loops()) {
2692 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2693 }
2694
2695 for (int i = C->predicate_count(); i > 0; i--) {
2696 Node * n = C->predicate_opaque1_node(i-1);
2697 assert(n->Opcode() == Op_Opaque1, "must be");
2698 if (!useful_predicates.member(n)) { // not in the useful list
2699 _igvn.replace_node(n, n->in(1));
2700 }
2701 }
2702 }
2703
2704 //------------------------process_expensive_nodes-----------------------------
2705 // Expensive nodes have their control input set to prevent the GVN
2706 // from commoning them and as a result forcing the resulting node to
2707 // be in a more frequent path. Use CFG information here, to change the
2708 // control inputs so that some expensive nodes can be commoned while
2709 // not executed more frequently.
2710 bool PhaseIdealLoop::process_expensive_nodes() {
2711 assert(OptimizeExpensiveOps, "optimization off?");
2712
2713 // Sort nodes to bring similar nodes together
2714 C->sort_expensive_nodes();
2715
2716 bool progress = false;
2717
2718 for (int i = 0; i < C->expensive_count(); ) {
2719 Node* n = C->expensive_node(i);
2720 int start = i;
2721 // Find nodes similar to n
2722 i++;
2723 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2724 int end = i;
2725 // And compare them two by two
2726 for (int j = start; j < end; j++) {
2727 Node* n1 = C->expensive_node(j);
2728 if (is_node_unreachable(n1)) {
2729 continue;
2730 }
2731 for (int k = j+1; k < end; k++) {
2732 Node* n2 = C->expensive_node(k);
2733 if (is_node_unreachable(n2)) {
2734 continue;
2735 }
2736
2737 assert(n1 != n2, "should be pair of nodes");
2738
2739 Node* c1 = n1->in(0);
2740 Node* c2 = n2->in(0);
2741
2742 Node* parent_c1 = c1;
2743 Node* parent_c2 = c2;
2744
2745 // The call to get_early_ctrl_for_expensive() moves the
2746 // expensive nodes up but stops at loops that are in a if
2747 // branch. See whether we can exit the loop and move above the
2748 // If.
2749 if (c1->is_Loop()) {
2750 parent_c1 = c1->in(1);
2751 }
2752 if (c2->is_Loop()) {
2753 parent_c2 = c2->in(1);
2754 }
2755
2756 if (parent_c1 == parent_c2) {
2757 _igvn._worklist.push(n1);
2758 _igvn._worklist.push(n2);
2759 continue;
2760 }
2761
2762 // Look for identical expensive node up the dominator chain.
2763 if (is_dominator(c1, c2)) {
2764 c2 = c1;
2765 } else if (is_dominator(c2, c1)) {
2766 c1 = c2;
2767 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2768 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2769 // Both branches have the same expensive node so move it up
2770 // before the if.
2771 c1 = c2 = idom(parent_c1->in(0));
2772 }
2773 // Do the actual moves
2774 if (n1->in(0) != c1) {
2775 _igvn.hash_delete(n1);
2776 n1->set_req(0, c1);
2777 _igvn.hash_insert(n1);
2778 _igvn._worklist.push(n1);
2779 progress = true;
2780 }
2781 if (n2->in(0) != c2) {
2782 _igvn.hash_delete(n2);
2783 n2->set_req(0, c2);
2784 _igvn.hash_insert(n2);
2785 _igvn._worklist.push(n2);
2786 progress = true;
2787 }
2788 }
2789 }
2790 }
2791
2792 return progress;
2793 }
2794
2795
2796 //=============================================================================
2797 //----------------------------build_and_optimize-------------------------------
2798 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2799 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2800 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
2801 bool do_split_ifs = (mode == LoopOptsDefault);
2802 bool skip_loop_opts = (mode == LoopOptsNone);
2803
2804 int old_progress = C->major_progress();
2805 uint orig_worklist_size = _igvn._worklist.size();
2806
2807 // Reset major-progress flag for the driver's heuristics
2808 C->clear_major_progress();
2809
2810 #ifndef PRODUCT
2811 // Capture for later assert
2812 uint unique = C->unique();
2813 _loop_invokes++;
2814 _loop_work += unique;
2815 #endif
2816
2817 // True if the method has at least 1 irreducible loop
2818 _has_irreducible_loops = false;
2819
2820 _created_loop_node = false;
2821
2822 Arena *a = Thread::current()->resource_area();
2823 VectorSet visited(a);
2824 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2825 _nodes.map(C->unique(), NULL);
2826 memset(_nodes.adr(), 0, wordSize * C->unique());
2827
2828 // Pre-build the top-level outermost loop tree entry
2829 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2830 // Do not need a safepoint at the top level
2831 _ltree_root->_has_sfpt = 1;
2832
2833 // Initialize Dominators.
2834 // Checked in clone_loop_predicate() during beautify_loops().
2835 _idom_size = 0;
2836 _idom = NULL;
2837 _dom_depth = NULL;
2838 _dom_stk = NULL;
2839
2840 // Empty pre-order array
2841 allocate_preorders();
2842
2843 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2844 // IdealLoopTree entries. Data nodes are NOT walked.
2845 build_loop_tree();
2846 // Check for bailout, and return
2847 if (C->failing()) {
2848 return;
2849 }
2850
2851 // No loops after all
2852 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2853
2854 // There should always be an outer loop containing the Root and Return nodes.
2855 // If not, we have a degenerate empty program. Bail out in this case.
2856 if (!has_node(C->root())) {
2857 if (!_verify_only) {
2858 C->clear_major_progress();
2859 C->record_method_not_compilable("empty program detected during loop optimization");
2860 }
2861 return;
2862 }
2863
2864 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2865 // Nothing to do, so get out
2866 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only &&
2867 !bs->is_gc_specific_loop_opts_pass(mode);
2868 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2869 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(mode);
2870 if (stop_early && !do_expensive_nodes) {
2871 _igvn.optimize(); // Cleanup NeverBranches
2872 return;
2873 }
2874
2875 // Set loop nesting depth
2876 _ltree_root->set_nest( 0 );
2877
2878 // Split shared headers and insert loop landing pads.
2879 // Do not bother doing this on the Root loop of course.
2880 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2881 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2882 if( _ltree_root->_child->beautify_loops( this ) ) {
2883 // Re-build loop tree!
2884 _ltree_root->_child = NULL;
2885 _nodes.clear();
2886 reallocate_preorders();
2887 build_loop_tree();
2888 // Check for bailout, and return
2889 if (C->failing()) {
2890 return;
2891 }
2892 // Reset loop nesting depth
2893 _ltree_root->set_nest( 0 );
2894
2895 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2896 }
2897 }
2898
2899 // Build Dominators for elision of NULL checks & loop finding.
2900 // Since nodes do not have a slot for immediate dominator, make
2901 // a persistent side array for that info indexed on node->_idx.
2902 _idom_size = C->unique();
2903 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2904 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2905 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2906 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2907
2908 Dominators();
2909
2910 if (!_verify_only) {
2911 // As a side effect, Dominators removed any unreachable CFG paths
2912 // into RegionNodes. It doesn't do this test against Root, so
2913 // we do it here.
2914 for( uint i = 1; i < C->root()->req(); i++ ) {
2915 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2916 _igvn.delete_input_of(C->root(), i);
2917 i--; // Rerun same iteration on compressed edges
2918 }
2919 }
2920
2921 // Given dominators, try to find inner loops with calls that must
2922 // always be executed (call dominates loop tail). These loops do
2923 // not need a separate safepoint.
2924 Node_List cisstack(a);
2925 _ltree_root->check_safepts(visited, cisstack);
2926 }
2927
2928 // Walk the DATA nodes and place into loops. Find earliest control
2929 // node. For CFG nodes, the _nodes array starts out and remains
2930 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2931 // _nodes array holds the earliest legal controlling CFG node.
2932
2933 // Allocate stack with enough space to avoid frequent realloc
2934 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2935 Node_Stack nstack( a, stack_size );
2936
2937 visited.clear();
2938 Node_List worklist(a);
2939 // Don't need C->root() on worklist since
2940 // it will be processed among C->top() inputs
2941 worklist.push(C->top());
2942 visited.set(C->top()->_idx); // Set C->top() as visited now
2943 build_loop_early( visited, worklist, nstack );
2944
2945 // Given early legal placement, try finding counted loops. This placement
2946 // is good enough to discover most loop invariants.
2947 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
2948 _ltree_root->counted_loop( this );
2949 }
2950
2951 // Find latest loop placement. Find ideal loop placement.
2952 visited.clear();
2953 init_dom_lca_tags();
2954 // Need C->root() on worklist when processing outs
2955 worklist.push(C->root());
2956 NOT_PRODUCT( C->verify_graph_edges(); )
2957 worklist.push(C->top());
2958 build_loop_late( visited, worklist, nstack );
2959
2960 if (_verify_only) {
2961 C->restore_major_progress(old_progress);
2962 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2963 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2964 return;
2965 }
2966
2967 // clear out the dead code after build_loop_late
2968 while (_deadlist.size()) {
2969 _igvn.remove_globally_dead_node(_deadlist.pop());
2970 }
2971
2972 if (stop_early) {
2973 assert(do_expensive_nodes, "why are we here?");
2974 if (process_expensive_nodes()) {
2975 // If we made some progress when processing expensive nodes then
2976 // the IGVN may modify the graph in a way that will allow us to
2977 // make some more progress: we need to try processing expensive
2978 // nodes again.
2979 C->set_major_progress();
2980 }
2981 _igvn.optimize();
2982 return;
2983 }
2984
2985 // Some parser-inserted loop predicates could never be used by loop
2986 // predication or they were moved away from loop during some optimizations.
2987 // For example, peeling. Eliminate them before next loop optimizations.
2988 eliminate_useless_predicates();
2989
2990 #ifndef PRODUCT
2991 C->verify_graph_edges();
2992 if (_verify_me) { // Nested verify pass?
2993 // Check to see if the verify mode is broken
2994 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2995 return;
2996 }
2997 if (VerifyLoopOptimizations) verify();
2998 if (TraceLoopOpts && C->has_loops()) {
2999 _ltree_root->dump();
3000 }
3001 #endif
3002
3003 if (skip_loop_opts) {
3004 // restore major progress flag
3005 C->restore_major_progress(old_progress);
3006
3007 // Cleanup any modified bits
3008 _igvn.optimize();
3009
3010 if (C->log() != NULL) {
3011 log_loop_tree(_ltree_root, _ltree_root, C->log());
3012 }
3013 return;
3014 }
3015
3016 if (bs->optimize_loops(this, mode, visited, nstack, worklist)) {
3017 _igvn.optimize();
3018 if (C->log() != NULL) {
3019 log_loop_tree(_ltree_root, _ltree_root, C->log());
3020 }
3021 return;
3022 }
3023
3024 if (ReassociateInvariants) {
3025 // Reassociate invariants and prep for split_thru_phi
3026 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3027 IdealLoopTree* lpt = iter.current();
3028 bool is_counted = lpt->is_counted();
3029 if (!is_counted || !lpt->is_innermost()) continue;
3030
3031 // check for vectorized loops, any reassociation of invariants was already done
3032 if (is_counted && lpt->_head->as_CountedLoop()->is_unroll_only()) {
3033 continue;
3034 } else {
3035 AutoNodeBudget node_budget(this);
3036 lpt->reassociate_invariants(this);
3037 }
3038 // Because RCE opportunities can be masked by split_thru_phi,
3039 // look for RCE candidates and inhibit split_thru_phi
3040 // on just their loop-phi's for this pass of loop opts
3041 if (SplitIfBlocks && do_split_ifs) {
3042 AutoNodeBudget node_budget(this, AutoNodeBudget::NO_BUDGET_CHECK);
3043 if (lpt->policy_range_check(this)) {
3044 lpt->_rce_candidate = 1; // = true
3045 }
3046 }
3047 }
3048 }
3049
3050 // Check for aggressive application of split-if and other transforms
3051 // that require basic-block info (like cloning through Phi's)
3052 if( SplitIfBlocks && do_split_ifs ) {
3053 visited.clear();
3054 split_if_with_blocks( visited, nstack);
3055 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
3056 }
3057
3058 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
3059 C->set_major_progress();
3060 }
3061
3062 // Perform loop predication before iteration splitting
3063 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
3064 _ltree_root->_child->loop_predication(this);
3065 }
3066
3067 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
3068 if (do_intrinsify_fill()) {
3069 C->set_major_progress();
3070 }
3071 }
3072
3073 // Perform iteration-splitting on inner loops. Split iterations to avoid
3074 // range checks or one-shot null checks.
3075
3076 // If split-if's didn't hack the graph too bad (no CFG changes)
3077 // then do loop opts.
3078 if (C->has_loops() && !C->major_progress()) {
3079 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
3080 _ltree_root->_child->iteration_split( this, worklist );
3081 // No verify after peeling! GCM has hoisted code out of the loop.
3082 // After peeling, the hoisted code could sink inside the peeled area.
3083 // The peeling code does not try to recompute the best location for
3084 // all the code before the peeled area, so the verify pass will always
3085 // complain about it.
3086 }
3087 // Do verify graph edges in any case
3088 NOT_PRODUCT( C->verify_graph_edges(); );
3089
3090 if (!do_split_ifs) {
3091 // We saw major progress in Split-If to get here. We forced a
3092 // pass with unrolling and not split-if, however more split-if's
3093 // might make progress. If the unrolling didn't make progress
3094 // then the major-progress flag got cleared and we won't try
3095 // another round of Split-If. In particular the ever-common
3096 // instance-of/check-cast pattern requires at least 2 rounds of
3097 // Split-If to clear out.
3098 C->set_major_progress();
3099 }
3100
3101 // Repeat loop optimizations if new loops were seen
3102 if (created_loop_node()) {
3103 C->set_major_progress();
3104 }
3105
3106 // Keep loop predicates and perform optimizations with them
3107 // until no more loop optimizations could be done.
3108 // After that switch predicates off and do more loop optimizations.
3109 if (!C->major_progress() && (C->predicate_count() > 0)) {
3110 C->cleanup_loop_predicates(_igvn);
3111 if (TraceLoopOpts) {
3112 tty->print_cr("PredicatesOff");
3113 }
3114 C->set_major_progress();
3115 }
3116
3117 // Convert scalar to superword operations at the end of all loop opts.
3118 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
3119 // SuperWord transform
3120 SuperWord sw(this);
3121 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3122 IdealLoopTree* lpt = iter.current();
3123 if (lpt->is_counted()) {
3124 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
3125
3126 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
3127 // Check that the rce'd post loop is encountered first, multiversion after all
3128 // major main loop optimization are concluded
3129 if (!C->major_progress()) {
3130 IdealLoopTree *lpt_next = lpt->_next;
3131 if (lpt_next && lpt_next->is_counted()) {
3132 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
3133 has_range_checks(lpt_next);
3134 if (cl->is_post_loop() && cl->range_checks_present()) {
3135 if (!cl->is_multiversioned()) {
3136 if (multi_version_post_loops(lpt, lpt_next) == false) {
3137 // Cause the rce loop to be optimized away if we fail
3138 cl->mark_is_multiversioned();
3139 cl->set_slp_max_unroll(0);
3140 poison_rce_post_loop(lpt);
3141 }
3142 }
3143 }
3144 }
3145 sw.transform_loop(lpt, true);
3146 }
3147 } else if (cl->is_main_loop()) {
3148 sw.transform_loop(lpt, true);
3149 }
3150 }
3151 }
3152 }
3153
3154 // Cleanup any modified bits
3155 _igvn.optimize();
3156
3157 // disable assert until issue with split_flow_path is resolved (6742111)
3158 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
3159 // "shouldn't introduce irreducible loops");
3160
3161 if (C->log() != NULL) {
3162 log_loop_tree(_ltree_root, _ltree_root, C->log());
3163 }
3164 }
3165
3166 #ifndef PRODUCT
3167 //------------------------------print_statistics-------------------------------
3168 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
3169 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
3170 void PhaseIdealLoop::print_statistics() {
3171 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
3172 }
3173
3174 //------------------------------verify-----------------------------------------
3175 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
3176 static int fail; // debug only, so its multi-thread dont care
3177 void PhaseIdealLoop::verify() const {
3178 int old_progress = C->major_progress();
3179 ResourceMark rm;
3180 PhaseIdealLoop loop_verify( _igvn, this );
3181 VectorSet visited(Thread::current()->resource_area());
3182
3183 fail = 0;
3184 verify_compare( C->root(), &loop_verify, visited );
3185 assert( fail == 0, "verify loops failed" );
3186 // Verify loop structure is the same
3187 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
3188 // Reset major-progress. It was cleared by creating a verify version of
3189 // PhaseIdealLoop.
3190 C->restore_major_progress(old_progress);
3191 }
3192
3193 //------------------------------verify_compare---------------------------------
3194 // Make sure me and the given PhaseIdealLoop agree on key data structures
3195 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
3196 if( !n ) return;
3197 if( visited.test_set( n->_idx ) ) return;
3198 if( !_nodes[n->_idx] ) { // Unreachable
3199 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
3200 return;
3201 }
3202
3203 uint i;
3204 for( i = 0; i < n->req(); i++ )
3205 verify_compare( n->in(i), loop_verify, visited );
3206
3207 // Check the '_nodes' block/loop structure
3208 i = n->_idx;
3209 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
3210 if( _nodes[i] != loop_verify->_nodes[i] &&
3211 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
3212 tty->print("Mismatched control setting for: ");
3213 n->dump();
3214 if( fail++ > 10 ) return;
3215 Node *c = get_ctrl_no_update(n);
3216 tty->print("We have it as: ");
3217 if( c->in(0) ) c->dump();
3218 else tty->print_cr("N%d",c->_idx);
3219 tty->print("Verify thinks: ");
3220 if( loop_verify->has_ctrl(n) )
3221 loop_verify->get_ctrl_no_update(n)->dump();
3222 else
3223 loop_verify->get_loop_idx(n)->dump();
3224 tty->cr();
3225 }
3226 } else { // We have a loop
3227 IdealLoopTree *us = get_loop_idx(n);
3228 if( loop_verify->has_ctrl(n) ) {
3229 tty->print("Mismatched loop setting for: ");
3230 n->dump();
3231 if( fail++ > 10 ) return;
3232 tty->print("We have it as: ");
3233 us->dump();
3234 tty->print("Verify thinks: ");
3235 loop_verify->get_ctrl_no_update(n)->dump();
3236 tty->cr();
3237 } else if (!C->major_progress()) {
3238 // Loop selection can be messed up if we did a major progress
3239 // operation, like split-if. Do not verify in that case.
3240 IdealLoopTree *them = loop_verify->get_loop_idx(n);
3241 if( us->_head != them->_head || us->_tail != them->_tail ) {
3242 tty->print("Unequals loops for: ");
3243 n->dump();
3244 if( fail++ > 10 ) return;
3245 tty->print("We have it as: ");
3246 us->dump();
3247 tty->print("Verify thinks: ");
3248 them->dump();
3249 tty->cr();
3250 }
3251 }
3252 }
3253
3254 // Check for immediate dominators being equal
3255 if( i >= _idom_size ) {
3256 if( !n->is_CFG() ) return;
3257 tty->print("CFG Node with no idom: ");
3258 n->dump();
3259 return;
3260 }
3261 if( !n->is_CFG() ) return;
3262 if( n == C->root() ) return; // No IDOM here
3263
3264 assert(n->_idx == i, "sanity");
3265 Node *id = idom_no_update(n);
3266 if( id != loop_verify->idom_no_update(n) ) {
3267 tty->print("Unequals idoms for: ");
3268 n->dump();
3269 if( fail++ > 10 ) return;
3270 tty->print("We have it as: ");
3271 id->dump();
3272 tty->print("Verify thinks: ");
3273 loop_verify->idom_no_update(n)->dump();
3274 tty->cr();
3275 }
3276
3277 }
3278
3279 //------------------------------verify_tree------------------------------------
3280 // Verify that tree structures match. Because the CFG can change, siblings
3281 // within the loop tree can be reordered. We attempt to deal with that by
3282 // reordering the verify's loop tree if possible.
3283 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
3284 assert( _parent == parent, "Badly formed loop tree" );
3285
3286 // Siblings not in same order? Attempt to re-order.
3287 if( _head != loop->_head ) {
3288 // Find _next pointer to update
3289 IdealLoopTree **pp = &loop->_parent->_child;
3290 while( *pp != loop )
3291 pp = &((*pp)->_next);
3292 // Find proper sibling to be next
3293 IdealLoopTree **nn = &loop->_next;
3294 while( (*nn) && (*nn)->_head != _head )
3295 nn = &((*nn)->_next);
3296
3297 // Check for no match.
3298 if( !(*nn) ) {
3299 // Annoyingly, irreducible loops can pick different headers
3300 // after a major_progress operation, so the rest of the loop
3301 // tree cannot be matched.
3302 if (_irreducible && Compile::current()->major_progress()) return;
3303 assert( 0, "failed to match loop tree" );
3304 }
3305
3306 // Move (*nn) to (*pp)
3307 IdealLoopTree *hit = *nn;
3308 *nn = hit->_next;
3309 hit->_next = loop;
3310 *pp = loop;
3311 loop = hit;
3312 // Now try again to verify
3313 }
3314
3315 assert( _head == loop->_head , "mismatched loop head" );
3316 Node *tail = _tail; // Inline a non-updating version of
3317 while( !tail->in(0) ) // the 'tail()' call.
3318 tail = tail->in(1);
3319 assert( tail == loop->_tail, "mismatched loop tail" );
3320
3321 // Counted loops that are guarded should be able to find their guards
3322 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
3323 CountedLoopNode *cl = _head->as_CountedLoop();
3324 Node *init = cl->init_trip();
3325 Node *ctrl = cl->in(LoopNode::EntryControl);
3326 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
3327 Node *iff = ctrl->in(0);
3328 assert( iff->Opcode() == Op_If, "" );
3329 Node *bol = iff->in(1);
3330 assert( bol->Opcode() == Op_Bool, "" );
3331 Node *cmp = bol->in(1);
3332 assert( cmp->Opcode() == Op_CmpI, "" );
3333 Node *add = cmp->in(1);
3334 Node *opaq;
3335 if( add->Opcode() == Op_Opaque1 ) {
3336 opaq = add;
3337 } else {
3338 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
3339 assert( add == init, "" );
3340 opaq = cmp->in(2);
3341 }
3342 assert( opaq->Opcode() == Op_Opaque1, "" );
3343
3344 }
3345
3346 if (_child != NULL) _child->verify_tree(loop->_child, this);
3347 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
3348 // Innermost loops need to verify loop bodies,
3349 // but only if no 'major_progress'
3350 int fail = 0;
3351 if (!Compile::current()->major_progress() && _child == NULL) {
3352 for( uint i = 0; i < _body.size(); i++ ) {
3353 Node *n = _body.at(i);
3354 if (n->outcnt() == 0) continue; // Ignore dead
3355 uint j;
3356 for( j = 0; j < loop->_body.size(); j++ )
3357 if( loop->_body.at(j) == n )
3358 break;
3359 if( j == loop->_body.size() ) { // Not found in loop body
3360 // Last ditch effort to avoid assertion: Its possible that we
3361 // have some users (so outcnt not zero) but are still dead.
3362 // Try to find from root.
3363 if (Compile::current()->root()->find(n->_idx)) {
3364 fail++;
3365 tty->print("We have that verify does not: ");
3366 n->dump();
3367 }
3368 }
3369 }
3370 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
3371 Node *n = loop->_body.at(i2);
3372 if (n->outcnt() == 0) continue; // Ignore dead
3373 uint j;
3374 for( j = 0; j < _body.size(); j++ )
3375 if( _body.at(j) == n )
3376 break;
3377 if( j == _body.size() ) { // Not found in loop body
3378 // Last ditch effort to avoid assertion: Its possible that we
3379 // have some users (so outcnt not zero) but are still dead.
3380 // Try to find from root.
3381 if (Compile::current()->root()->find(n->_idx)) {
3382 fail++;
3383 tty->print("Verify has that we do not: ");
3384 n->dump();
3385 }
3386 }
3387 }
3388 assert( !fail, "loop body mismatch" );
3389 }
3390 }
3391
3392 #endif
3393
3394 //------------------------------set_idom---------------------------------------
3395 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
3396 uint idx = d->_idx;
3397 if (idx >= _idom_size) {
3398 uint newsize = next_power_of_2(idx);
3399 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
3400 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
3401 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
3402 _idom_size = newsize;
3403 }
3404 _idom[idx] = n;
3405 _dom_depth[idx] = dom_depth;
3406 }
3407
3408 //------------------------------recompute_dom_depth---------------------------------------
3409 // The dominator tree is constructed with only parent pointers.
3410 // This recomputes the depth in the tree by first tagging all
3411 // nodes as "no depth yet" marker. The next pass then runs up
3412 // the dom tree from each node marked "no depth yet", and computes
3413 // the depth on the way back down.
3414 void PhaseIdealLoop::recompute_dom_depth() {
3415 uint no_depth_marker = C->unique();
3416 uint i;
3417 // Initialize depth to "no depth yet" and realize all lazy updates
3418 for (i = 0; i < _idom_size; i++) {
3419 // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized).
3420 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
3421 _dom_depth[i] = no_depth_marker;
3422
3423 // heal _idom if it has a fwd mapping in _nodes
3424 if (_idom[i]->in(0) == NULL) {
3425 idom(i);
3426 }
3427 }
3428 }
3429 if (_dom_stk == NULL) {
3430 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
3431 if (init_size < 10) init_size = 10;
3432 _dom_stk = new GrowableArray<uint>(init_size);
3433 }
3434 // Compute new depth for each node.
3435 for (i = 0; i < _idom_size; i++) {
3436 uint j = i;
3437 // Run up the dom tree to find a node with a depth
3438 while (_dom_depth[j] == no_depth_marker) {
3439 _dom_stk->push(j);
3440 j = _idom[j]->_idx;
3441 }
3442 // Compute the depth on the way back down this tree branch
3443 uint dd = _dom_depth[j] + 1;
3444 while (_dom_stk->length() > 0) {
3445 uint j = _dom_stk->pop();
3446 _dom_depth[j] = dd;
3447 dd++;
3448 }
3449 }
3450 }
3451
3452 //------------------------------sort-------------------------------------------
3453 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
3454 // loop tree, not the root.
3455 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
3456 if( !innermost ) return loop; // New innermost loop
3457
3458 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
3459 assert( loop_preorder, "not yet post-walked loop" );
3460 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
3461 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
3462
3463 // Insert at start of list
3464 while( l ) { // Insertion sort based on pre-order
3465 if( l == loop ) return innermost; // Already on list!
3466 int l_preorder = get_preorder(l->_head); // Cache pre-order number
3467 assert( l_preorder, "not yet post-walked l" );
3468 // Check header pre-order number to figure proper nesting
3469 if( loop_preorder > l_preorder )
3470 break; // End of insertion
3471 // If headers tie (e.g., shared headers) check tail pre-order numbers.
3472 // Since I split shared headers, you'd think this could not happen.
3473 // BUT: I must first do the preorder numbering before I can discover I
3474 // have shared headers, so the split headers all get the same preorder
3475 // number as the RegionNode they split from.
3476 if( loop_preorder == l_preorder &&
3477 get_preorder(loop->_tail) < get_preorder(l->_tail) )
3478 break; // Also check for shared headers (same pre#)
3479 pp = &l->_parent; // Chain up list
3480 l = *pp;
3481 }
3482 // Link into list
3483 // Point predecessor to me
3484 *pp = loop;
3485 // Point me to successor
3486 IdealLoopTree *p = loop->_parent;
3487 loop->_parent = l; // Point me to successor
3488 if( p ) sort( p, innermost ); // Insert my parents into list as well
3489 return innermost;
3490 }
3491
3492 //------------------------------build_loop_tree--------------------------------
3493 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
3494 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
3495 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
3496 // tightest enclosing IdealLoopTree for post-walked.
3497 //
3498 // During my forward walk I do a short 1-layer lookahead to see if I can find
3499 // a loop backedge with that doesn't have any work on the backedge. This
3500 // helps me construct nested loops with shared headers better.
3501 //
3502 // Once I've done the forward recursion, I do the post-work. For each child
3503 // I check to see if there is a backedge. Backedges define a loop! I
3504 // insert an IdealLoopTree at the target of the backedge.
3505 //
3506 // During the post-work I also check to see if I have several children
3507 // belonging to different loops. If so, then this Node is a decision point
3508 // where control flow can choose to change loop nests. It is at this
3509 // decision point where I can figure out how loops are nested. At this
3510 // time I can properly order the different loop nests from my children.
3511 // Note that there may not be any backedges at the decision point!
3512 //
3513 // Since the decision point can be far removed from the backedges, I can't
3514 // order my loops at the time I discover them. Thus at the decision point
3515 // I need to inspect loop header pre-order numbers to properly nest my
3516 // loops. This means I need to sort my childrens' loops by pre-order.
3517 // The sort is of size number-of-control-children, which generally limits
3518 // it to size 2 (i.e., I just choose between my 2 target loops).
3519 void PhaseIdealLoop::build_loop_tree() {
3520 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3521 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
3522 Node *n = C->root();
3523 bltstack.push(n);
3524 int pre_order = 1;
3525 int stack_size;
3526
3527 while ( ( stack_size = bltstack.length() ) != 0 ) {
3528 n = bltstack.top(); // Leave node on stack
3529 if ( !is_visited(n) ) {
3530 // ---- Pre-pass Work ----
3531 // Pre-walked but not post-walked nodes need a pre_order number.
3532
3533 set_preorder_visited( n, pre_order ); // set as visited
3534
3535 // ---- Scan over children ----
3536 // Scan first over control projections that lead to loop headers.
3537 // This helps us find inner-to-outer loops with shared headers better.
3538
3539 // Scan children's children for loop headers.
3540 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
3541 Node* m = n->raw_out(i); // Child
3542 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
3543 // Scan over children's children to find loop
3544 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3545 Node* l = m->fast_out(j);
3546 if( is_visited(l) && // Been visited?
3547 !is_postvisited(l) && // But not post-visited
3548 get_preorder(l) < pre_order ) { // And smaller pre-order
3549 // Found! Scan the DFS down this path before doing other paths
3550 bltstack.push(m);
3551 break;
3552 }
3553 }
3554 }
3555 }
3556 pre_order++;
3557 }
3558 else if ( !is_postvisited(n) ) {
3559 // Note: build_loop_tree_impl() adds out edges on rare occasions,
3560 // such as com.sun.rsasign.am::a.
3561 // For non-recursive version, first, process current children.
3562 // On next iteration, check if additional children were added.
3563 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
3564 Node* u = n->raw_out(k);
3565 if ( u->is_CFG() && !is_visited(u) ) {
3566 bltstack.push(u);
3567 }
3568 }
3569 if ( bltstack.length() == stack_size ) {
3570 // There were no additional children, post visit node now
3571 (void)bltstack.pop(); // Remove node from stack
3572 pre_order = build_loop_tree_impl( n, pre_order );
3573 // Check for bailout
3574 if (C->failing()) {
3575 return;
3576 }
3577 // Check to grow _preorders[] array for the case when
3578 // build_loop_tree_impl() adds new nodes.
3579 check_grow_preorders();
3580 }
3581 }
3582 else {
3583 (void)bltstack.pop(); // Remove post-visited node from stack
3584 }
3585 }
3586 }
3587
3588 //------------------------------build_loop_tree_impl---------------------------
3589 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
3590 // ---- Post-pass Work ----
3591 // Pre-walked but not post-walked nodes need a pre_order number.
3592
3593 // Tightest enclosing loop for this Node
3594 IdealLoopTree *innermost = NULL;
3595
3596 // For all children, see if any edge is a backedge. If so, make a loop
3597 // for it. Then find the tightest enclosing loop for the self Node.
3598 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3599 Node* m = n->fast_out(i); // Child
3600 if( n == m ) continue; // Ignore control self-cycles
3601 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
3602
3603 IdealLoopTree *l; // Child's loop
3604 if( !is_postvisited(m) ) { // Child visited but not post-visited?
3605 // Found a backedge
3606 assert( get_preorder(m) < pre_order, "should be backedge" );
3607 // Check for the RootNode, which is already a LoopNode and is allowed
3608 // to have multiple "backedges".
3609 if( m == C->root()) { // Found the root?
3610 l = _ltree_root; // Root is the outermost LoopNode
3611 } else { // Else found a nested loop
3612 // Insert a LoopNode to mark this loop.
3613 l = new IdealLoopTree(this, m, n);
3614 } // End of Else found a nested loop
3615 if( !has_loop(m) ) // If 'm' does not already have a loop set
3616 set_loop(m, l); // Set loop header to loop now
3617
3618 } else { // Else not a nested loop
3619 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
3620 l = get_loop(m); // Get previously determined loop
3621 // If successor is header of a loop (nest), move up-loop till it
3622 // is a member of some outer enclosing loop. Since there are no
3623 // shared headers (I've split them already) I only need to go up
3624 // at most 1 level.
3625 while( l && l->_head == m ) // Successor heads loop?
3626 l = l->_parent; // Move up 1 for me
3627 // If this loop is not properly parented, then this loop
3628 // has no exit path out, i.e. its an infinite loop.
3629 if( !l ) {
3630 // Make loop "reachable" from root so the CFG is reachable. Basically
3631 // insert a bogus loop exit that is never taken. 'm', the loop head,
3632 // points to 'n', one (of possibly many) fall-in paths. There may be
3633 // many backedges as well.
3634
3635 // Here I set the loop to be the root loop. I could have, after
3636 // inserting a bogus loop exit, restarted the recursion and found my
3637 // new loop exit. This would make the infinite loop a first-class
3638 // loop and it would then get properly optimized. What's the use of
3639 // optimizing an infinite loop?
3640 l = _ltree_root; // Oops, found infinite loop
3641
3642 if (!_verify_only) {
3643 // Insert the NeverBranch between 'm' and it's control user.
3644 NeverBranchNode *iff = new NeverBranchNode( m );
3645 _igvn.register_new_node_with_optimizer(iff);
3646 set_loop(iff, l);
3647 Node *if_t = new CProjNode( iff, 0 );
3648 _igvn.register_new_node_with_optimizer(if_t);
3649 set_loop(if_t, l);
3650
3651 Node* cfg = NULL; // Find the One True Control User of m
3652 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3653 Node* x = m->fast_out(j);
3654 if (x->is_CFG() && x != m && x != iff)
3655 { cfg = x; break; }
3656 }
3657 assert(cfg != NULL, "must find the control user of m");
3658 uint k = 0; // Probably cfg->in(0)
3659 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
3660 cfg->set_req( k, if_t ); // Now point to NeverBranch
3661 _igvn._worklist.push(cfg);
3662
3663 // Now create the never-taken loop exit
3664 Node *if_f = new CProjNode( iff, 1 );
3665 _igvn.register_new_node_with_optimizer(if_f);
3666 set_loop(if_f, l);
3667 // Find frame ptr for Halt. Relies on the optimizer
3668 // V-N'ing. Easier and quicker than searching through
3669 // the program structure.
3670 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3671 _igvn.register_new_node_with_optimizer(frame);
3672 // Halt & Catch Fire
3673 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
3674 _igvn.register_new_node_with_optimizer(halt);
3675 set_loop(halt, l);
3676 C->root()->add_req(halt);
3677 }
3678 set_loop(C->root(), _ltree_root);
3679 }
3680 }
3681 // Weeny check for irreducible. This child was already visited (this
3682 // IS the post-work phase). Is this child's loop header post-visited
3683 // as well? If so, then I found another entry into the loop.
3684 if (!_verify_only) {
3685 while( is_postvisited(l->_head) ) {
3686 // found irreducible
3687 l->_irreducible = 1; // = true
3688 l = l->_parent;
3689 _has_irreducible_loops = true;
3690 // Check for bad CFG here to prevent crash, and bailout of compile
3691 if (l == NULL) {
3692 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3693 return pre_order;
3694 }
3695 }
3696 C->set_has_irreducible_loop(_has_irreducible_loops);
3697 }
3698
3699 // This Node might be a decision point for loops. It is only if
3700 // it's children belong to several different loops. The sort call
3701 // does a trivial amount of work if there is only 1 child or all
3702 // children belong to the same loop. If however, the children
3703 // belong to different loops, the sort call will properly set the
3704 // _parent pointers to show how the loops nest.
3705 //
3706 // In any case, it returns the tightest enclosing loop.
3707 innermost = sort( l, innermost );
3708 }
3709
3710 // Def-use info will have some dead stuff; dead stuff will have no
3711 // loop decided on.
3712
3713 // Am I a loop header? If so fix up my parent's child and next ptrs.
3714 if( innermost && innermost->_head == n ) {
3715 assert( get_loop(n) == innermost, "" );
3716 IdealLoopTree *p = innermost->_parent;
3717 IdealLoopTree *l = innermost;
3718 while( p && l->_head == n ) {
3719 l->_next = p->_child; // Put self on parents 'next child'
3720 p->_child = l; // Make self as first child of parent
3721 l = p; // Now walk up the parent chain
3722 p = l->_parent;
3723 }
3724 } else {
3725 // Note that it is possible for a LoopNode to reach here, if the
3726 // backedge has been made unreachable (hence the LoopNode no longer
3727 // denotes a Loop, and will eventually be removed).
3728
3729 // Record tightest enclosing loop for self. Mark as post-visited.
3730 set_loop(n, innermost);
3731 // Also record has_call flag early on
3732 if( innermost ) {
3733 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3734 // Do not count uncommon calls
3735 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3736 Node *iff = n->in(0)->in(0);
3737 // No any calls for vectorized loops.
3738 if( UseSuperWord || !iff->is_If() ||
3739 (n->in(0)->Opcode() == Op_IfFalse &&
3740 (1.0 - iff->as_If()->_prob) >= 0.01) ||
3741 (iff->as_If()->_prob >= 0.01) )
3742 innermost->_has_call = 1;
3743 }
3744 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3745 // Disable loop optimizations if the loop has a scalar replaceable
3746 // allocation. This disabling may cause a potential performance lost
3747 // if the allocation is not eliminated for some reason.
3748 innermost->_allow_optimizations = false;
3749 innermost->_has_call = 1; // = true
3750 } else if (n->Opcode() == Op_SafePoint) {
3751 // Record all safepoints in this loop.
3752 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3753 innermost->_safepts->push(n);
3754 }
3755 }
3756 }
3757
3758 // Flag as post-visited now
3759 set_postvisited(n);
3760 return pre_order;
3761 }
3762
3763
3764 //------------------------------build_loop_early-------------------------------
3765 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3766 // First pass computes the earliest controlling node possible. This is the
3767 // controlling input with the deepest dominating depth.
3768 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3769 while (worklist.size() != 0) {
3770 // Use local variables nstack_top_n & nstack_top_i to cache values
3771 // on nstack's top.
3772 Node *nstack_top_n = worklist.pop();
3773 uint nstack_top_i = 0;
3774 //while_nstack_nonempty:
3775 while (true) {
3776 // Get parent node and next input's index from stack's top.
3777 Node *n = nstack_top_n;
3778 uint i = nstack_top_i;
3779 uint cnt = n->req(); // Count of inputs
3780 if (i == 0) { // Pre-process the node.
3781 if( has_node(n) && // Have either loop or control already?
3782 !has_ctrl(n) ) { // Have loop picked out already?
3783 // During "merge_many_backedges" we fold up several nested loops
3784 // into a single loop. This makes the members of the original
3785 // loop bodies pointing to dead loops; they need to move up
3786 // to the new UNION'd larger loop. I set the _head field of these
3787 // dead loops to NULL and the _parent field points to the owning
3788 // loop. Shades of UNION-FIND algorithm.
3789 IdealLoopTree *ilt;
3790 while( !(ilt = get_loop(n))->_head ) {
3791 // Normally I would use a set_loop here. But in this one special
3792 // case, it is legal (and expected) to change what loop a Node
3793 // belongs to.
3794 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3795 }
3796 // Remove safepoints ONLY if I've already seen I don't need one.
3797 // (the old code here would yank a 2nd safepoint after seeing a
3798 // first one, even though the 1st did not dominate in the loop body
3799 // and thus could be avoided indefinitely)
3800 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3801 is_deleteable_safept(n)) {
3802 Node *in = n->in(TypeFunc::Control);
3803 lazy_replace(n,in); // Pull safepoint now
3804 if (ilt->_safepts != NULL) {
3805 ilt->_safepts->yank(n);
3806 }
3807 // Carry on with the recursion "as if" we are walking
3808 // only the control input
3809 if( !visited.test_set( in->_idx ) ) {
3810 worklist.push(in); // Visit this guy later, using worklist
3811 }
3812 // Get next node from nstack:
3813 // - skip n's inputs processing by setting i > cnt;
3814 // - we also will not call set_early_ctrl(n) since
3815 // has_node(n) == true (see the condition above).
3816 i = cnt + 1;
3817 }
3818 }
3819 } // if (i == 0)
3820
3821 // Visit all inputs
3822 bool done = true; // Assume all n's inputs will be processed
3823 while (i < cnt) {
3824 Node *in = n->in(i);
3825 ++i;
3826 if (in == NULL) continue;
3827 if (in->pinned() && !in->is_CFG())
3828 set_ctrl(in, in->in(0));
3829 int is_visited = visited.test_set( in->_idx );
3830 if (!has_node(in)) { // No controlling input yet?
3831 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3832 assert( !is_visited, "visit only once" );
3833 nstack.push(n, i); // Save parent node and next input's index.
3834 nstack_top_n = in; // Process current input now.
3835 nstack_top_i = 0;
3836 done = false; // Not all n's inputs processed.
3837 break; // continue while_nstack_nonempty;
3838 } else if (!is_visited) {
3839 // This guy has a location picked out for him, but has not yet
3840 // been visited. Happens to all CFG nodes, for instance.
3841 // Visit him using the worklist instead of recursion, to break
3842 // cycles. Since he has a location already we do not need to
3843 // find his location before proceeding with the current Node.
3844 worklist.push(in); // Visit this guy later, using worklist
3845 }
3846 }
3847 if (done) {
3848 // All of n's inputs have been processed, complete post-processing.
3849
3850 // Compute earliest point this Node can go.
3851 // CFG, Phi, pinned nodes already know their controlling input.
3852 if (!has_node(n)) {
3853 // Record earliest legal location
3854 set_early_ctrl( n );
3855 }
3856 if (nstack.is_empty()) {
3857 // Finished all nodes on stack.
3858 // Process next node on the worklist.
3859 break;
3860 }
3861 // Get saved parent node and next input's index.
3862 nstack_top_n = nstack.node();
3863 nstack_top_i = nstack.index();
3864 nstack.pop();
3865 }
3866 } // while (true)
3867 }
3868 }
3869
3870 //------------------------------dom_lca_internal--------------------------------
3871 // Pair-wise LCA
3872 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3873 if( !n1 ) return n2; // Handle NULL original LCA
3874 assert( n1->is_CFG(), "" );
3875 assert( n2->is_CFG(), "" );
3876 // find LCA of all uses
3877 uint d1 = dom_depth(n1);
3878 uint d2 = dom_depth(n2);
3879 while (n1 != n2) {
3880 if (d1 > d2) {
3881 n1 = idom(n1);
3882 d1 = dom_depth(n1);
3883 } else if (d1 < d2) {
3884 n2 = idom(n2);
3885 d2 = dom_depth(n2);
3886 } else {
3887 // Here d1 == d2. Due to edits of the dominator-tree, sections
3888 // of the tree might have the same depth. These sections have
3889 // to be searched more carefully.
3890
3891 // Scan up all the n1's with equal depth, looking for n2.
3892 Node *t1 = idom(n1);
3893 while (dom_depth(t1) == d1) {
3894 if (t1 == n2) return n2;
3895 t1 = idom(t1);
3896 }
3897 // Scan up all the n2's with equal depth, looking for n1.
3898 Node *t2 = idom(n2);
3899 while (dom_depth(t2) == d2) {
3900 if (t2 == n1) return n1;
3901 t2 = idom(t2);
3902 }
3903 // Move up to a new dominator-depth value as well as up the dom-tree.
3904 n1 = t1;
3905 n2 = t2;
3906 d1 = dom_depth(n1);
3907 d2 = dom_depth(n2);
3908 }
3909 }
3910 return n1;
3911 }
3912
3913 //------------------------------compute_idom-----------------------------------
3914 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3915 // IDOMs are correct.
3916 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3917 assert( region->is_Region(), "" );
3918 Node *LCA = NULL;
3919 for( uint i = 1; i < region->req(); i++ ) {
3920 if( region->in(i) != C->top() )
3921 LCA = dom_lca( LCA, region->in(i) );
3922 }
3923 return LCA;
3924 }
3925
3926 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3927 bool had_error = false;
3928 #ifdef ASSERT
3929 if (early != C->root()) {
3930 // Make sure that there's a dominance path from LCA to early
3931 Node* d = LCA;
3932 while (d != early) {
3933 if (d == C->root()) {
3934 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3935 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3936 had_error = true;
3937 break;
3938 }
3939 d = idom(d);
3940 }
3941 }
3942 #endif
3943 return had_error;
3944 }
3945
3946
3947 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3948 // Compute LCA over list of uses
3949 bool had_error = false;
3950 Node *LCA = NULL;
3951 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3952 Node* c = n->fast_out(i);
3953 if (_nodes[c->_idx] == NULL)
3954 continue; // Skip the occasional dead node
3955 if( c->is_Phi() ) { // For Phis, we must land above on the path
3956 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3957 if( c->in(j) == n ) { // Found matching input?
3958 Node *use = c->in(0)->in(j);
3959 if (_verify_only && use->is_top()) continue;
3960 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3961 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3962 }
3963 }
3964 } else {
3965 // For CFG data-users, use is in the block just prior
3966 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3967 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3968 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3969 }
3970 }
3971 assert(!had_error, "bad dominance");
3972 return LCA;
3973 }
3974
3975 // Check the shape of the graph at the loop entry. In some cases,
3976 // the shape of the graph does not match the shape outlined below.
3977 // That is caused by the Opaque1 node "protecting" the shape of
3978 // the graph being removed by, for example, the IGVN performed
3979 // in PhaseIdealLoop::build_and_optimize().
3980 //
3981 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3982 // loop unswitching, and IGVN, or a combination of them) can freely change
3983 // the graph's shape. As a result, the graph shape outlined below cannot
3984 // be guaranteed anymore.
3985 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3986 if (!cl->is_main_loop() && !cl->is_post_loop()) {
3987 return false;
3988 }
3989 Node* ctrl = cl->skip_predicates();
3990
3991 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3992 return false;
3993 }
3994 Node* iffm = ctrl->in(0);
3995 if (iffm == NULL || !iffm->is_If()) {
3996 return false;
3997 }
3998 Node* bolzm = iffm->in(1);
3999 if (bolzm == NULL || !bolzm->is_Bool()) {
4000 return false;
4001 }
4002 Node* cmpzm = bolzm->in(1);
4003 if (cmpzm == NULL || !cmpzm->is_Cmp()) {
4004 return false;
4005 }
4006 // compares can get conditionally flipped
4007 bool found_opaque = false;
4008 for (uint i = 1; i < cmpzm->req(); i++) {
4009 Node* opnd = cmpzm->in(i);
4010 if (opnd && opnd->Opcode() == Op_Opaque1) {
4011 found_opaque = true;
4012 break;
4013 }
4014 }
4015 if (!found_opaque) {
4016 return false;
4017 }
4018 return true;
4019 }
4020
4021 //------------------------------get_late_ctrl----------------------------------
4022 // Compute latest legal control.
4023 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
4024 assert(early != NULL, "early control should not be NULL");
4025
4026 Node* LCA = compute_lca_of_uses(n, early);
4027 #ifdef ASSERT
4028 if (LCA == C->root() && LCA != early) {
4029 // def doesn't dominate uses so print some useful debugging output
4030 compute_lca_of_uses(n, early, true);
4031 }
4032 #endif
4033
4034 // if this is a load, check for anti-dependent stores
4035 // We use a conservative algorithm to identify potential interfering
4036 // instructions and for rescheduling the load. The users of the memory
4037 // input of this load are examined. Any use which is not a load and is
4038 // dominated by early is considered a potentially interfering store.
4039 // This can produce false positives.
4040 if (n->is_Load() && LCA != early) {
4041 int load_alias_idx = C->get_alias_index(n->adr_type());
4042 if (C->alias_type(load_alias_idx)->is_rewritable()) {
4043
4044 Node_List worklist;
4045
4046 Node *mem = n->in(MemNode::Memory);
4047 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
4048 Node* s = mem->fast_out(i);
4049 worklist.push(s);
4050 }
4051 while(worklist.size() != 0 && LCA != early) {
4052 Node* s = worklist.pop();
4053 if (s->is_Load() || s->Opcode() == Op_SafePoint ||
4054 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) {
4055 continue;
4056 } else if (s->is_MergeMem()) {
4057 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
4058 Node* s1 = s->fast_out(i);
4059 worklist.push(s1);
4060 }
4061 } else {
4062 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
4063 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
4064 if (sctrl != NULL && !sctrl->is_top() && C->can_alias(s->adr_type(), load_alias_idx) && is_dominator(early, sctrl)) {
4065 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
4066 }
4067 }
4068 }
4069 }
4070 }
4071
4072 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
4073 return LCA;
4074 }
4075
4076 // true if CFG node d dominates CFG node n
4077 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
4078 if (d == n)
4079 return true;
4080 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
4081 uint dd = dom_depth(d);
4082 while (dom_depth(n) >= dd) {
4083 if (n == d)
4084 return true;
4085 n = idom(n);
4086 }
4087 return false;
4088 }
4089
4090 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
4091 // Pair-wise LCA with tags.
4092 // Tag each index with the node 'tag' currently being processed
4093 // before advancing up the dominator chain using idom().
4094 // Later calls that find a match to 'tag' know that this path has already
4095 // been considered in the current LCA (which is input 'n1' by convention).
4096 // Since get_late_ctrl() is only called once for each node, the tag array
4097 // does not need to be cleared between calls to get_late_ctrl().
4098 // Algorithm trades a larger constant factor for better asymptotic behavior
4099 //
4100 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
4101 uint d1 = dom_depth(n1);
4102 uint d2 = dom_depth(n2);
4103
4104 do {
4105 if (d1 > d2) {
4106 // current lca is deeper than n2
4107 _dom_lca_tags.map(n1->_idx, tag);
4108 n1 = idom(n1);
4109 d1 = dom_depth(n1);
4110 } else if (d1 < d2) {
4111 // n2 is deeper than current lca
4112 Node *memo = _dom_lca_tags[n2->_idx];
4113 if( memo == tag ) {
4114 return n1; // Return the current LCA
4115 }
4116 _dom_lca_tags.map(n2->_idx, tag);
4117 n2 = idom(n2);
4118 d2 = dom_depth(n2);
4119 } else {
4120 // Here d1 == d2. Due to edits of the dominator-tree, sections
4121 // of the tree might have the same depth. These sections have
4122 // to be searched more carefully.
4123
4124 // Scan up all the n1's with equal depth, looking for n2.
4125 _dom_lca_tags.map(n1->_idx, tag);
4126 Node *t1 = idom(n1);
4127 while (dom_depth(t1) == d1) {
4128 if (t1 == n2) return n2;
4129 _dom_lca_tags.map(t1->_idx, tag);
4130 t1 = idom(t1);
4131 }
4132 // Scan up all the n2's with equal depth, looking for n1.
4133 _dom_lca_tags.map(n2->_idx, tag);
4134 Node *t2 = idom(n2);
4135 while (dom_depth(t2) == d2) {
4136 if (t2 == n1) return n1;
4137 _dom_lca_tags.map(t2->_idx, tag);
4138 t2 = idom(t2);
4139 }
4140 // Move up to a new dominator-depth value as well as up the dom-tree.
4141 n1 = t1;
4142 n2 = t2;
4143 d1 = dom_depth(n1);
4144 d2 = dom_depth(n2);
4145 }
4146 } while (n1 != n2);
4147 return n1;
4148 }
4149
4150 //------------------------------init_dom_lca_tags------------------------------
4151 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4152 // Intended use does not involve any growth for the array, so it could
4153 // be of fixed size.
4154 void PhaseIdealLoop::init_dom_lca_tags() {
4155 uint limit = C->unique() + 1;
4156 _dom_lca_tags.map( limit, NULL );
4157 #ifdef ASSERT
4158 for( uint i = 0; i < limit; ++i ) {
4159 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4160 }
4161 #endif // ASSERT
4162 }
4163
4164 //------------------------------clear_dom_lca_tags------------------------------
4165 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4166 // Intended use does not involve any growth for the array, so it could
4167 // be of fixed size.
4168 void PhaseIdealLoop::clear_dom_lca_tags() {
4169 uint limit = C->unique() + 1;
4170 _dom_lca_tags.map( limit, NULL );
4171 _dom_lca_tags.clear();
4172 #ifdef ASSERT
4173 for( uint i = 0; i < limit; ++i ) {
4174 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4175 }
4176 #endif // ASSERT
4177 }
4178
4179 //------------------------------build_loop_late--------------------------------
4180 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4181 // Second pass finds latest legal placement, and ideal loop placement.
4182 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
4183 while (worklist.size() != 0) {
4184 Node *n = worklist.pop();
4185 // Only visit once
4186 if (visited.test_set(n->_idx)) continue;
4187 uint cnt = n->outcnt();
4188 uint i = 0;
4189 while (true) {
4190 assert( _nodes[n->_idx], "no dead nodes" );
4191 // Visit all children
4192 if (i < cnt) {
4193 Node* use = n->raw_out(i);
4194 ++i;
4195 // Check for dead uses. Aggressively prune such junk. It might be
4196 // dead in the global sense, but still have local uses so I cannot
4197 // easily call 'remove_dead_node'.
4198 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
4199 // Due to cycles, we might not hit the same fixed point in the verify
4200 // pass as we do in the regular pass. Instead, visit such phis as
4201 // simple uses of the loop head.
4202 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
4203 if( !visited.test(use->_idx) )
4204 worklist.push(use);
4205 } else if( !visited.test_set(use->_idx) ) {
4206 nstack.push(n, i); // Save parent and next use's index.
4207 n = use; // Process all children of current use.
4208 cnt = use->outcnt();
4209 i = 0;
4210 }
4211 } else {
4212 // Do not visit around the backedge of loops via data edges.
4213 // push dead code onto a worklist
4214 _deadlist.push(use);
4215 }
4216 } else {
4217 // All of n's children have been processed, complete post-processing.
4218 build_loop_late_post(n);
4219 if (nstack.is_empty()) {
4220 // Finished all nodes on stack.
4221 // Process next node on the worklist.
4222 break;
4223 }
4224 // Get saved parent node and next use's index. Visit the rest of uses.
4225 n = nstack.node();
4226 cnt = n->outcnt();
4227 i = nstack.index();
4228 nstack.pop();
4229 }
4230 }
4231 }
4232 }
4233
4234 // Verify that no data node is scheduled in the outer loop of a strip
4235 // mined loop.
4236 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
4237 #ifdef ASSERT
4238 if (get_loop(least)->_nest == 0) {
4239 return;
4240 }
4241 IdealLoopTree* loop = get_loop(least);
4242 Node* head = loop->_head;
4243 if (head->is_OuterStripMinedLoop() &&
4244 // Verification can't be applied to fully built strip mined loops
4245 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
4246 Node* sfpt = head->as_Loop()->outer_safepoint();
4247 ResourceMark rm;
4248 Unique_Node_List wq;
4249 wq.push(sfpt);
4250 for (uint i = 0; i < wq.size(); i++) {
4251 Node *m = wq.at(i);
4252 for (uint i = 1; i < m->req(); i++) {
4253 Node* nn = m->in(i);
4254 if (nn == n) {
4255 return;
4256 }
4257 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
4258 wq.push(nn);
4259 }
4260 }
4261 }
4262 ShouldNotReachHere();
4263 }
4264 #endif
4265 }
4266
4267
4268 //------------------------------build_loop_late_post---------------------------
4269 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4270 // Second pass finds latest legal placement, and ideal loop placement.
4271 void PhaseIdealLoop::build_loop_late_post(Node *n) {
4272 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4273
4274 if (bs->build_loop_late_post(this, n)) {
4275 return;
4276 }
4277
4278 build_loop_late_post_work(n, true);
4279 }
4280
4281 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
4282
4283 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
4284 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
4285 }
4286
4287 #ifdef ASSERT
4288 if (_verify_only && !n->is_CFG()) {
4289 // Check def-use domination.
4290 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
4291 }
4292 #endif
4293
4294 // CFG and pinned nodes already handled
4295 if( n->in(0) ) {
4296 if( n->in(0)->is_top() ) return; // Dead?
4297
4298 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
4299 // _must_ be pinned (they have to observe their control edge of course).
4300 // Unlike Stores (which modify an unallocable resource, the memory
4301 // state), Mods/Loads can float around. So free them up.
4302 switch( n->Opcode() ) {
4303 case Op_DivI:
4304 case Op_DivF:
4305 case Op_DivD:
4306 case Op_ModI:
4307 case Op_ModF:
4308 case Op_ModD:
4309 case Op_LoadB: // Same with Loads; they can sink
4310 case Op_LoadUB: // during loop optimizations.
4311 case Op_LoadUS:
4312 case Op_LoadD:
4313 case Op_LoadF:
4314 case Op_LoadI:
4315 case Op_LoadKlass:
4316 case Op_LoadNKlass:
4317 case Op_LoadL:
4318 case Op_LoadS:
4319 case Op_LoadP:
4320 case Op_LoadN:
4321 case Op_LoadRange:
4322 case Op_LoadD_unaligned:
4323 case Op_LoadL_unaligned:
4324 case Op_StrComp: // Does a bunch of load-like effects
4325 case Op_StrEquals:
4326 case Op_StrIndexOf:
4327 case Op_StrIndexOfChar:
4328 case Op_AryEq:
4329 case Op_HasNegatives:
4330 pinned = false;
4331 }
4332 if( pinned ) {
4333 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
4334 if( !chosen_loop->_child ) // Inner loop?
4335 chosen_loop->_body.push(n); // Collect inner loops
4336 return;
4337 }
4338 } else { // No slot zero
4339 if( n->is_CFG() ) { // CFG with no slot 0 is dead
4340 _nodes.map(n->_idx,0); // No block setting, it's globally dead
4341 return;
4342 }
4343 assert(!n->is_CFG() || n->outcnt() == 0, "");
4344 }
4345
4346 // Do I have a "safe range" I can select over?
4347 Node *early = get_ctrl(n);// Early location already computed
4348
4349 // Compute latest point this Node can go
4350 Node *LCA = get_late_ctrl( n, early );
4351 // LCA is NULL due to uses being dead
4352 if( LCA == NULL ) {
4353 #ifdef ASSERT
4354 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
4355 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
4356 }
4357 #endif
4358 _nodes.map(n->_idx, 0); // This node is useless
4359 _deadlist.push(n);
4360 return;
4361 }
4362 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
4363
4364 Node *legal = LCA; // Walk 'legal' up the IDOM chain
4365 Node *least = legal; // Best legal position so far
4366 while( early != legal ) { // While not at earliest legal
4367 #ifdef ASSERT
4368 if (legal->is_Start() && !early->is_Root()) {
4369 // Bad graph. Print idom path and fail.
4370 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
4371 assert(false, "Bad graph detected in build_loop_late");
4372 }
4373 #endif
4374 // Find least loop nesting depth
4375 legal = idom(legal); // Bump up the IDOM tree
4376 // Check for lower nesting depth
4377 if( get_loop(legal)->_nest < get_loop(least)->_nest )
4378 least = legal;
4379 }
4380 assert(early == legal || legal != C->root(), "bad dominance of inputs");
4381
4382 // Try not to place code on a loop entry projection
4383 // which can inhibit range check elimination.
4384 if (least != early) {
4385 Node* ctrl_out = least->unique_ctrl_out();
4386 if (ctrl_out && ctrl_out->is_Loop() &&
4387 least == ctrl_out->in(LoopNode::EntryControl)) {
4388 // Move the node above predicates as far up as possible so a
4389 // following pass of loop predication doesn't hoist a predicate
4390 // that depends on it above that node.
4391 Node* new_ctrl = least;
4392 for (;;) {
4393 if (!new_ctrl->is_Proj()) {
4394 break;
4395 }
4396 CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
4397 if (call == NULL) {
4398 break;
4399 }
4400 int req = call->uncommon_trap_request();
4401 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
4402 if (trap_reason != Deoptimization::Reason_loop_limit_check &&
4403 trap_reason != Deoptimization::Reason_predicate &&
4404 trap_reason != Deoptimization::Reason_profile_predicate) {
4405 break;
4406 }
4407 Node* c = new_ctrl->in(0)->in(0);
4408 if (is_dominator(c, early) && c != early) {
4409 break;
4410 }
4411 new_ctrl = c;
4412 }
4413 least = new_ctrl;
4414 }
4415 }
4416
4417 #ifdef ASSERT
4418 // If verifying, verify that 'verify_me' has a legal location
4419 // and choose it as our location.
4420 if( _verify_me ) {
4421 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
4422 Node *legal = LCA;
4423 while( early != legal ) { // While not at earliest legal
4424 if( legal == v_ctrl ) break; // Check for prior good location
4425 legal = idom(legal) ;// Bump up the IDOM tree
4426 }
4427 // Check for prior good location
4428 if( legal == v_ctrl ) least = legal; // Keep prior if found
4429 }
4430 #endif
4431
4432 // Assign discovered "here or above" point
4433 least = find_non_split_ctrl(least);
4434 verify_strip_mined_scheduling(n, least);
4435 set_ctrl(n, least);
4436
4437 // Collect inner loop bodies
4438 IdealLoopTree *chosen_loop = get_loop(least);
4439 if( !chosen_loop->_child ) // Inner loop?
4440 chosen_loop->_body.push(n);// Collect inner loops
4441 }
4442
4443 #ifdef ASSERT
4444 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
4445 tty->print_cr("%s", msg);
4446 tty->print("n: "); n->dump();
4447 tty->print("early(n): "); early->dump();
4448 if (n->in(0) != NULL && !n->in(0)->is_top() &&
4449 n->in(0) != early && !n->in(0)->is_Root()) {
4450 tty->print("n->in(0): "); n->in(0)->dump();
4451 }
4452 for (uint i = 1; i < n->req(); i++) {
4453 Node* in1 = n->in(i);
4454 if (in1 != NULL && in1 != n && !in1->is_top()) {
4455 tty->print("n->in(%d): ", i); in1->dump();
4456 Node* in1_early = get_ctrl(in1);
4457 tty->print("early(n->in(%d)): ", i); in1_early->dump();
4458 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
4459 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
4460 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
4461 }
4462 for (uint j = 1; j < in1->req(); j++) {
4463 Node* in2 = in1->in(j);
4464 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
4465 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
4466 Node* in2_early = get_ctrl(in2);
4467 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
4468 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
4469 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
4470 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
4471 }
4472 }
4473 }
4474 }
4475 }
4476 tty->cr();
4477 tty->print("LCA(n): "); LCA->dump();
4478 for (uint i = 0; i < n->outcnt(); i++) {
4479 Node* u1 = n->raw_out(i);
4480 if (u1 == n)
4481 continue;
4482 tty->print("n->out(%d): ", i); u1->dump();
4483 if (u1->is_CFG()) {
4484 for (uint j = 0; j < u1->outcnt(); j++) {
4485 Node* u2 = u1->raw_out(j);
4486 if (u2 != u1 && u2 != n && u2->is_CFG()) {
4487 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4488 }
4489 }
4490 } else {
4491 Node* u1_later = get_ctrl(u1);
4492 tty->print("later(n->out(%d)): ", i); u1_later->dump();
4493 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
4494 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
4495 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
4496 }
4497 for (uint j = 0; j < u1->outcnt(); j++) {
4498 Node* u2 = u1->raw_out(j);
4499 if (u2 == n || u2 == u1)
4500 continue;
4501 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4502 if (!u2->is_CFG()) {
4503 Node* u2_later = get_ctrl(u2);
4504 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
4505 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
4506 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
4507 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
4508 }
4509 }
4510 }
4511 }
4512 }
4513 tty->cr();
4514 int ct = 0;
4515 Node *dbg_legal = LCA;
4516 while(!dbg_legal->is_Start() && ct < 100) {
4517 tty->print("idom[%d] ",ct); dbg_legal->dump();
4518 ct++;
4519 dbg_legal = idom(dbg_legal);
4520 }
4521 tty->cr();
4522 }
4523 #endif
4524
4525 #ifndef PRODUCT
4526 //------------------------------dump-------------------------------------------
4527 void PhaseIdealLoop::dump() const {
4528 ResourceMark rm;
4529 Arena* arena = Thread::current()->resource_area();
4530 Node_Stack stack(arena, C->live_nodes() >> 2);
4531 Node_List rpo_list;
4532 VectorSet visited(arena);
4533 visited.set(C->top()->_idx);
4534 rpo(C->root(), stack, visited, rpo_list);
4535 // Dump root loop indexed by last element in PO order
4536 dump(_ltree_root, rpo_list.size(), rpo_list);
4537 }
4538
4539 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {
4540 loop->dump_head();
4541
4542 // Now scan for CFG nodes in the same loop
4543 for (uint j = idx; j > 0; j--) {
4544 Node* n = rpo_list[j-1];
4545 if (!_nodes[n->_idx]) // Skip dead nodes
4546 continue;
4547
4548 if (get_loop(n) != loop) { // Wrong loop nest
4549 if (get_loop(n)->_head == n && // Found nested loop?
4550 get_loop(n)->_parent == loop)
4551 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
4552 continue;
4553 }
4554
4555 // Dump controlling node
4556 tty->sp(2 * loop->_nest);
4557 tty->print("C");
4558 if (n == C->root()) {
4559 n->dump();
4560 } else {
4561 Node* cached_idom = idom_no_update(n);
4562 Node* computed_idom = n->in(0);
4563 if (n->is_Region()) {
4564 computed_idom = compute_idom(n);
4565 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
4566 // any MultiBranch ctrl node), so apply a similar transform to
4567 // the cached idom returned from idom_no_update.
4568 cached_idom = find_non_split_ctrl(cached_idom);
4569 }
4570 tty->print(" ID:%d", computed_idom->_idx);
4571 n->dump();
4572 if (cached_idom != computed_idom) {
4573 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
4574 computed_idom->_idx, cached_idom->_idx);
4575 }
4576 }
4577 // Dump nodes it controls
4578 for (uint k = 0; k < _nodes.Size(); k++) {
4579 // (k < C->unique() && get_ctrl(find(k)) == n)
4580 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
4581 Node* m = C->root()->find(k);
4582 if (m && m->outcnt() > 0) {
4583 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
4584 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
4585 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
4586 }
4587 tty->sp(2 * loop->_nest + 1);
4588 m->dump();
4589 }
4590 }
4591 }
4592 }
4593 }
4594 #endif
4595
4596 // Collect a R-P-O for the whole CFG.
4597 // Result list is in post-order (scan backwards for RPO)
4598 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
4599 stk.push(start, 0);
4600 visited.set(start->_idx);
4601
4602 while (stk.is_nonempty()) {
4603 Node* m = stk.node();
4604 uint idx = stk.index();
4605 if (idx < m->outcnt()) {
4606 stk.set_index(idx + 1);
4607 Node* n = m->raw_out(idx);
4608 if (n->is_CFG() && !visited.test_set(n->_idx)) {
4609 stk.push(n, 0);
4610 }
4611 } else {
4612 rpo_list.push(m);
4613 stk.pop();
4614 }
4615 }
4616 }
4617
4618
4619 //=============================================================================
4620 //------------------------------LoopTreeIterator-------------------------------
4621
4622 // Advance to next loop tree using a preorder, left-to-right traversal.
4623 void LoopTreeIterator::next() {
4624 assert(!done(), "must not be done.");
4625 if (_curnt->_child != NULL) {
4626 _curnt = _curnt->_child;
4627 } else if (_curnt->_next != NULL) {
4628 _curnt = _curnt->_next;
4629 } else {
4630 while (_curnt != _root && _curnt->_next == NULL) {
4631 _curnt = _curnt->_parent;
4632 }
4633 if (_curnt == _root) {
4634 _curnt = NULL;
4635 assert(done(), "must be done.");
4636 } else {
4637 assert(_curnt->_next != NULL, "must be more to do");
4638 _curnt = _curnt->_next;
4639 }
4640 }
4641 }
--- EOF ---