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