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 if (LoopLimitCheck) {
467
468 // ===================================================
469 // Generate loop limit check to avoid integer overflow
470 // in cases like next (cyclic loops):
471 //
472 // for (i=0; i <= max_jint; i++) {}
473 // for (i=0; i < max_jint; i+=2) {}
474 //
475 //
476 // Limit check predicate depends on the loop test:
477 //
478 // for(;i != limit; i++) --> limit <= (max_jint)
479 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
480 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
481 //
482
483 // Check if limit is excluded to do more precise int overflow check.
484 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
485 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
486
487 // If compare points directly to the phi we need to adjust
488 // the compare so that it points to the incr. Limit have
489 // to be adjusted to keep trip count the same and the
490 // adjusted limit should be checked for int overflow.
491 if (phi_incr != NULL) {
492 stride_m += stride_con;
493 }
494
495 if (limit->is_Con()) {
496 int limit_con = limit->get_int();
497 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
498 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
499 // Bailout: it could be integer overflow.
500 return false;
501 }
502 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
503 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
504 // Limit's type may satisfy the condition, for example,
505 // when it is an array length.
506 } else {
507 // Generate loop's limit check.
508 // Loop limit check predicate should be near the loop.
509 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
510 if (!limit_check_proj) {
511 // The limit check predicate is not generated if this method trapped here before.
512 #ifdef ASSERT
513 if (TraceLoopLimitCheck) {
514 tty->print("missing loop limit check:");
515 loop->dump_head();
516 x->dump(1);
517 }
518 #endif
519 return false;
520 }
521
522 IfNode* check_iff = limit_check_proj->in(0)->as_If();
523 Node* cmp_limit;
524 Node* bol;
525
526 if (stride_con > 0) {
527 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
528 bol = new BoolNode(cmp_limit, BoolTest::le);
529 } else {
530 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
531 bol = new BoolNode(cmp_limit, BoolTest::ge);
532 }
533 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
534 bol = _igvn.register_new_node_with_optimizer(bol);
535 set_subtree_ctrl(bol);
536
537 // Replace condition in original predicate but preserve Opaque node
538 // so that previous predicates could be found.
539 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
540 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
541 Node* opq = check_iff->in(1)->in(1);
542 _igvn.replace_input_of(opq, 1, bol);
543 // Update ctrl.
544 set_ctrl(opq, check_iff->in(0));
545 set_ctrl(check_iff->in(1), check_iff->in(0));
546
547 #ifndef PRODUCT
548 // report that the loop predication has been actually performed
549 // for this loop
550 if (TraceLoopLimitCheck) {
551 tty->print_cr("Counted Loop Limit Check generated:");
552 debug_only( bol->dump(2); )
553 }
554 #endif
555 }
556
557 if (phi_incr != NULL) {
558 // If compare points directly to the phi we need to adjust
559 // the compare so that it points to the incr. Limit have
560 // to be adjusted to keep trip count the same and we
561 // should avoid int overflow.
562 //
563 // i = init; do {} while(i++ < limit);
564 // is converted to
565 // i = init; do {} while(++i < limit+1);
566 //
567 limit = gvn->transform(new AddINode(limit, stride));
568 }
569
570 // Now we need to canonicalize loop condition.
571 if (bt == BoolTest::ne) {
572 assert(stride_con == 1 || stride_con == -1, "simple increment only");
573 // 'ne' can be replaced with 'lt' only when init < limit.
574 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
575 bt = BoolTest::lt;
576 // 'ne' can be replaced with 'gt' only when init > limit.
577 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
578 bt = BoolTest::gt;
579 }
580
581 if (incl_limit) {
582 // The limit check guaranties that 'limit <= (max_jint - stride)' so
583 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
584 //
585 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
586 limit = gvn->transform(new AddINode(limit, one));
587 if (bt == BoolTest::le)
588 bt = BoolTest::lt;
589 else if (bt == BoolTest::ge)
590 bt = BoolTest::gt;
591 else
592 ShouldNotReachHere();
593 }
594 set_subtree_ctrl( limit );
595
596 } else { // LoopLimitCheck
597
598 // If compare points to incr, we are ok. Otherwise the compare
599 // can directly point to the phi; in this case adjust the compare so that
600 // it points to the incr by adjusting the limit.
601 if (cmp->in(1) == phi || cmp->in(2) == phi)
602 limit = gvn->transform(new AddINode(limit,stride));
603
604 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
605 // Final value for iterator should be: trip_count * stride + init_trip.
606 Node *one_p = gvn->intcon( 1);
607 Node *one_m = gvn->intcon(-1);
608
609 Node *trip_count = NULL;
610 switch( bt ) {
611 case BoolTest::eq:
612 ShouldNotReachHere();
613 case BoolTest::ne: // Ahh, the case we desire
614 if (stride_con == 1)
615 trip_count = gvn->transform(new SubINode(limit,init_trip));
616 else if (stride_con == -1)
617 trip_count = gvn->transform(new SubINode(init_trip,limit));
618 else
619 ShouldNotReachHere();
620 set_subtree_ctrl(trip_count);
621 //_loop.map(trip_count->_idx,loop(limit));
622 break;
623 case BoolTest::le: // Maybe convert to '<' case
624 limit = gvn->transform(new AddINode(limit,one_p));
625 set_subtree_ctrl( limit );
626 hook->init_req(4, limit);
627
628 bt = BoolTest::lt;
629 // Make the new limit be in the same loop nest as the old limit
630 //_loop.map(limit->_idx,limit_loop);
631 // Fall into next case
632 case BoolTest::lt: { // Maybe convert to '!=' case
633 if (stride_con < 0) // Count down loop rolls through MAXINT
634 ShouldNotReachHere();
635 Node *range = gvn->transform(new SubINode(limit,init_trip));
636 set_subtree_ctrl( range );
637 hook->init_req(0, range);
638
639 Node *bias = gvn->transform(new AddINode(range,stride));
640 set_subtree_ctrl( bias );
641 hook->init_req(1, bias);
642
643 Node *bias1 = gvn->transform(new AddINode(bias,one_m));
644 set_subtree_ctrl( bias1 );
645 hook->init_req(2, bias1);
646
647 trip_count = gvn->transform(new DivINode(0,bias1,stride));
648 set_subtree_ctrl( trip_count );
649 hook->init_req(3, trip_count);
650 break;
651 }
652
653 case BoolTest::ge: // Maybe convert to '>' case
654 limit = gvn->transform(new AddINode(limit,one_m));
655 set_subtree_ctrl( limit );
656 hook->init_req(4 ,limit);
657
658 bt = BoolTest::gt;
659 // Make the new limit be in the same loop nest as the old limit
660 //_loop.map(limit->_idx,limit_loop);
661 // Fall into next case
662 case BoolTest::gt: { // Maybe convert to '!=' case
663 if (stride_con > 0) // count up loop rolls through MININT
664 ShouldNotReachHere();
665 Node *range = gvn->transform(new SubINode(limit,init_trip));
666 set_subtree_ctrl( range );
667 hook->init_req(0, range);
668
669 Node *bias = gvn->transform(new AddINode(range,stride));
670 set_subtree_ctrl( bias );
671 hook->init_req(1, bias);
672
673 Node *bias1 = gvn->transform(new AddINode(bias,one_p));
674 set_subtree_ctrl( bias1 );
675 hook->init_req(2, bias1);
676
677 trip_count = gvn->transform(new DivINode(0,bias1,stride));
678 set_subtree_ctrl( trip_count );
679 hook->init_req(3, trip_count);
680 break;
681 }
682 } // switch( bt )
683
684 Node *span = gvn->transform(new MulINode(trip_count,stride));
685 set_subtree_ctrl( span );
686 hook->init_req(5, span);
687
688 limit = gvn->transform(new AddINode(span,init_trip));
689 set_subtree_ctrl( limit );
690
691 } // LoopLimitCheck
692
693 if (!UseCountedLoopSafepoints) {
694 // Check for SafePoint on backedge and remove
695 Node *sfpt = x->in(LoopNode::LoopBackControl);
696 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
697 lazy_replace( sfpt, iftrue );
698 if (loop->_safepts != NULL) {
699 loop->_safepts->yank(sfpt);
700 }
701 loop->_tail = iftrue;
702 }
703 }
704
705 // Build a canonical trip test.
706 // Clone code, as old values may be in use.
707 incr = incr->clone();
708 incr->set_req(1,phi);
709 incr->set_req(2,stride);
710 incr = _igvn.register_new_node_with_optimizer(incr);
711 set_early_ctrl( incr );
712 _igvn.rehash_node_delayed(phi);
713 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
714
715 // If phi type is more restrictive than Int, raise to
716 // Int to prevent (almost) infinite recursion in igvn
717 // which can only handle integer types for constants or minint..maxint.
718 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
719 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
720 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
721 nphi = _igvn.register_new_node_with_optimizer(nphi);
722 set_ctrl(nphi, get_ctrl(phi));
723 _igvn.replace_node(phi, nphi);
724 phi = nphi->as_Phi();
725 }
726 cmp = cmp->clone();
727 cmp->set_req(1,incr);
728 cmp->set_req(2,limit);
729 cmp = _igvn.register_new_node_with_optimizer(cmp);
730 set_ctrl(cmp, iff->in(0));
731
732 test = test->clone()->as_Bool();
733 (*(BoolTest*)&test->_test)._test = bt;
734 test->set_req(1,cmp);
735 _igvn.register_new_node_with_optimizer(test);
736 set_ctrl(test, iff->in(0));
737
738 // Replace the old IfNode with a new LoopEndNode
739 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
740 IfNode *le = lex->as_If();
741 uint dd = dom_depth(iff);
742 set_idom(le, le->in(0), dd); // Update dominance for loop exit
743 set_loop(le, loop);
744
745 // Get the loop-exit control
746 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
747
748 // Need to swap loop-exit and loop-back control?
749 if (iftrue_op == Op_IfFalse) {
750 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
751 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
752
753 loop->_tail = back_control = ift2;
754 set_loop(ift2, loop);
755 set_loop(iff2, get_loop(iffalse));
756
757 // Lazy update of 'get_ctrl' mechanism.
758 lazy_replace(iffalse, iff2);
759 lazy_replace(iftrue, ift2);
760
761 // Swap names
762 iffalse = iff2;
763 iftrue = ift2;
764 } else {
765 _igvn.rehash_node_delayed(iffalse);
766 _igvn.rehash_node_delayed(iftrue);
767 iffalse->set_req_X( 0, le, &_igvn );
768 iftrue ->set_req_X( 0, le, &_igvn );
769 }
770
771 set_idom(iftrue, le, dd+1);
772 set_idom(iffalse, le, dd+1);
773 assert(iff->outcnt() == 0, "should be dead now");
774 lazy_replace( iff, le ); // fix 'get_ctrl'
775
776 // Now setup a new CountedLoopNode to replace the existing LoopNode
777 CountedLoopNode *l = new CountedLoopNode(init_control, back_control);
778 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
779 // The following assert is approximately true, and defines the intention
780 // of can_be_counted_loop. It fails, however, because phase->type
781 // is not yet initialized for this loop and its parts.
782 //assert(l->can_be_counted_loop(this), "sanity");
783 _igvn.register_new_node_with_optimizer(l);
784 set_loop(l, loop);
785 loop->_head = l;
786 // Fix all data nodes placed at the old loop head.
787 // Uses the lazy-update mechanism of 'get_ctrl'.
788 lazy_replace( x, l );
789 set_idom(l, init_control, dom_depth(x));
790
791 if (!UseCountedLoopSafepoints) {
792 // Check for immediately preceding SafePoint and remove
793 Node *sfpt2 = le->in(0);
794 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
795 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
796 if (loop->_safepts != NULL) {
797 loop->_safepts->yank(sfpt2);
798 }
799 }
800 }
801
802 // Free up intermediate goo
803 _igvn.remove_dead_node(hook);
804
805 #ifdef ASSERT
806 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
807 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
808 #endif
809 #ifndef PRODUCT
810 if (TraceLoopOpts) {
811 tty->print("Counted ");
812 loop->dump_head();
813 }
814 #endif
815
816 C->print_method(PHASE_AFTER_CLOOPS, 3);
817
818 // Capture bounds of the loop in the induction variable Phi before
819 // subsequent transformation (iteration splitting) obscures the
820 // bounds
821 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
822
823 return true;
824 }
825
826 //----------------------exact_limit-------------------------------------------
827 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
828 assert(loop->_head->is_CountedLoop(), "");
829 CountedLoopNode *cl = loop->_head->as_CountedLoop();
830 assert(cl->is_valid_counted_loop(), "");
831
832 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
833 cl->limit()->Opcode() == Op_LoopLimit) {
834 // Old code has exact limit (it could be incorrect in case of int overflow).
835 // Loop limit is exact with stride == 1. And loop may already have exact limit.
836 return cl->limit();
837 }
838 Node *limit = NULL;
839 #ifdef ASSERT
840 BoolTest::mask bt = cl->loopexit()->test_trip();
841 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
842 #endif
843 if (cl->has_exact_trip_count()) {
844 // Simple case: loop has constant boundaries.
845 // Use jlongs to avoid integer overflow.
846 int stride_con = cl->stride_con();
847 jlong init_con = cl->init_trip()->get_int();
848 jlong limit_con = cl->limit()->get_int();
849 julong trip_cnt = cl->trip_count();
850 jlong final_con = init_con + trip_cnt*stride_con;
851 int final_int = (int)final_con;
852 // The final value should be in integer range since the loop
853 // is counted and the limit was checked for overflow.
854 assert(final_con == (jlong)final_int, "final value should be integer");
855 limit = _igvn.intcon(final_int);
856 } else {
857 // Create new LoopLimit node to get exact limit (final iv value).
858 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
859 register_new_node(limit, cl->in(LoopNode::EntryControl));
860 }
861 assert(limit != NULL, "sanity");
862 return limit;
863 }
864
865 //------------------------------Ideal------------------------------------------
866 // Return a node which is more "ideal" than the current node.
867 // Attempt to convert into a counted-loop.
868 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
869 if (!can_be_counted_loop(phase)) {
870 phase->C->set_major_progress();
871 }
872 return RegionNode::Ideal(phase, can_reshape);
873 }
874
875
876 //=============================================================================
877 //------------------------------Ideal------------------------------------------
878 // Return a node which is more "ideal" than the current node.
879 // Attempt to convert into a counted-loop.
880 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
881 return RegionNode::Ideal(phase, can_reshape);
882 }
883
884 //------------------------------dump_spec--------------------------------------
885 // Dump special per-node info
886 #ifndef PRODUCT
887 void CountedLoopNode::dump_spec(outputStream *st) const {
888 LoopNode::dump_spec(st);
889 if (stride_is_con()) {
890 st->print("stride: %d ",stride_con());
891 }
892 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
893 if (is_main_loop()) st->print("main of N%d", _idx);
894 if (is_post_loop()) st->print("post of N%d", _main_idx);
895 }
896 #endif
897
898 //=============================================================================
899 int CountedLoopEndNode::stride_con() const {
900 return stride()->bottom_type()->is_int()->get_con();
901 }
902
903 //=============================================================================
904 //------------------------------Value-----------------------------------------
905 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
906 const Type* init_t = phase->type(in(Init));
907 const Type* limit_t = phase->type(in(Limit));
908 const Type* stride_t = phase->type(in(Stride));
909 // Either input is TOP ==> the result is TOP
910 if (init_t == Type::TOP) return Type::TOP;
911 if (limit_t == Type::TOP) return Type::TOP;
912 if (stride_t == Type::TOP) return Type::TOP;
913
914 int stride_con = stride_t->is_int()->get_con();
915 if (stride_con == 1)
916 return NULL; // Identity
917
918 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
919 // Use jlongs to avoid integer overflow.
920 jlong init_con = init_t->is_int()->get_con();
921 jlong limit_con = limit_t->is_int()->get_con();
922 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
923 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
924 jlong final_con = init_con + stride_con*trip_count;
925 int final_int = (int)final_con;
926 // The final value should be in integer range since the loop
927 // is counted and the limit was checked for overflow.
928 assert(final_con == (jlong)final_int, "final value should be integer");
929 return TypeInt::make(final_int);
930 }
931
932 return bottom_type(); // TypeInt::INT
933 }
934
935 //------------------------------Ideal------------------------------------------
936 // Return a node which is more "ideal" than the current node.
937 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
938 if (phase->type(in(Init)) == Type::TOP ||
939 phase->type(in(Limit)) == Type::TOP ||
940 phase->type(in(Stride)) == Type::TOP)
941 return NULL; // Dead
942
943 int stride_con = phase->type(in(Stride))->is_int()->get_con();
944 if (stride_con == 1)
945 return NULL; // Identity
946
947 if (in(Init)->is_Con() && in(Limit)->is_Con())
948 return NULL; // Value
949
950 // Delay following optimizations until all loop optimizations
951 // done to keep Ideal graph simple.
952 if (!can_reshape || phase->C->major_progress())
953 return NULL;
954
955 const TypeInt* init_t = phase->type(in(Init) )->is_int();
956 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
957 int stride_p;
958 jlong lim, ini;
959 julong max;
960 if (stride_con > 0) {
961 stride_p = stride_con;
962 lim = limit_t->_hi;
963 ini = init_t->_lo;
964 max = (julong)max_jint;
965 } else {
966 stride_p = -stride_con;
967 lim = init_t->_hi;
968 ini = limit_t->_lo;
969 max = (julong)min_jint;
970 }
971 julong range = lim - ini + stride_p;
972 if (range <= max) {
973 // Convert to integer expression if it is not overflow.
974 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
975 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
976 Node *bias = phase->transform(new AddINode(range, stride_m));
977 Node *trip = phase->transform(new DivINode(0, bias, in(Stride)));
978 Node *span = phase->transform(new MulINode(trip, in(Stride)));
979 return new AddINode(span, in(Init)); // exact limit
980 }
981
982 if (is_power_of_2(stride_p) || // divisor is 2^n
983 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
984 // Convert to long expression to avoid integer overflow
985 // and let igvn optimizer convert this division.
986 //
987 Node* init = phase->transform( new ConvI2LNode(in(Init)));
988 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
989 Node* stride = phase->longcon(stride_con);
990 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
991
992 Node *range = phase->transform(new SubLNode(limit, init));
993 Node *bias = phase->transform(new AddLNode(range, stride_m));
994 Node *span;
995 if (stride_con > 0 && is_power_of_2(stride_p)) {
996 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
997 // and avoid generating rounding for division. Zero trip guard should
998 // guarantee that init < limit but sometimes the guard is missing and
999 // we can get situation when init > limit. Note, for the empty loop
1000 // optimization zero trip guard is generated explicitly which leaves
1001 // only RCE predicate where exact limit is used and the predicate
1002 // will simply fail forcing recompilation.
1003 Node* neg_stride = phase->longcon(-stride_con);
1004 span = phase->transform(new AndLNode(bias, neg_stride));
1005 } else {
1006 Node *trip = phase->transform(new DivLNode(0, bias, stride));
1007 span = phase->transform(new MulLNode(trip, stride));
1008 }
1009 // Convert back to int
1010 Node *span_int = phase->transform(new ConvL2INode(span));
1011 return new AddINode(span_int, in(Init)); // exact limit
1012 }
1013
1014 return NULL; // No progress
1015 }
1016
1017 //------------------------------Identity---------------------------------------
1018 // If stride == 1 return limit node.
1019 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1020 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1021 if (stride_con == 1 || stride_con == -1)
1022 return in(Limit);
1023 return this;
1024 }
1025
1026 //=============================================================================
1027 //----------------------match_incr_with_optional_truncation--------------------
1028 // Match increment with optional truncation:
1029 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1030 // Return NULL for failure. Success returns the increment node.
1031 Node* CountedLoopNode::match_incr_with_optional_truncation(
1032 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1033 // Quick cutouts:
1034 if (expr == NULL || expr->req() != 3) return NULL;
1035
1036 Node *t1 = NULL;
1037 Node *t2 = NULL;
1038 const TypeInt* trunc_t = TypeInt::INT;
1039 Node* n1 = expr;
1040 int n1op = n1->Opcode();
1041
1042 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1043 if (n1op == Op_AndI &&
1044 n1->in(2)->is_Con() &&
1045 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1046 // %%% This check should match any mask of 2**K-1.
1047 t1 = n1;
1048 n1 = t1->in(1);
1049 n1op = n1->Opcode();
1050 trunc_t = TypeInt::CHAR;
1051 } else if (n1op == Op_RShiftI &&
1052 n1->in(1) != NULL &&
1053 n1->in(1)->Opcode() == Op_LShiftI &&
1054 n1->in(2) == n1->in(1)->in(2) &&
1055 n1->in(2)->is_Con()) {
1056 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1057 // %%% This check should match any shift in [1..31].
1058 if (shift == 16 || shift == 8) {
1059 t1 = n1;
1060 t2 = t1->in(1);
1061 n1 = t2->in(1);
1062 n1op = n1->Opcode();
1063 if (shift == 16) {
1064 trunc_t = TypeInt::SHORT;
1065 } else if (shift == 8) {
1066 trunc_t = TypeInt::BYTE;
1067 }
1068 }
1069 }
1070
1071 // If (maybe after stripping) it is an AddI, we won:
1072 if (n1op == Op_AddI) {
1073 *trunc1 = t1;
1074 *trunc2 = t2;
1075 *trunc_type = trunc_t;
1076 return n1;
1077 }
1078
1079 // failed
1080 return NULL;
1081 }
1082
1083
1084 //------------------------------filtered_type--------------------------------
1085 // Return a type based on condition control flow
1086 // A successful return will be a type that is restricted due
1087 // to a series of dominating if-tests, such as:
1088 // if (i < 10) {
1089 // if (i > 0) {
1090 // here: "i" type is [1..10)
1091 // }
1092 // }
1093 // or a control flow merge
1094 // if (i < 10) {
1095 // do {
1096 // phi( , ) -- at top of loop type is [min_int..10)
1097 // i = ?
1098 // } while ( i < 10)
1099 //
1100 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1101 assert(n && n->bottom_type()->is_int(), "must be int");
1102 const TypeInt* filtered_t = NULL;
1103 if (!n->is_Phi()) {
1104 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1105 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1106
1107 } else {
1108 Node* phi = n->as_Phi();
1109 Node* region = phi->in(0);
1110 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1111 if (region && region != C->top()) {
1112 for (uint i = 1; i < phi->req(); i++) {
1113 Node* val = phi->in(i);
1114 Node* use_c = region->in(i);
1115 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1116 if (val_t != NULL) {
1117 if (filtered_t == NULL) {
1118 filtered_t = val_t;
1119 } else {
1120 filtered_t = filtered_t->meet(val_t)->is_int();
1121 }
1122 }
1123 }
1124 }
1125 }
1126 const TypeInt* n_t = _igvn.type(n)->is_int();
1127 if (filtered_t != NULL) {
1128 n_t = n_t->join(filtered_t)->is_int();
1129 }
1130 return n_t;
1131 }
1132
1133
1134 //------------------------------filtered_type_from_dominators--------------------------------
1135 // Return a possibly more restrictive type for val based on condition control flow of dominators
1136 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1137 if (val->is_Con()) {
1138 return val->bottom_type()->is_int();
1139 }
1140 uint if_limit = 10; // Max number of dominating if's visited
1141 const TypeInt* rtn_t = NULL;
1142
1143 if (use_ctrl && use_ctrl != C->top()) {
1144 Node* val_ctrl = get_ctrl(val);
1145 uint val_dom_depth = dom_depth(val_ctrl);
1146 Node* pred = use_ctrl;
1147 uint if_cnt = 0;
1148 while (if_cnt < if_limit) {
1149 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1150 if_cnt++;
1151 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1152 if (if_t != NULL) {
1153 if (rtn_t == NULL) {
1154 rtn_t = if_t;
1155 } else {
1156 rtn_t = rtn_t->join(if_t)->is_int();
1157 }
1158 }
1159 }
1160 pred = idom(pred);
1161 if (pred == NULL || pred == C->top()) {
1162 break;
1163 }
1164 // Stop if going beyond definition block of val
1165 if (dom_depth(pred) < val_dom_depth) {
1166 break;
1167 }
1168 }
1169 }
1170 return rtn_t;
1171 }
1172
1173
1174 //------------------------------dump_spec--------------------------------------
1175 // Dump special per-node info
1176 #ifndef PRODUCT
1177 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1178 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1179 BoolTest bt( test_trip()); // Added this for g++.
1180
1181 st->print("[");
1182 bt.dump_on(st);
1183 st->print("]");
1184 }
1185 st->print(" ");
1186 IfNode::dump_spec(st);
1187 }
1188 #endif
1189
1190 //=============================================================================
1191 //------------------------------is_member--------------------------------------
1192 // Is 'l' a member of 'this'?
1193 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1194 while( l->_nest > _nest ) l = l->_parent;
1195 return l == this;
1196 }
1197
1198 //------------------------------set_nest---------------------------------------
1199 // Set loop tree nesting depth. Accumulate _has_call bits.
1200 int IdealLoopTree::set_nest( uint depth ) {
1201 _nest = depth;
1202 int bits = _has_call;
1203 if( _child ) bits |= _child->set_nest(depth+1);
1204 if( bits ) _has_call = 1;
1205 if( _next ) bits |= _next ->set_nest(depth );
1206 return bits;
1207 }
1208
1209 //------------------------------split_fall_in----------------------------------
1210 // Split out multiple fall-in edges from the loop header. Move them to a
1211 // private RegionNode before the loop. This becomes the loop landing pad.
1212 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1213 PhaseIterGVN &igvn = phase->_igvn;
1214 uint i;
1215
1216 // Make a new RegionNode to be the landing pad.
1217 Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1218 phase->set_loop(landing_pad,_parent);
1219 // Gather all the fall-in control paths into the landing pad
1220 uint icnt = fall_in_cnt;
1221 uint oreq = _head->req();
1222 for( i = oreq-1; i>0; i-- )
1223 if( !phase->is_member( this, _head->in(i) ) )
1224 landing_pad->set_req(icnt--,_head->in(i));
1225
1226 // Peel off PhiNode edges as well
1227 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1228 Node *oj = _head->fast_out(j);
1229 if( oj->is_Phi() ) {
1230 PhiNode* old_phi = oj->as_Phi();
1231 assert( old_phi->region() == _head, "" );
1232 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1233 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1234 uint icnt = fall_in_cnt;
1235 for( i = oreq-1; i>0; i-- ) {
1236 if( !phase->is_member( this, _head->in(i) ) ) {
1237 p->init_req(icnt--, old_phi->in(i));
1238 // Go ahead and clean out old edges from old phi
1239 old_phi->del_req(i);
1240 }
1241 }
1242 // Search for CSE's here, because ZKM.jar does a lot of
1243 // loop hackery and we need to be a little incremental
1244 // with the CSE to avoid O(N^2) node blow-up.
1245 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1246 if( p2 ) { // Found CSE
1247 p->destruct(); // Recover useless new node
1248 p = p2; // Use old node
1249 } else {
1250 igvn.register_new_node_with_optimizer(p, old_phi);
1251 }
1252 // Make old Phi refer to new Phi.
1253 old_phi->add_req(p);
1254 // Check for the special case of making the old phi useless and
1255 // disappear it. In JavaGrande I have a case where this useless
1256 // Phi is the loop limit and prevents recognizing a CountedLoop
1257 // which in turn prevents removing an empty loop.
1258 Node *id_old_phi = old_phi->Identity( &igvn );
1259 if( id_old_phi != old_phi ) { // Found a simple identity?
1260 // Note that I cannot call 'replace_node' here, because
1261 // that will yank the edge from old_phi to the Region and
1262 // I'm mid-iteration over the Region's uses.
1263 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1264 Node* use = old_phi->last_out(i);
1265 igvn.rehash_node_delayed(use);
1266 uint uses_found = 0;
1267 for (uint j = 0; j < use->len(); j++) {
1268 if (use->in(j) == old_phi) {
1269 if (j < use->req()) use->set_req (j, id_old_phi);
1270 else use->set_prec(j, id_old_phi);
1271 uses_found++;
1272 }
1273 }
1274 i -= uses_found; // we deleted 1 or more copies of this edge
1275 }
1276 }
1277 igvn._worklist.push(old_phi);
1278 }
1279 }
1280 // Finally clean out the fall-in edges from the RegionNode
1281 for( i = oreq-1; i>0; i-- ) {
1282 if( !phase->is_member( this, _head->in(i) ) ) {
1283 _head->del_req(i);
1284 }
1285 }
1286 igvn.rehash_node_delayed(_head);
1287 // Transform landing pad
1288 igvn.register_new_node_with_optimizer(landing_pad, _head);
1289 // Insert landing pad into the header
1290 _head->add_req(landing_pad);
1291 }
1292
1293 //------------------------------split_outer_loop-------------------------------
1294 // Split out the outermost loop from this shared header.
1295 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1296 PhaseIterGVN &igvn = phase->_igvn;
1297
1298 // Find index of outermost loop; it should also be my tail.
1299 uint outer_idx = 1;
1300 while( _head->in(outer_idx) != _tail ) outer_idx++;
1301
1302 // Make a LoopNode for the outermost loop.
1303 Node *ctl = _head->in(LoopNode::EntryControl);
1304 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1305 outer = igvn.register_new_node_with_optimizer(outer, _head);
1306 phase->set_created_loop_node();
1307
1308 // Outermost loop falls into '_head' loop
1309 _head->set_req(LoopNode::EntryControl, outer);
1310 _head->del_req(outer_idx);
1311 // Split all the Phis up between '_head' loop and 'outer' loop.
1312 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1313 Node *out = _head->fast_out(j);
1314 if( out->is_Phi() ) {
1315 PhiNode *old_phi = out->as_Phi();
1316 assert( old_phi->region() == _head, "" );
1317 Node *phi = PhiNode::make_blank(outer, old_phi);
1318 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1319 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1320 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1321 // Make old Phi point to new Phi on the fall-in path
1322 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1323 old_phi->del_req(outer_idx);
1324 }
1325 }
1326
1327 // Use the new loop head instead of the old shared one
1328 _head = outer;
1329 phase->set_loop(_head, this);
1330 }
1331
1332 //------------------------------fix_parent-------------------------------------
1333 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1334 loop->_parent = parent;
1335 if( loop->_child ) fix_parent( loop->_child, loop );
1336 if( loop->_next ) fix_parent( loop->_next , parent );
1337 }
1338
1339 //------------------------------estimate_path_freq-----------------------------
1340 static float estimate_path_freq( Node *n ) {
1341 // Try to extract some path frequency info
1342 IfNode *iff;
1343 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1344 uint nop = n->Opcode();
1345 if( nop == Op_SafePoint ) { // Skip any safepoint
1346 n = n->in(0);
1347 continue;
1348 }
1349 if( nop == Op_CatchProj ) { // Get count from a prior call
1350 // Assume call does not always throw exceptions: means the call-site
1351 // count is also the frequency of the fall-through path.
1352 assert( n->is_CatchProj(), "" );
1353 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1354 return 0.0f; // Assume call exception path is rare
1355 Node *call = n->in(0)->in(0)->in(0);
1356 assert( call->is_Call(), "expect a call here" );
1357 const JVMState *jvms = ((CallNode*)call)->jvms();
1358 ciMethodData* methodData = jvms->method()->method_data();
1359 if (!methodData->is_mature()) return 0.0f; // No call-site data
1360 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1361 if ((data == NULL) || !data->is_CounterData()) {
1362 // no call profile available, try call's control input
1363 n = n->in(0);
1364 continue;
1365 }
1366 return data->as_CounterData()->count()/FreqCountInvocations;
1367 }
1368 // See if there's a gating IF test
1369 Node *n_c = n->in(0);
1370 if( !n_c->is_If() ) break; // No estimate available
1371 iff = n_c->as_If();
1372 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1373 // Compute how much count comes on this path
1374 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1375 // Have no count info. Skip dull uncommon-trap like branches.
1376 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1377 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1378 break;
1379 // Skip through never-taken branch; look for a real loop exit.
1380 n = iff->in(0);
1381 }
1382 return 0.0f; // No estimate available
1383 }
1384
1385 //------------------------------merge_many_backedges---------------------------
1386 // Merge all the backedges from the shared header into a private Region.
1387 // Feed that region as the one backedge to this loop.
1388 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1389 uint i;
1390
1391 // Scan for the top 2 hottest backedges
1392 float hotcnt = 0.0f;
1393 float warmcnt = 0.0f;
1394 uint hot_idx = 0;
1395 // Loop starts at 2 because slot 1 is the fall-in path
1396 for( i = 2; i < _head->req(); i++ ) {
1397 float cnt = estimate_path_freq(_head->in(i));
1398 if( cnt > hotcnt ) { // Grab hottest path
1399 warmcnt = hotcnt;
1400 hotcnt = cnt;
1401 hot_idx = i;
1402 } else if( cnt > warmcnt ) { // And 2nd hottest path
1403 warmcnt = cnt;
1404 }
1405 }
1406
1407 // See if the hottest backedge is worthy of being an inner loop
1408 // by being much hotter than the next hottest backedge.
1409 if( hotcnt <= 0.0001 ||
1410 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1411
1412 // Peel out the backedges into a private merge point; peel
1413 // them all except optionally hot_idx.
1414 PhaseIterGVN &igvn = phase->_igvn;
1415
1416 Node *hot_tail = NULL;
1417 // Make a Region for the merge point
1418 Node *r = new RegionNode(1);
1419 for( i = 2; i < _head->req(); i++ ) {
1420 if( i != hot_idx )
1421 r->add_req( _head->in(i) );
1422 else hot_tail = _head->in(i);
1423 }
1424 igvn.register_new_node_with_optimizer(r, _head);
1425 // Plug region into end of loop _head, followed by hot_tail
1426 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1427 igvn.replace_input_of(_head, 2, r);
1428 if( hot_idx ) _head->add_req(hot_tail);
1429
1430 // Split all the Phis up between '_head' loop and the Region 'r'
1431 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1432 Node *out = _head->fast_out(j);
1433 if( out->is_Phi() ) {
1434 PhiNode* n = out->as_Phi();
1435 igvn.hash_delete(n); // Delete from hash before hacking edges
1436 Node *hot_phi = NULL;
1437 Node *phi = new PhiNode(r, n->type(), n->adr_type());
1438 // Check all inputs for the ones to peel out
1439 uint j = 1;
1440 for( uint i = 2; i < n->req(); i++ ) {
1441 if( i != hot_idx )
1442 phi->set_req( j++, n->in(i) );
1443 else hot_phi = n->in(i);
1444 }
1445 // Register the phi but do not transform until whole place transforms
1446 igvn.register_new_node_with_optimizer(phi, n);
1447 // Add the merge phi to the old Phi
1448 while( n->req() > 3 ) n->del_req( n->req()-1 );
1449 igvn.replace_input_of(n, 2, phi);
1450 if( hot_idx ) n->add_req(hot_phi);
1451 }
1452 }
1453
1454
1455 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1456 // of self loop tree. Turn self into a loop headed by _head and with
1457 // tail being the new merge point.
1458 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1459 phase->set_loop(_tail,ilt); // Adjust tail
1460 _tail = r; // Self's tail is new merge point
1461 phase->set_loop(r,this);
1462 ilt->_child = _child; // New guy has my children
1463 _child = ilt; // Self has new guy as only child
1464 ilt->_parent = this; // new guy has self for parent
1465 ilt->_nest = _nest; // Same nesting depth (for now)
1466
1467 // Starting with 'ilt', look for child loop trees using the same shared
1468 // header. Flatten these out; they will no longer be loops in the end.
1469 IdealLoopTree **pilt = &_child;
1470 while( ilt ) {
1471 if( ilt->_head == _head ) {
1472 uint i;
1473 for( i = 2; i < _head->req(); i++ )
1474 if( _head->in(i) == ilt->_tail )
1475 break; // Still a loop
1476 if( i == _head->req() ) { // No longer a loop
1477 // Flatten ilt. Hang ilt's "_next" list from the end of
1478 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1479 IdealLoopTree **cp = &ilt->_child;
1480 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1481 *cp = ilt->_next; // Hang next list at end of child list
1482 *pilt = ilt->_child; // Move child up to replace ilt
1483 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1484 ilt = ilt->_child; // Repeat using new ilt
1485 continue; // do not advance over ilt->_child
1486 }
1487 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1488 phase->set_loop(_head,ilt);
1489 }
1490 pilt = &ilt->_child; // Advance to next
1491 ilt = *pilt;
1492 }
1493
1494 if( _child ) fix_parent( _child, this );
1495 }
1496
1497 //------------------------------beautify_loops---------------------------------
1498 // Split shared headers and insert loop landing pads.
1499 // Insert a LoopNode to replace the RegionNode.
1500 // Return TRUE if loop tree is structurally changed.
1501 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1502 bool result = false;
1503 // Cache parts in locals for easy
1504 PhaseIterGVN &igvn = phase->_igvn;
1505
1506 igvn.hash_delete(_head); // Yank from hash before hacking edges
1507
1508 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1509 int fall_in_cnt = 0;
1510 for( uint i = 1; i < _head->req(); i++ )
1511 if( !phase->is_member( this, _head->in(i) ) )
1512 fall_in_cnt++;
1513 assert( fall_in_cnt, "at least 1 fall-in path" );
1514 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1515 split_fall_in( phase, fall_in_cnt );
1516
1517 // Swap inputs to the _head and all Phis to move the fall-in edge to
1518 // the left.
1519 fall_in_cnt = 1;
1520 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1521 fall_in_cnt++;
1522 if( fall_in_cnt > 1 ) {
1523 // Since I am just swapping inputs I do not need to update def-use info
1524 Node *tmp = _head->in(1);
1525 igvn.rehash_node_delayed(_head);
1526 _head->set_req( 1, _head->in(fall_in_cnt) );
1527 _head->set_req( fall_in_cnt, tmp );
1528 // Swap also all Phis
1529 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1530 Node* phi = _head->fast_out(i);
1531 if( phi->is_Phi() ) {
1532 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
1533 tmp = phi->in(1);
1534 phi->set_req( 1, phi->in(fall_in_cnt) );
1535 phi->set_req( fall_in_cnt, tmp );
1536 }
1537 }
1538 }
1539 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1540 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1541
1542 // If I am a shared header (multiple backedges), peel off the many
1543 // backedges into a private merge point and use the merge point as
1544 // the one true backedge.
1545 if( _head->req() > 3 ) {
1546 // Merge the many backedges into a single backedge but leave
1547 // the hottest backedge as separate edge for the following peel.
1548 merge_many_backedges( phase );
1549 result = true;
1550 }
1551
1552 // If I have one hot backedge, peel off myself loop.
1553 // I better be the outermost loop.
1554 if (_head->req() > 3 && !_irreducible) {
1555 split_outer_loop( phase );
1556 result = true;
1557
1558 } else if (!_head->is_Loop() && !_irreducible) {
1559 // Make a new LoopNode to replace the old loop head
1560 Node *l = new LoopNode( _head->in(1), _head->in(2) );
1561 l = igvn.register_new_node_with_optimizer(l, _head);
1562 phase->set_created_loop_node();
1563 // Go ahead and replace _head
1564 phase->_igvn.replace_node( _head, l );
1565 _head = l;
1566 phase->set_loop(_head, this);
1567 }
1568
1569 // Now recursively beautify nested loops
1570 if( _child ) result |= _child->beautify_loops( phase );
1571 if( _next ) result |= _next ->beautify_loops( phase );
1572 return result;
1573 }
1574
1575 //------------------------------allpaths_check_safepts----------------------------
1576 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1577 // encountered. Helper for check_safepts.
1578 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1579 assert(stack.size() == 0, "empty stack");
1580 stack.push(_tail);
1581 visited.Clear();
1582 visited.set(_tail->_idx);
1583 while (stack.size() > 0) {
1584 Node* n = stack.pop();
1585 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1586 // Terminate this path
1587 } else if (n->Opcode() == Op_SafePoint) {
1588 if (_phase->get_loop(n) != this) {
1589 if (_required_safept == NULL) _required_safept = new Node_List();
1590 _required_safept->push(n); // save the one closest to the tail
1591 }
1592 // Terminate this path
1593 } else {
1594 uint start = n->is_Region() ? 1 : 0;
1595 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1596 for (uint i = start; i < end; i++) {
1597 Node* in = n->in(i);
1598 assert(in->is_CFG(), "must be");
1599 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1600 stack.push(in);
1601 }
1602 }
1603 }
1604 }
1605 }
1606
1607 //------------------------------check_safepts----------------------------
1608 // Given dominators, try to find loops with calls that must always be
1609 // executed (call dominates loop tail). These loops do not need non-call
1610 // safepoints (ncsfpt).
1611 //
1612 // A complication is that a safepoint in a inner loop may be needed
1613 // by an outer loop. In the following, the inner loop sees it has a
1614 // call (block 3) on every path from the head (block 2) to the
1615 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1616 // in block 2, _but_ this leaves the outer loop without a safepoint.
1617 //
1618 // entry 0
1619 // |
1620 // v
1621 // outer 1,2 +->1
1622 // | |
1623 // | v
1624 // | 2<---+ ncsfpt in 2
1625 // |_/|\ |
1626 // | v |
1627 // inner 2,3 / 3 | call in 3
1628 // / | |
1629 // v +--+
1630 // exit 4
1631 //
1632 //
1633 // This method creates a list (_required_safept) of ncsfpt nodes that must
1634 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1635 // is first looked for in the lists for the outer loops of the current loop.
1636 //
1637 // The insights into the problem:
1638 // A) counted loops are okay
1639 // B) innermost loops are okay (only an inner loop can delete
1640 // a ncsfpt needed by an outer loop)
1641 // C) a loop is immune from an inner loop deleting a safepoint
1642 // if the loop has a call on the idom-path
1643 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1644 // idom-path that is not in a nested loop
1645 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1646 // loop needs to be prevented from deletion by an inner loop
1647 //
1648 // There are two analyses:
1649 // 1) The first, and cheaper one, scans the loop body from
1650 // tail to head following the idom (immediate dominator)
1651 // chain, looking for the cases (C,D,E) above.
1652 // Since inner loops are scanned before outer loops, there is summary
1653 // information about inner loops. Inner loops can be skipped over
1654 // when the tail of an inner loop is encountered.
1655 //
1656 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1657 // the idom path (which is rare), scans all predecessor control paths
1658 // from the tail to the head, terminating a path when a call or sfpt
1659 // is encountered, to find the ncsfpt's that are closest to the tail.
1660 //
1661 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1662 // Bottom up traversal
1663 IdealLoopTree* ch = _child;
1664 if (_child) _child->check_safepts(visited, stack);
1665 if (_next) _next ->check_safepts(visited, stack);
1666
1667 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1668 bool has_call = false; // call on dom-path
1669 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1670 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1671 // Scan the dom-path nodes from tail to head
1672 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1673 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1674 has_call = true;
1675 _has_sfpt = 1; // Then no need for a safept!
1676 break;
1677 } else if (n->Opcode() == Op_SafePoint) {
1678 if (_phase->get_loop(n) == this) {
1679 has_local_ncsfpt = true;
1680 break;
1681 }
1682 if (nonlocal_ncsfpt == NULL) {
1683 nonlocal_ncsfpt = n; // save the one closest to the tail
1684 }
1685 } else {
1686 IdealLoopTree* nlpt = _phase->get_loop(n);
1687 if (this != nlpt) {
1688 // If at an inner loop tail, see if the inner loop has already
1689 // recorded seeing a call on the dom-path (and stop.) If not,
1690 // jump to the head of the inner loop.
1691 assert(is_member(nlpt), "nested loop");
1692 Node* tail = nlpt->_tail;
1693 if (tail->in(0)->is_If()) tail = tail->in(0);
1694 if (n == tail) {
1695 // If inner loop has call on dom-path, so does outer loop
1696 if (nlpt->_has_sfpt) {
1697 has_call = true;
1698 _has_sfpt = 1;
1699 break;
1700 }
1701 // Skip to head of inner loop
1702 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1703 n = nlpt->_head;
1704 }
1705 }
1706 }
1707 }
1708 // Record safept's that this loop needs preserved when an
1709 // inner loop attempts to delete it's safepoints.
1710 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1711 if (nonlocal_ncsfpt != NULL) {
1712 if (_required_safept == NULL) _required_safept = new Node_List();
1713 _required_safept->push(nonlocal_ncsfpt);
1714 } else {
1715 // Failed to find a suitable safept on the dom-path. Now use
1716 // an all paths walk from tail to head, looking for safepoints to preserve.
1717 allpaths_check_safepts(visited, stack);
1718 }
1719 }
1720 }
1721 }
1722
1723 //---------------------------is_deleteable_safept----------------------------
1724 // Is safept not required by an outer loop?
1725 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1726 assert(sfpt->Opcode() == Op_SafePoint, "");
1727 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1728 while (lp != NULL) {
1729 Node_List* sfpts = lp->_required_safept;
1730 if (sfpts != NULL) {
1731 for (uint i = 0; i < sfpts->size(); i++) {
1732 if (sfpt == sfpts->at(i))
1733 return false;
1734 }
1735 }
1736 lp = lp->_parent;
1737 }
1738 return true;
1739 }
1740
1741 //---------------------------replace_parallel_iv-------------------------------
1742 // Replace parallel induction variable (parallel to trip counter)
1743 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1744 assert(loop->_head->is_CountedLoop(), "");
1745 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1746 if (!cl->is_valid_counted_loop())
1747 return; // skip malformed counted loop
1748 Node *incr = cl->incr();
1749 if (incr == NULL)
1750 return; // Dead loop?
1751 Node *init = cl->init_trip();
1752 Node *phi = cl->phi();
1753 int stride_con = cl->stride_con();
1754
1755 // Visit all children, looking for Phis
1756 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1757 Node *out = cl->out(i);
1758 // Look for other phis (secondary IVs). Skip dead ones
1759 if (!out->is_Phi() || out == phi || !has_node(out))
1760 continue;
1761 PhiNode* phi2 = out->as_Phi();
1762 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1763 // Look for induction variables of the form: X += constant
1764 if (phi2->region() != loop->_head ||
1765 incr2->req() != 3 ||
1766 incr2->in(1) != phi2 ||
1767 incr2 == incr ||
1768 incr2->Opcode() != Op_AddI ||
1769 !incr2->in(2)->is_Con())
1770 continue;
1771
1772 // Check for parallel induction variable (parallel to trip counter)
1773 // via an affine function. In particular, count-down loops with
1774 // count-up array indices are common. We only RCE references off
1775 // the trip-counter, so we need to convert all these to trip-counter
1776 // expressions.
1777 Node *init2 = phi2->in( LoopNode::EntryControl );
1778 int stride_con2 = incr2->in(2)->get_int();
1779
1780 // The general case here gets a little tricky. We want to find the
1781 // GCD of all possible parallel IV's and make a new IV using this
1782 // GCD for the loop. Then all possible IVs are simple multiples of
1783 // the GCD. In practice, this will cover very few extra loops.
1784 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1785 // where +/-1 is the common case, but other integer multiples are
1786 // also easy to handle.
1787 int ratio_con = stride_con2/stride_con;
1788
1789 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1790 #ifndef PRODUCT
1791 if (TraceLoopOpts) {
1792 tty->print("Parallel IV: %d ", phi2->_idx);
1793 loop->dump_head();
1794 }
1795 #endif
1796 // Convert to using the trip counter. The parallel induction
1797 // variable differs from the trip counter by a loop-invariant
1798 // amount, the difference between their respective initial values.
1799 // It is scaled by the 'ratio_con'.
1800 Node* ratio = _igvn.intcon(ratio_con);
1801 set_ctrl(ratio, C->root());
1802 Node* ratio_init = new MulINode(init, ratio);
1803 _igvn.register_new_node_with_optimizer(ratio_init, init);
1804 set_early_ctrl(ratio_init);
1805 Node* diff = new SubINode(init2, ratio_init);
1806 _igvn.register_new_node_with_optimizer(diff, init2);
1807 set_early_ctrl(diff);
1808 Node* ratio_idx = new MulINode(phi, ratio);
1809 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1810 set_ctrl(ratio_idx, cl);
1811 Node* add = new AddINode(ratio_idx, diff);
1812 _igvn.register_new_node_with_optimizer(add);
1813 set_ctrl(add, cl);
1814 _igvn.replace_node( phi2, add );
1815 // Sometimes an induction variable is unused
1816 if (add->outcnt() == 0) {
1817 _igvn.remove_dead_node(add);
1818 }
1819 --i; // deleted this phi; rescan starting with next position
1820 continue;
1821 }
1822 }
1823 }
1824
1825 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
1826 Node* keep = NULL;
1827 if (keep_one) {
1828 // Look for a safepoint on the idom-path.
1829 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
1830 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
1831 keep = i;
1832 break; // Found one
1833 }
1834 }
1835 }
1836
1837 // Don't remove any safepoints if it is requested to keep a single safepoint and
1838 // no safepoint was found on idom-path. It is not safe to remove any safepoint
1839 // in this case since there's no safepoint dominating all paths in the loop body.
1840 bool prune = !keep_one || keep != NULL;
1841
1842 // Delete other safepoints in this loop.
1843 Node_List* sfpts = _safepts;
1844 if (prune && sfpts != NULL) {
1845 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
1846 for (uint i = 0; i < sfpts->size(); i++) {
1847 Node* n = sfpts->at(i);
1848 assert(phase->get_loop(n) == this, "");
1849 if (n != keep && phase->is_deleteable_safept(n)) {
1850 phase->lazy_replace(n, n->in(TypeFunc::Control));
1851 }
1852 }
1853 }
1854 }
1855
1856 //------------------------------counted_loop-----------------------------------
1857 // Convert to counted loops where possible
1858 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1859
1860 // For grins, set the inner-loop flag here
1861 if (!_child) {
1862 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1863 }
1864
1865 if (_head->is_CountedLoop() ||
1866 phase->is_counted_loop(_head, this)) {
1867
1868 if (!UseCountedLoopSafepoints) {
1869 // Indicate we do not need a safepoint here
1870 _has_sfpt = 1;
1871 }
1872
1873 // Remove safepoints
1874 bool keep_one_sfpt = !(_has_call || _has_sfpt);
1875 remove_safepoints(phase, keep_one_sfpt);
1876
1877 // Look for induction variables
1878 phase->replace_parallel_iv(this);
1879
1880 } else if (_parent != NULL && !_irreducible) {
1881 // Not a counted loop. Keep one safepoint.
1882 bool keep_one_sfpt = true;
1883 remove_safepoints(phase, keep_one_sfpt);
1884 }
1885
1886 // Recursively
1887 if (_child) _child->counted_loop( phase );
1888 if (_next) _next ->counted_loop( phase );
1889 }
1890
1891 #ifndef PRODUCT
1892 //------------------------------dump_head--------------------------------------
1893 // Dump 1 liner for loop header info
1894 void IdealLoopTree::dump_head( ) const {
1895 for (uint i=0; i<_nest; i++)
1896 tty->print(" ");
1897 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1898 if (_irreducible) tty->print(" IRREDUCIBLE");
1899 Node* entry = _head->in(LoopNode::EntryControl);
1900 if (LoopLimitCheck) {
1901 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1902 if (predicate != NULL ) {
1903 tty->print(" limit_check");
1904 entry = entry->in(0)->in(0);
1905 }
1906 }
1907 if (UseLoopPredicate) {
1908 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1909 if (entry != NULL) {
1910 tty->print(" predicated");
1911 }
1912 }
1913 if (_head->is_CountedLoop()) {
1914 CountedLoopNode *cl = _head->as_CountedLoop();
1915 tty->print(" counted");
1916
1917 Node* init_n = cl->init_trip();
1918 if (init_n != NULL && init_n->is_Con())
1919 tty->print(" [%d,", cl->init_trip()->get_int());
1920 else
1921 tty->print(" [int,");
1922 Node* limit_n = cl->limit();
1923 if (limit_n != NULL && limit_n->is_Con())
1924 tty->print("%d),", cl->limit()->get_int());
1925 else
1926 tty->print("int),");
1927 int stride_con = cl->stride_con();
1928 if (stride_con > 0) tty->print("+");
1929 tty->print("%d", stride_con);
1930
1931 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
1932
1933 if (cl->is_pre_loop ()) tty->print(" pre" );
1934 if (cl->is_main_loop()) tty->print(" main");
1935 if (cl->is_post_loop()) tty->print(" post");
1936 if (cl->is_vectorized_loop()) tty->print(" vector");
1937 if (cl->has_range_checks()) tty->print(" rc ");
1938 if (cl->is_multiversioned()) tty->print(" multi ");
1939 }
1940 if (_has_call) tty->print(" has_call");
1941 if (_has_sfpt) tty->print(" has_sfpt");
1942 if (_rce_candidate) tty->print(" rce");
1943 if (_safepts != NULL && _safepts->size() > 0) {
1944 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
1945 }
1946 if (_required_safept != NULL && _required_safept->size() > 0) {
1947 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
1948 }
1949 tty->cr();
1950 }
1951
1952 //------------------------------dump-------------------------------------------
1953 // Dump loops by loop tree
1954 void IdealLoopTree::dump( ) const {
1955 dump_head();
1956 if (_child) _child->dump();
1957 if (_next) _next ->dump();
1958 }
1959
1960 #endif
1961
1962 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1963 if (loop == root) {
1964 if (loop->_child != NULL) {
1965 log->begin_head("loop_tree");
1966 log->end_head();
1967 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1968 log->tail("loop_tree");
1969 assert(loop->_next == NULL, "what?");
1970 }
1971 } else {
1972 Node* head = loop->_head;
1973 log->begin_head("loop");
1974 log->print(" idx='%d' ", head->_idx);
1975 if (loop->_irreducible) log->print("irreducible='1' ");
1976 if (head->is_Loop()) {
1977 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1978 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1979 }
1980 if (head->is_CountedLoop()) {
1981 CountedLoopNode* cl = head->as_CountedLoop();
1982 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1983 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1984 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1985 }
1986 log->end_head();
1987 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1988 log->tail("loop");
1989 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1990 }
1991 }
1992
1993 //---------------------collect_potentially_useful_predicates-----------------------
1994 // Helper function to collect potentially useful predicates to prevent them from
1995 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1996 void PhaseIdealLoop::collect_potentially_useful_predicates(
1997 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1998 if (loop->_child) { // child
1999 collect_potentially_useful_predicates(loop->_child, useful_predicates);
2000 }
2001
2002 // self (only loops that we can apply loop predication may use their predicates)
2003 if (loop->_head->is_Loop() &&
2004 !loop->_irreducible &&
2005 !loop->tail()->is_top()) {
2006 LoopNode* lpn = loop->_head->as_Loop();
2007 Node* entry = lpn->in(LoopNode::EntryControl);
2008 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2009 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2010 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2011 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2012 entry = entry->in(0)->in(0);
2013 }
2014 predicate_proj = find_predicate(entry); // Predicate
2015 if (predicate_proj != NULL ) {
2016 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2017 }
2018 }
2019
2020 if (loop->_next) { // sibling
2021 collect_potentially_useful_predicates(loop->_next, useful_predicates);
2022 }
2023 }
2024
2025 //------------------------eliminate_useless_predicates-----------------------------
2026 // Eliminate all inserted predicates if they could not be used by loop predication.
2027 // Note: it will also eliminates loop limits check predicate since it also uses
2028 // Opaque1 node (see Parse::add_predicate()).
2029 void PhaseIdealLoop::eliminate_useless_predicates() {
2030 if (C->predicate_count() == 0)
2031 return; // no predicate left
2032
2033 Unique_Node_List useful_predicates; // to store useful predicates
2034 if (C->has_loops()) {
2035 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2036 }
2037
2038 for (int i = C->predicate_count(); i > 0; i--) {
2039 Node * n = C->predicate_opaque1_node(i-1);
2040 assert(n->Opcode() == Op_Opaque1, "must be");
2041 if (!useful_predicates.member(n)) { // not in the useful list
2042 _igvn.replace_node(n, n->in(1));
2043 }
2044 }
2045 }
2046
2047 //------------------------process_expensive_nodes-----------------------------
2048 // Expensive nodes have their control input set to prevent the GVN
2049 // from commoning them and as a result forcing the resulting node to
2050 // be in a more frequent path. Use CFG information here, to change the
2051 // control inputs so that some expensive nodes can be commoned while
2052 // not executed more frequently.
2053 bool PhaseIdealLoop::process_expensive_nodes() {
2054 assert(OptimizeExpensiveOps, "optimization off?");
2055
2056 // Sort nodes to bring similar nodes together
2057 C->sort_expensive_nodes();
2058
2059 bool progress = false;
2060
2061 for (int i = 0; i < C->expensive_count(); ) {
2062 Node* n = C->expensive_node(i);
2063 int start = i;
2064 // Find nodes similar to n
2065 i++;
2066 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2067 int end = i;
2068 // And compare them two by two
2069 for (int j = start; j < end; j++) {
2070 Node* n1 = C->expensive_node(j);
2071 if (is_node_unreachable(n1)) {
2072 continue;
2073 }
2074 for (int k = j+1; k < end; k++) {
2075 Node* n2 = C->expensive_node(k);
2076 if (is_node_unreachable(n2)) {
2077 continue;
2078 }
2079
2080 assert(n1 != n2, "should be pair of nodes");
2081
2082 Node* c1 = n1->in(0);
2083 Node* c2 = n2->in(0);
2084
2085 Node* parent_c1 = c1;
2086 Node* parent_c2 = c2;
2087
2088 // The call to get_early_ctrl_for_expensive() moves the
2089 // expensive nodes up but stops at loops that are in a if
2090 // branch. See whether we can exit the loop and move above the
2091 // If.
2092 if (c1->is_Loop()) {
2093 parent_c1 = c1->in(1);
2094 }
2095 if (c2->is_Loop()) {
2096 parent_c2 = c2->in(1);
2097 }
2098
2099 if (parent_c1 == parent_c2) {
2100 _igvn._worklist.push(n1);
2101 _igvn._worklist.push(n2);
2102 continue;
2103 }
2104
2105 // Look for identical expensive node up the dominator chain.
2106 if (is_dominator(c1, c2)) {
2107 c2 = c1;
2108 } else if (is_dominator(c2, c1)) {
2109 c1 = c2;
2110 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2111 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2112 // Both branches have the same expensive node so move it up
2113 // before the if.
2114 c1 = c2 = idom(parent_c1->in(0));
2115 }
2116 // Do the actual moves
2117 if (n1->in(0) != c1) {
2118 _igvn.hash_delete(n1);
2119 n1->set_req(0, c1);
2120 _igvn.hash_insert(n1);
2121 _igvn._worklist.push(n1);
2122 progress = true;
2123 }
2124 if (n2->in(0) != c2) {
2125 _igvn.hash_delete(n2);
2126 n2->set_req(0, c2);
2127 _igvn.hash_insert(n2);
2128 _igvn._worklist.push(n2);
2129 progress = true;
2130 }
2131 }
2132 }
2133 }
2134
2135 return progress;
2136 }
2137
2138
2139 //=============================================================================
2140 //----------------------------build_and_optimize-------------------------------
2141 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2142 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2143 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
2144 ResourceMark rm;
2145
2146 int old_progress = C->major_progress();
2147 uint orig_worklist_size = _igvn._worklist.size();
2148
2149 // Reset major-progress flag for the driver's heuristics
2150 C->clear_major_progress();
2151
2152 #ifndef PRODUCT
2153 // Capture for later assert
2154 uint unique = C->unique();
2155 _loop_invokes++;
2156 _loop_work += unique;
2157 #endif
2158
2159 // True if the method has at least 1 irreducible loop
2160 _has_irreducible_loops = false;
2161
2162 _created_loop_node = false;
2163
2164 Arena *a = Thread::current()->resource_area();
2165 VectorSet visited(a);
2166 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2167 _nodes.map(C->unique(), NULL);
2168 memset(_nodes.adr(), 0, wordSize * C->unique());
2169
2170 // Pre-build the top-level outermost loop tree entry
2171 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2172 // Do not need a safepoint at the top level
2173 _ltree_root->_has_sfpt = 1;
2174
2175 // Initialize Dominators.
2176 // Checked in clone_loop_predicate() during beautify_loops().
2177 _idom_size = 0;
2178 _idom = NULL;
2179 _dom_depth = NULL;
2180 _dom_stk = NULL;
2181
2182 // Empty pre-order array
2183 allocate_preorders();
2184
2185 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2186 // IdealLoopTree entries. Data nodes are NOT walked.
2187 build_loop_tree();
2188 // Check for bailout, and return
2189 if (C->failing()) {
2190 return;
2191 }
2192
2193 // No loops after all
2194 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2195
2196 // There should always be an outer loop containing the Root and Return nodes.
2197 // If not, we have a degenerate empty program. Bail out in this case.
2198 if (!has_node(C->root())) {
2199 if (!_verify_only) {
2200 C->clear_major_progress();
2201 C->record_method_not_compilable("empty program detected during loop optimization");
2202 }
2203 return;
2204 }
2205
2206 // Nothing to do, so get out
2207 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
2208 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2209 if (stop_early && !do_expensive_nodes) {
2210 _igvn.optimize(); // Cleanup NeverBranches
2211 return;
2212 }
2213
2214 // Set loop nesting depth
2215 _ltree_root->set_nest( 0 );
2216
2217 // Split shared headers and insert loop landing pads.
2218 // Do not bother doing this on the Root loop of course.
2219 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2220 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2221 if( _ltree_root->_child->beautify_loops( this ) ) {
2222 // Re-build loop tree!
2223 _ltree_root->_child = NULL;
2224 _nodes.clear();
2225 reallocate_preorders();
2226 build_loop_tree();
2227 // Check for bailout, and return
2228 if (C->failing()) {
2229 return;
2230 }
2231 // Reset loop nesting depth
2232 _ltree_root->set_nest( 0 );
2233
2234 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2235 }
2236 }
2237
2238 // Build Dominators for elision of NULL checks & loop finding.
2239 // Since nodes do not have a slot for immediate dominator, make
2240 // a persistent side array for that info indexed on node->_idx.
2241 _idom_size = C->unique();
2242 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2243 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2244 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2245 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2246
2247 Dominators();
2248
2249 if (!_verify_only) {
2250 // As a side effect, Dominators removed any unreachable CFG paths
2251 // into RegionNodes. It doesn't do this test against Root, so
2252 // we do it here.
2253 for( uint i = 1; i < C->root()->req(); i++ ) {
2254 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2255 _igvn.delete_input_of(C->root(), i);
2256 i--; // Rerun same iteration on compressed edges
2257 }
2258 }
2259
2260 // Given dominators, try to find inner loops with calls that must
2261 // always be executed (call dominates loop tail). These loops do
2262 // not need a separate safepoint.
2263 Node_List cisstack(a);
2264 _ltree_root->check_safepts(visited, cisstack);
2265 }
2266
2267 // Walk the DATA nodes and place into loops. Find earliest control
2268 // node. For CFG nodes, the _nodes array starts out and remains
2269 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2270 // _nodes array holds the earliest legal controlling CFG node.
2271
2272 // Allocate stack with enough space to avoid frequent realloc
2273 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2274 Node_Stack nstack( a, stack_size );
2275
2276 visited.Clear();
2277 Node_List worklist(a);
2278 // Don't need C->root() on worklist since
2279 // it will be processed among C->top() inputs
2280 worklist.push( C->top() );
2281 visited.set( C->top()->_idx ); // Set C->top() as visited now
2282 build_loop_early( visited, worklist, nstack );
2283
2284 // Given early legal placement, try finding counted loops. This placement
2285 // is good enough to discover most loop invariants.
2286 if( !_verify_me && !_verify_only )
2287 _ltree_root->counted_loop( this );
2288
2289 // Find latest loop placement. Find ideal loop placement.
2290 visited.Clear();
2291 init_dom_lca_tags();
2292 // Need C->root() on worklist when processing outs
2293 worklist.push( C->root() );
2294 NOT_PRODUCT( C->verify_graph_edges(); )
2295 worklist.push( C->top() );
2296 build_loop_late( visited, worklist, nstack );
2297
2298 if (_verify_only) {
2299 // restore major progress flag
2300 for (int i = 0; i < old_progress; i++)
2301 C->set_major_progress();
2302 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2303 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2304 return;
2305 }
2306
2307 // clear out the dead code after build_loop_late
2308 while (_deadlist.size()) {
2309 _igvn.remove_globally_dead_node(_deadlist.pop());
2310 }
2311
2312 if (stop_early) {
2313 assert(do_expensive_nodes, "why are we here?");
2314 if (process_expensive_nodes()) {
2315 // If we made some progress when processing expensive nodes then
2316 // the IGVN may modify the graph in a way that will allow us to
2317 // make some more progress: we need to try processing expensive
2318 // nodes again.
2319 C->set_major_progress();
2320 }
2321 _igvn.optimize();
2322 return;
2323 }
2324
2325 // Some parser-inserted loop predicates could never be used by loop
2326 // predication or they were moved away from loop during some optimizations.
2327 // For example, peeling. Eliminate them before next loop optimizations.
2328 if (UseLoopPredicate || LoopLimitCheck) {
2329 eliminate_useless_predicates();
2330 }
2331
2332 #ifndef PRODUCT
2333 C->verify_graph_edges();
2334 if (_verify_me) { // Nested verify pass?
2335 // Check to see if the verify mode is broken
2336 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2337 return;
2338 }
2339 if(VerifyLoopOptimizations) verify();
2340 if(TraceLoopOpts && C->has_loops()) {
2341 _ltree_root->dump();
2342 }
2343 #endif
2344
2345 if (skip_loop_opts) {
2346 // restore major progress flag
2347 for (int i = 0; i < old_progress; i++) {
2348 C->set_major_progress();
2349 }
2350
2351 // Cleanup any modified bits
2352 _igvn.optimize();
2353
2354 if (C->log() != NULL) {
2355 log_loop_tree(_ltree_root, _ltree_root, C->log());
2356 }
2357 return;
2358 }
2359
2360 if (ReassociateInvariants) {
2361 // Reassociate invariants and prep for split_thru_phi
2362 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2363 IdealLoopTree* lpt = iter.current();
2364 bool is_counted = lpt->is_counted();
2365 if (!is_counted || !lpt->is_inner()) continue;
2366
2367 // check for vectorized loops, any reassociation of invariants was already done
2368 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2369
2370 lpt->reassociate_invariants(this);
2371
2372 // Because RCE opportunities can be masked by split_thru_phi,
2373 // look for RCE candidates and inhibit split_thru_phi
2374 // on just their loop-phi's for this pass of loop opts
2375 if (SplitIfBlocks && do_split_ifs) {
2376 if (lpt->policy_range_check(this)) {
2377 lpt->_rce_candidate = 1; // = true
2378 }
2379 }
2380 }
2381 }
2382
2383 // Check for aggressive application of split-if and other transforms
2384 // that require basic-block info (like cloning through Phi's)
2385 if( SplitIfBlocks && do_split_ifs ) {
2386 visited.Clear();
2387 split_if_with_blocks( visited, nstack );
2388 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2389 }
2390
2391 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2392 C->set_major_progress();
2393 }
2394
2395 // Perform loop predication before iteration splitting
2396 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2397 _ltree_root->_child->loop_predication(this);
2398 }
2399
2400 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2401 if (do_intrinsify_fill()) {
2402 C->set_major_progress();
2403 }
2404 }
2405
2406 // Perform iteration-splitting on inner loops. Split iterations to avoid
2407 // range checks or one-shot null checks.
2408
2409 // If split-if's didn't hack the graph too bad (no CFG changes)
2410 // then do loop opts.
2411 if (C->has_loops() && !C->major_progress()) {
2412 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2413 _ltree_root->_child->iteration_split( this, worklist );
2414 // No verify after peeling! GCM has hoisted code out of the loop.
2415 // After peeling, the hoisted code could sink inside the peeled area.
2416 // The peeling code does not try to recompute the best location for
2417 // all the code before the peeled area, so the verify pass will always
2418 // complain about it.
2419 }
2420 // Do verify graph edges in any case
2421 NOT_PRODUCT( C->verify_graph_edges(); );
2422
2423 if (!do_split_ifs) {
2424 // We saw major progress in Split-If to get here. We forced a
2425 // pass with unrolling and not split-if, however more split-if's
2426 // might make progress. If the unrolling didn't make progress
2427 // then the major-progress flag got cleared and we won't try
2428 // another round of Split-If. In particular the ever-common
2429 // instance-of/check-cast pattern requires at least 2 rounds of
2430 // Split-If to clear out.
2431 C->set_major_progress();
2432 }
2433
2434 // Repeat loop optimizations if new loops were seen
2435 if (created_loop_node()) {
2436 C->set_major_progress();
2437 }
2438
2439 // Keep loop predicates and perform optimizations with them
2440 // until no more loop optimizations could be done.
2441 // After that switch predicates off and do more loop optimizations.
2442 if (!C->major_progress() && (C->predicate_count() > 0)) {
2443 C->cleanup_loop_predicates(_igvn);
2444 if (TraceLoopOpts) {
2445 tty->print_cr("PredicatesOff");
2446 }
2447 C->set_major_progress();
2448 }
2449
2450 // Convert scalar to superword operations at the end of all loop opts.
2451 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2452 // SuperWord transform
2453 SuperWord sw(this);
2454 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2455 IdealLoopTree* lpt = iter.current();
2456 if (lpt->is_counted()) {
2457 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
2458
2459 if (PostLoopMultiversioning && cl->is_post_loop() && !cl->is_vectorized_loop()) {
2460 // Check that the rce'd post loop is encountered first, multiversion after all
2461 // major main loop optimization are concluded
2462 has_range_checks(lpt);
2463 if (!cl->has_range_checks() && !C->major_progress()) {
2464 IdealLoopTree *lpt_next = lpt->_next;
2465 if (lpt_next && lpt_next->is_counted()) {
2466 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
2467 has_range_checks(lpt_next);
2468 if (cl->is_post_loop() && cl->has_range_checks()) {
2469 if (!cl->is_multiversioned()) {
2470 if (multi_version_post_loops(lpt, lpt_next) == false) {
2471 // Cause the rce loop to be optimized away if we fail
2472 cl->mark_is_multiversioned();
2473 poison_rce_post_loop(lpt);
2474 }
2475 }
2476 }
2477 }
2478 }
2479 } else if (cl->is_main_loop()) {
2480 sw.transform_loop(lpt, true);
2481 }
2482 }
2483 }
2484 }
2485
2486 // Cleanup any modified bits
2487 _igvn.optimize();
2488
2489 // disable assert until issue with split_flow_path is resolved (6742111)
2490 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2491 // "shouldn't introduce irreducible loops");
2492
2493 if (C->log() != NULL) {
2494 log_loop_tree(_ltree_root, _ltree_root, C->log());
2495 }
2496 }
2497
2498 #ifndef PRODUCT
2499 //------------------------------print_statistics-------------------------------
2500 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2501 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2502 void PhaseIdealLoop::print_statistics() {
2503 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2504 }
2505
2506 //------------------------------verify-----------------------------------------
2507 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2508 static int fail; // debug only, so its multi-thread dont care
2509 void PhaseIdealLoop::verify() const {
2510 int old_progress = C->major_progress();
2511 ResourceMark rm;
2512 PhaseIdealLoop loop_verify( _igvn, this );
2513 VectorSet visited(Thread::current()->resource_area());
2514
2515 fail = 0;
2516 verify_compare( C->root(), &loop_verify, visited );
2517 assert( fail == 0, "verify loops failed" );
2518 // Verify loop structure is the same
2519 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2520 // Reset major-progress. It was cleared by creating a verify version of
2521 // PhaseIdealLoop.
2522 for( int i=0; i<old_progress; i++ )
2523 C->set_major_progress();
2524 }
2525
2526 //------------------------------verify_compare---------------------------------
2527 // Make sure me and the given PhaseIdealLoop agree on key data structures
2528 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2529 if( !n ) return;
2530 if( visited.test_set( n->_idx ) ) return;
2531 if( !_nodes[n->_idx] ) { // Unreachable
2532 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2533 return;
2534 }
2535
2536 uint i;
2537 for( i = 0; i < n->req(); i++ )
2538 verify_compare( n->in(i), loop_verify, visited );
2539
2540 // Check the '_nodes' block/loop structure
2541 i = n->_idx;
2542 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2543 if( _nodes[i] != loop_verify->_nodes[i] &&
2544 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2545 tty->print("Mismatched control setting for: ");
2546 n->dump();
2547 if( fail++ > 10 ) return;
2548 Node *c = get_ctrl_no_update(n);
2549 tty->print("We have it as: ");
2550 if( c->in(0) ) c->dump();
2551 else tty->print_cr("N%d",c->_idx);
2552 tty->print("Verify thinks: ");
2553 if( loop_verify->has_ctrl(n) )
2554 loop_verify->get_ctrl_no_update(n)->dump();
2555 else
2556 loop_verify->get_loop_idx(n)->dump();
2557 tty->cr();
2558 }
2559 } else { // We have a loop
2560 IdealLoopTree *us = get_loop_idx(n);
2561 if( loop_verify->has_ctrl(n) ) {
2562 tty->print("Mismatched loop setting for: ");
2563 n->dump();
2564 if( fail++ > 10 ) return;
2565 tty->print("We have it as: ");
2566 us->dump();
2567 tty->print("Verify thinks: ");
2568 loop_verify->get_ctrl_no_update(n)->dump();
2569 tty->cr();
2570 } else if (!C->major_progress()) {
2571 // Loop selection can be messed up if we did a major progress
2572 // operation, like split-if. Do not verify in that case.
2573 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2574 if( us->_head != them->_head || us->_tail != them->_tail ) {
2575 tty->print("Unequals loops for: ");
2576 n->dump();
2577 if( fail++ > 10 ) return;
2578 tty->print("We have it as: ");
2579 us->dump();
2580 tty->print("Verify thinks: ");
2581 them->dump();
2582 tty->cr();
2583 }
2584 }
2585 }
2586
2587 // Check for immediate dominators being equal
2588 if( i >= _idom_size ) {
2589 if( !n->is_CFG() ) return;
2590 tty->print("CFG Node with no idom: ");
2591 n->dump();
2592 return;
2593 }
2594 if( !n->is_CFG() ) return;
2595 if( n == C->root() ) return; // No IDOM here
2596
2597 assert(n->_idx == i, "sanity");
2598 Node *id = idom_no_update(n);
2599 if( id != loop_verify->idom_no_update(n) ) {
2600 tty->print("Unequals idoms for: ");
2601 n->dump();
2602 if( fail++ > 10 ) return;
2603 tty->print("We have it as: ");
2604 id->dump();
2605 tty->print("Verify thinks: ");
2606 loop_verify->idom_no_update(n)->dump();
2607 tty->cr();
2608 }
2609
2610 }
2611
2612 //------------------------------verify_tree------------------------------------
2613 // Verify that tree structures match. Because the CFG can change, siblings
2614 // within the loop tree can be reordered. We attempt to deal with that by
2615 // reordering the verify's loop tree if possible.
2616 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2617 assert( _parent == parent, "Badly formed loop tree" );
2618
2619 // Siblings not in same order? Attempt to re-order.
2620 if( _head != loop->_head ) {
2621 // Find _next pointer to update
2622 IdealLoopTree **pp = &loop->_parent->_child;
2623 while( *pp != loop )
2624 pp = &((*pp)->_next);
2625 // Find proper sibling to be next
2626 IdealLoopTree **nn = &loop->_next;
2627 while( (*nn) && (*nn)->_head != _head )
2628 nn = &((*nn)->_next);
2629
2630 // Check for no match.
2631 if( !(*nn) ) {
2632 // Annoyingly, irreducible loops can pick different headers
2633 // after a major_progress operation, so the rest of the loop
2634 // tree cannot be matched.
2635 if (_irreducible && Compile::current()->major_progress()) return;
2636 assert( 0, "failed to match loop tree" );
2637 }
2638
2639 // Move (*nn) to (*pp)
2640 IdealLoopTree *hit = *nn;
2641 *nn = hit->_next;
2642 hit->_next = loop;
2643 *pp = loop;
2644 loop = hit;
2645 // Now try again to verify
2646 }
2647
2648 assert( _head == loop->_head , "mismatched loop head" );
2649 Node *tail = _tail; // Inline a non-updating version of
2650 while( !tail->in(0) ) // the 'tail()' call.
2651 tail = tail->in(1);
2652 assert( tail == loop->_tail, "mismatched loop tail" );
2653
2654 // Counted loops that are guarded should be able to find their guards
2655 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2656 CountedLoopNode *cl = _head->as_CountedLoop();
2657 Node *init = cl->init_trip();
2658 Node *ctrl = cl->in(LoopNode::EntryControl);
2659 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2660 Node *iff = ctrl->in(0);
2661 assert( iff->Opcode() == Op_If, "" );
2662 Node *bol = iff->in(1);
2663 assert( bol->Opcode() == Op_Bool, "" );
2664 Node *cmp = bol->in(1);
2665 assert( cmp->Opcode() == Op_CmpI, "" );
2666 Node *add = cmp->in(1);
2667 Node *opaq;
2668 if( add->Opcode() == Op_Opaque1 ) {
2669 opaq = add;
2670 } else {
2671 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2672 assert( add == init, "" );
2673 opaq = cmp->in(2);
2674 }
2675 assert( opaq->Opcode() == Op_Opaque1, "" );
2676
2677 }
2678
2679 if (_child != NULL) _child->verify_tree(loop->_child, this);
2680 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2681 // Innermost loops need to verify loop bodies,
2682 // but only if no 'major_progress'
2683 int fail = 0;
2684 if (!Compile::current()->major_progress() && _child == NULL) {
2685 for( uint i = 0; i < _body.size(); i++ ) {
2686 Node *n = _body.at(i);
2687 if (n->outcnt() == 0) continue; // Ignore dead
2688 uint j;
2689 for( j = 0; j < loop->_body.size(); j++ )
2690 if( loop->_body.at(j) == n )
2691 break;
2692 if( j == loop->_body.size() ) { // Not found in loop body
2693 // Last ditch effort to avoid assertion: Its possible that we
2694 // have some users (so outcnt not zero) but are still dead.
2695 // Try to find from root.
2696 if (Compile::current()->root()->find(n->_idx)) {
2697 fail++;
2698 tty->print("We have that verify does not: ");
2699 n->dump();
2700 }
2701 }
2702 }
2703 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2704 Node *n = loop->_body.at(i2);
2705 if (n->outcnt() == 0) continue; // Ignore dead
2706 uint j;
2707 for( j = 0; j < _body.size(); j++ )
2708 if( _body.at(j) == n )
2709 break;
2710 if( j == _body.size() ) { // Not found in loop body
2711 // Last ditch effort to avoid assertion: Its possible that we
2712 // have some users (so outcnt not zero) but are still dead.
2713 // Try to find from root.
2714 if (Compile::current()->root()->find(n->_idx)) {
2715 fail++;
2716 tty->print("Verify has that we do not: ");
2717 n->dump();
2718 }
2719 }
2720 }
2721 assert( !fail, "loop body mismatch" );
2722 }
2723 }
2724
2725 #endif
2726
2727 //------------------------------set_idom---------------------------------------
2728 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2729 uint idx = d->_idx;
2730 if (idx >= _idom_size) {
2731 uint newsize = _idom_size<<1;
2732 while( idx >= newsize ) {
2733 newsize <<= 1;
2734 }
2735 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2736 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2737 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2738 _idom_size = newsize;
2739 }
2740 _idom[idx] = n;
2741 _dom_depth[idx] = dom_depth;
2742 }
2743
2744 //------------------------------recompute_dom_depth---------------------------------------
2745 // The dominator tree is constructed with only parent pointers.
2746 // This recomputes the depth in the tree by first tagging all
2747 // nodes as "no depth yet" marker. The next pass then runs up
2748 // the dom tree from each node marked "no depth yet", and computes
2749 // the depth on the way back down.
2750 void PhaseIdealLoop::recompute_dom_depth() {
2751 uint no_depth_marker = C->unique();
2752 uint i;
2753 // Initialize depth to "no depth yet"
2754 for (i = 0; i < _idom_size; i++) {
2755 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2756 _dom_depth[i] = no_depth_marker;
2757 }
2758 }
2759 if (_dom_stk == NULL) {
2760 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
2761 if (init_size < 10) init_size = 10;
2762 _dom_stk = new GrowableArray<uint>(init_size);
2763 }
2764 // Compute new depth for each node.
2765 for (i = 0; i < _idom_size; i++) {
2766 uint j = i;
2767 // Run up the dom tree to find a node with a depth
2768 while (_dom_depth[j] == no_depth_marker) {
2769 _dom_stk->push(j);
2770 j = _idom[j]->_idx;
2771 }
2772 // Compute the depth on the way back down this tree branch
2773 uint dd = _dom_depth[j] + 1;
2774 while (_dom_stk->length() > 0) {
2775 uint j = _dom_stk->pop();
2776 _dom_depth[j] = dd;
2777 dd++;
2778 }
2779 }
2780 }
2781
2782 //------------------------------sort-------------------------------------------
2783 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2784 // loop tree, not the root.
2785 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2786 if( !innermost ) return loop; // New innermost loop
2787
2788 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2789 assert( loop_preorder, "not yet post-walked loop" );
2790 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2791 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2792
2793 // Insert at start of list
2794 while( l ) { // Insertion sort based on pre-order
2795 if( l == loop ) return innermost; // Already on list!
2796 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2797 assert( l_preorder, "not yet post-walked l" );
2798 // Check header pre-order number to figure proper nesting
2799 if( loop_preorder > l_preorder )
2800 break; // End of insertion
2801 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2802 // Since I split shared headers, you'd think this could not happen.
2803 // BUT: I must first do the preorder numbering before I can discover I
2804 // have shared headers, so the split headers all get the same preorder
2805 // number as the RegionNode they split from.
2806 if( loop_preorder == l_preorder &&
2807 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2808 break; // Also check for shared headers (same pre#)
2809 pp = &l->_parent; // Chain up list
2810 l = *pp;
2811 }
2812 // Link into list
2813 // Point predecessor to me
2814 *pp = loop;
2815 // Point me to successor
2816 IdealLoopTree *p = loop->_parent;
2817 loop->_parent = l; // Point me to successor
2818 if( p ) sort( p, innermost ); // Insert my parents into list as well
2819 return innermost;
2820 }
2821
2822 //------------------------------build_loop_tree--------------------------------
2823 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2824 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2825 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2826 // tightest enclosing IdealLoopTree for post-walked.
2827 //
2828 // During my forward walk I do a short 1-layer lookahead to see if I can find
2829 // a loop backedge with that doesn't have any work on the backedge. This
2830 // helps me construct nested loops with shared headers better.
2831 //
2832 // Once I've done the forward recursion, I do the post-work. For each child
2833 // I check to see if there is a backedge. Backedges define a loop! I
2834 // insert an IdealLoopTree at the target of the backedge.
2835 //
2836 // During the post-work I also check to see if I have several children
2837 // belonging to different loops. If so, then this Node is a decision point
2838 // where control flow can choose to change loop nests. It is at this
2839 // decision point where I can figure out how loops are nested. At this
2840 // time I can properly order the different loop nests from my children.
2841 // Note that there may not be any backedges at the decision point!
2842 //
2843 // Since the decision point can be far removed from the backedges, I can't
2844 // order my loops at the time I discover them. Thus at the decision point
2845 // I need to inspect loop header pre-order numbers to properly nest my
2846 // loops. This means I need to sort my childrens' loops by pre-order.
2847 // The sort is of size number-of-control-children, which generally limits
2848 // it to size 2 (i.e., I just choose between my 2 target loops).
2849 void PhaseIdealLoop::build_loop_tree() {
2850 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
2851 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
2852 Node *n = C->root();
2853 bltstack.push(n);
2854 int pre_order = 1;
2855 int stack_size;
2856
2857 while ( ( stack_size = bltstack.length() ) != 0 ) {
2858 n = bltstack.top(); // Leave node on stack
2859 if ( !is_visited(n) ) {
2860 // ---- Pre-pass Work ----
2861 // Pre-walked but not post-walked nodes need a pre_order number.
2862
2863 set_preorder_visited( n, pre_order ); // set as visited
2864
2865 // ---- Scan over children ----
2866 // Scan first over control projections that lead to loop headers.
2867 // This helps us find inner-to-outer loops with shared headers better.
2868
2869 // Scan children's children for loop headers.
2870 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2871 Node* m = n->raw_out(i); // Child
2872 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2873 // Scan over children's children to find loop
2874 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2875 Node* l = m->fast_out(j);
2876 if( is_visited(l) && // Been visited?
2877 !is_postvisited(l) && // But not post-visited
2878 get_preorder(l) < pre_order ) { // And smaller pre-order
2879 // Found! Scan the DFS down this path before doing other paths
2880 bltstack.push(m);
2881 break;
2882 }
2883 }
2884 }
2885 }
2886 pre_order++;
2887 }
2888 else if ( !is_postvisited(n) ) {
2889 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2890 // such as com.sun.rsasign.am::a.
2891 // For non-recursive version, first, process current children.
2892 // On next iteration, check if additional children were added.
2893 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2894 Node* u = n->raw_out(k);
2895 if ( u->is_CFG() && !is_visited(u) ) {
2896 bltstack.push(u);
2897 }
2898 }
2899 if ( bltstack.length() == stack_size ) {
2900 // There were no additional children, post visit node now
2901 (void)bltstack.pop(); // Remove node from stack
2902 pre_order = build_loop_tree_impl( n, pre_order );
2903 // Check for bailout
2904 if (C->failing()) {
2905 return;
2906 }
2907 // Check to grow _preorders[] array for the case when
2908 // build_loop_tree_impl() adds new nodes.
2909 check_grow_preorders();
2910 }
2911 }
2912 else {
2913 (void)bltstack.pop(); // Remove post-visited node from stack
2914 }
2915 }
2916 }
2917
2918 //------------------------------build_loop_tree_impl---------------------------
2919 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2920 // ---- Post-pass Work ----
2921 // Pre-walked but not post-walked nodes need a pre_order number.
2922
2923 // Tightest enclosing loop for this Node
2924 IdealLoopTree *innermost = NULL;
2925
2926 // For all children, see if any edge is a backedge. If so, make a loop
2927 // for it. Then find the tightest enclosing loop for the self Node.
2928 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2929 Node* m = n->fast_out(i); // Child
2930 if( n == m ) continue; // Ignore control self-cycles
2931 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2932
2933 IdealLoopTree *l; // Child's loop
2934 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2935 // Found a backedge
2936 assert( get_preorder(m) < pre_order, "should be backedge" );
2937 // Check for the RootNode, which is already a LoopNode and is allowed
2938 // to have multiple "backedges".
2939 if( m == C->root()) { // Found the root?
2940 l = _ltree_root; // Root is the outermost LoopNode
2941 } else { // Else found a nested loop
2942 // Insert a LoopNode to mark this loop.
2943 l = new IdealLoopTree(this, m, n);
2944 } // End of Else found a nested loop
2945 if( !has_loop(m) ) // If 'm' does not already have a loop set
2946 set_loop(m, l); // Set loop header to loop now
2947
2948 } else { // Else not a nested loop
2949 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2950 l = get_loop(m); // Get previously determined loop
2951 // If successor is header of a loop (nest), move up-loop till it
2952 // is a member of some outer enclosing loop. Since there are no
2953 // shared headers (I've split them already) I only need to go up
2954 // at most 1 level.
2955 while( l && l->_head == m ) // Successor heads loop?
2956 l = l->_parent; // Move up 1 for me
2957 // If this loop is not properly parented, then this loop
2958 // has no exit path out, i.e. its an infinite loop.
2959 if( !l ) {
2960 // Make loop "reachable" from root so the CFG is reachable. Basically
2961 // insert a bogus loop exit that is never taken. 'm', the loop head,
2962 // points to 'n', one (of possibly many) fall-in paths. There may be
2963 // many backedges as well.
2964
2965 // Here I set the loop to be the root loop. I could have, after
2966 // inserting a bogus loop exit, restarted the recursion and found my
2967 // new loop exit. This would make the infinite loop a first-class
2968 // loop and it would then get properly optimized. What's the use of
2969 // optimizing an infinite loop?
2970 l = _ltree_root; // Oops, found infinite loop
2971
2972 if (!_verify_only) {
2973 // Insert the NeverBranch between 'm' and it's control user.
2974 NeverBranchNode *iff = new NeverBranchNode( m );
2975 _igvn.register_new_node_with_optimizer(iff);
2976 set_loop(iff, l);
2977 Node *if_t = new CProjNode( iff, 0 );
2978 _igvn.register_new_node_with_optimizer(if_t);
2979 set_loop(if_t, l);
2980
2981 Node* cfg = NULL; // Find the One True Control User of m
2982 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2983 Node* x = m->fast_out(j);
2984 if (x->is_CFG() && x != m && x != iff)
2985 { cfg = x; break; }
2986 }
2987 assert(cfg != NULL, "must find the control user of m");
2988 uint k = 0; // Probably cfg->in(0)
2989 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2990 cfg->set_req( k, if_t ); // Now point to NeverBranch
2991 _igvn._worklist.push(cfg);
2992
2993 // Now create the never-taken loop exit
2994 Node *if_f = new CProjNode( iff, 1 );
2995 _igvn.register_new_node_with_optimizer(if_f);
2996 set_loop(if_f, l);
2997 // Find frame ptr for Halt. Relies on the optimizer
2998 // V-N'ing. Easier and quicker than searching through
2999 // the program structure.
3000 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3001 _igvn.register_new_node_with_optimizer(frame);
3002 // Halt & Catch Fire
3003 Node *halt = new HaltNode( if_f, frame );
3004 _igvn.register_new_node_with_optimizer(halt);
3005 set_loop(halt, l);
3006 C->root()->add_req(halt);
3007 }
3008 set_loop(C->root(), _ltree_root);
3009 }
3010 }
3011 // Weeny check for irreducible. This child was already visited (this
3012 // IS the post-work phase). Is this child's loop header post-visited
3013 // as well? If so, then I found another entry into the loop.
3014 if (!_verify_only) {
3015 while( is_postvisited(l->_head) ) {
3016 // found irreducible
3017 l->_irreducible = 1; // = true
3018 l = l->_parent;
3019 _has_irreducible_loops = true;
3020 // Check for bad CFG here to prevent crash, and bailout of compile
3021 if (l == NULL) {
3022 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3023 return pre_order;
3024 }
3025 }
3026 C->set_has_irreducible_loop(_has_irreducible_loops);
3027 }
3028
3029 // This Node might be a decision point for loops. It is only if
3030 // it's children belong to several different loops. The sort call
3031 // does a trivial amount of work if there is only 1 child or all
3032 // children belong to the same loop. If however, the children
3033 // belong to different loops, the sort call will properly set the
3034 // _parent pointers to show how the loops nest.
3035 //
3036 // In any case, it returns the tightest enclosing loop.
3037 innermost = sort( l, innermost );
3038 }
3039
3040 // Def-use info will have some dead stuff; dead stuff will have no
3041 // loop decided on.
3042
3043 // Am I a loop header? If so fix up my parent's child and next ptrs.
3044 if( innermost && innermost->_head == n ) {
3045 assert( get_loop(n) == innermost, "" );
3046 IdealLoopTree *p = innermost->_parent;
3047 IdealLoopTree *l = innermost;
3048 while( p && l->_head == n ) {
3049 l->_next = p->_child; // Put self on parents 'next child'
3050 p->_child = l; // Make self as first child of parent
3051 l = p; // Now walk up the parent chain
3052 p = l->_parent;
3053 }
3054 } else {
3055 // Note that it is possible for a LoopNode to reach here, if the
3056 // backedge has been made unreachable (hence the LoopNode no longer
3057 // denotes a Loop, and will eventually be removed).
3058
3059 // Record tightest enclosing loop for self. Mark as post-visited.
3060 set_loop(n, innermost);
3061 // Also record has_call flag early on
3062 if( innermost ) {
3063 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3064 // Do not count uncommon calls
3065 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3066 Node *iff = n->in(0)->in(0);
3067 // No any calls for vectorized loops.
3068 if( UseSuperWord || !iff->is_If() ||
3069 (n->in(0)->Opcode() == Op_IfFalse &&
3070 (1.0 - iff->as_If()->_prob) >= 0.01) ||
3071 (iff->as_If()->_prob >= 0.01) )
3072 innermost->_has_call = 1;
3073 }
3074 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3075 // Disable loop optimizations if the loop has a scalar replaceable
3076 // allocation. This disabling may cause a potential performance lost
3077 // if the allocation is not eliminated for some reason.
3078 innermost->_allow_optimizations = false;
3079 innermost->_has_call = 1; // = true
3080 } else if (n->Opcode() == Op_SafePoint) {
3081 // Record all safepoints in this loop.
3082 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3083 innermost->_safepts->push(n);
3084 }
3085 }
3086 }
3087
3088 // Flag as post-visited now
3089 set_postvisited(n);
3090 return pre_order;
3091 }
3092
3093
3094 //------------------------------build_loop_early-------------------------------
3095 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3096 // First pass computes the earliest controlling node possible. This is the
3097 // controlling input with the deepest dominating depth.
3098 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3099 while (worklist.size() != 0) {
3100 // Use local variables nstack_top_n & nstack_top_i to cache values
3101 // on nstack's top.
3102 Node *nstack_top_n = worklist.pop();
3103 uint nstack_top_i = 0;
3104 //while_nstack_nonempty:
3105 while (true) {
3106 // Get parent node and next input's index from stack's top.
3107 Node *n = nstack_top_n;
3108 uint i = nstack_top_i;
3109 uint cnt = n->req(); // Count of inputs
3110 if (i == 0) { // Pre-process the node.
3111 if( has_node(n) && // Have either loop or control already?
3112 !has_ctrl(n) ) { // Have loop picked out already?
3113 // During "merge_many_backedges" we fold up several nested loops
3114 // into a single loop. This makes the members of the original
3115 // loop bodies pointing to dead loops; they need to move up
3116 // to the new UNION'd larger loop. I set the _head field of these
3117 // dead loops to NULL and the _parent field points to the owning
3118 // loop. Shades of UNION-FIND algorithm.
3119 IdealLoopTree *ilt;
3120 while( !(ilt = get_loop(n))->_head ) {
3121 // Normally I would use a set_loop here. But in this one special
3122 // case, it is legal (and expected) to change what loop a Node
3123 // belongs to.
3124 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3125 }
3126 // Remove safepoints ONLY if I've already seen I don't need one.
3127 // (the old code here would yank a 2nd safepoint after seeing a
3128 // first one, even though the 1st did not dominate in the loop body
3129 // and thus could be avoided indefinitely)
3130 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3131 is_deleteable_safept(n)) {
3132 Node *in = n->in(TypeFunc::Control);
3133 lazy_replace(n,in); // Pull safepoint now
3134 if (ilt->_safepts != NULL) {
3135 ilt->_safepts->yank(n);
3136 }
3137 // Carry on with the recursion "as if" we are walking
3138 // only the control input
3139 if( !visited.test_set( in->_idx ) ) {
3140 worklist.push(in); // Visit this guy later, using worklist
3141 }
3142 // Get next node from nstack:
3143 // - skip n's inputs processing by setting i > cnt;
3144 // - we also will not call set_early_ctrl(n) since
3145 // has_node(n) == true (see the condition above).
3146 i = cnt + 1;
3147 }
3148 }
3149 } // if (i == 0)
3150
3151 // Visit all inputs
3152 bool done = true; // Assume all n's inputs will be processed
3153 while (i < cnt) {
3154 Node *in = n->in(i);
3155 ++i;
3156 if (in == NULL) continue;
3157 if (in->pinned() && !in->is_CFG())
3158 set_ctrl(in, in->in(0));
3159 int is_visited = visited.test_set( in->_idx );
3160 if (!has_node(in)) { // No controlling input yet?
3161 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3162 assert( !is_visited, "visit only once" );
3163 nstack.push(n, i); // Save parent node and next input's index.
3164 nstack_top_n = in; // Process current input now.
3165 nstack_top_i = 0;
3166 done = false; // Not all n's inputs processed.
3167 break; // continue while_nstack_nonempty;
3168 } else if (!is_visited) {
3169 // This guy has a location picked out for him, but has not yet
3170 // been visited. Happens to all CFG nodes, for instance.
3171 // Visit him using the worklist instead of recursion, to break
3172 // cycles. Since he has a location already we do not need to
3173 // find his location before proceeding with the current Node.
3174 worklist.push(in); // Visit this guy later, using worklist
3175 }
3176 }
3177 if (done) {
3178 // All of n's inputs have been processed, complete post-processing.
3179
3180 // Compute earliest point this Node can go.
3181 // CFG, Phi, pinned nodes already know their controlling input.
3182 if (!has_node(n)) {
3183 // Record earliest legal location
3184 set_early_ctrl( n );
3185 }
3186 if (nstack.is_empty()) {
3187 // Finished all nodes on stack.
3188 // Process next node on the worklist.
3189 break;
3190 }
3191 // Get saved parent node and next input's index.
3192 nstack_top_n = nstack.node();
3193 nstack_top_i = nstack.index();
3194 nstack.pop();
3195 }
3196 } // while (true)
3197 }
3198 }
3199
3200 //------------------------------dom_lca_internal--------------------------------
3201 // Pair-wise LCA
3202 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3203 if( !n1 ) return n2; // Handle NULL original LCA
3204 assert( n1->is_CFG(), "" );
3205 assert( n2->is_CFG(), "" );
3206 // find LCA of all uses
3207 uint d1 = dom_depth(n1);
3208 uint d2 = dom_depth(n2);
3209 while (n1 != n2) {
3210 if (d1 > d2) {
3211 n1 = idom(n1);
3212 d1 = dom_depth(n1);
3213 } else if (d1 < d2) {
3214 n2 = idom(n2);
3215 d2 = dom_depth(n2);
3216 } else {
3217 // Here d1 == d2. Due to edits of the dominator-tree, sections
3218 // of the tree might have the same depth. These sections have
3219 // to be searched more carefully.
3220
3221 // Scan up all the n1's with equal depth, looking for n2.
3222 Node *t1 = idom(n1);
3223 while (dom_depth(t1) == d1) {
3224 if (t1 == n2) return n2;
3225 t1 = idom(t1);
3226 }
3227 // Scan up all the n2's with equal depth, looking for n1.
3228 Node *t2 = idom(n2);
3229 while (dom_depth(t2) == d2) {
3230 if (t2 == n1) return n1;
3231 t2 = idom(t2);
3232 }
3233 // Move up to a new dominator-depth value as well as up the dom-tree.
3234 n1 = t1;
3235 n2 = t2;
3236 d1 = dom_depth(n1);
3237 d2 = dom_depth(n2);
3238 }
3239 }
3240 return n1;
3241 }
3242
3243 //------------------------------compute_idom-----------------------------------
3244 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3245 // IDOMs are correct.
3246 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3247 assert( region->is_Region(), "" );
3248 Node *LCA = NULL;
3249 for( uint i = 1; i < region->req(); i++ ) {
3250 if( region->in(i) != C->top() )
3251 LCA = dom_lca( LCA, region->in(i) );
3252 }
3253 return LCA;
3254 }
3255
3256 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3257 bool had_error = false;
3258 #ifdef ASSERT
3259 if (early != C->root()) {
3260 // Make sure that there's a dominance path from LCA to early
3261 Node* d = LCA;
3262 while (d != early) {
3263 if (d == C->root()) {
3264 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3265 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3266 had_error = true;
3267 break;
3268 }
3269 d = idom(d);
3270 }
3271 }
3272 #endif
3273 return had_error;
3274 }
3275
3276
3277 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3278 // Compute LCA over list of uses
3279 bool had_error = false;
3280 Node *LCA = NULL;
3281 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3282 Node* c = n->fast_out(i);
3283 if (_nodes[c->_idx] == NULL)
3284 continue; // Skip the occasional dead node
3285 if( c->is_Phi() ) { // For Phis, we must land above on the path
3286 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3287 if( c->in(j) == n ) { // Found matching input?
3288 Node *use = c->in(0)->in(j);
3289 if (_verify_only && use->is_top()) continue;
3290 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3291 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3292 }
3293 }
3294 } else {
3295 // For CFG data-users, use is in the block just prior
3296 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3297 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3298 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3299 }
3300 }
3301 assert(!had_error, "bad dominance");
3302 return LCA;
3303 }
3304
3305 //------------------------------get_late_ctrl----------------------------------
3306 // Compute latest legal control.
3307 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3308 assert(early != NULL, "early control should not be NULL");
3309
3310 Node* LCA = compute_lca_of_uses(n, early);
3311 #ifdef ASSERT
3312 if (LCA == C->root() && LCA != early) {
3313 // def doesn't dominate uses so print some useful debugging output
3314 compute_lca_of_uses(n, early, true);
3315 }
3316 #endif
3317
3318 // if this is a load, check for anti-dependent stores
3319 // We use a conservative algorithm to identify potential interfering
3320 // instructions and for rescheduling the load. The users of the memory
3321 // input of this load are examined. Any use which is not a load and is
3322 // dominated by early is considered a potentially interfering store.
3323 // This can produce false positives.
3324 if (n->is_Load() && LCA != early) {
3325 Node_List worklist;
3326
3327 Node *mem = n->in(MemNode::Memory);
3328 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3329 Node* s = mem->fast_out(i);
3330 worklist.push(s);
3331 }
3332 while(worklist.size() != 0 && LCA != early) {
3333 Node* s = worklist.pop();
3334 if (s->is_Load()) {
3335 continue;
3336 } else if (s->is_MergeMem()) {
3337 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3338 Node* s1 = s->fast_out(i);
3339 worklist.push(s1);
3340 }
3341 } else {
3342 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3343 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3344 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3345 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3346 }
3347 }
3348 }
3349 }
3350
3351 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3352 return LCA;
3353 }
3354
3355 // true if CFG node d dominates CFG node n
3356 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3357 if (d == n)
3358 return true;
3359 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3360 uint dd = dom_depth(d);
3361 while (dom_depth(n) >= dd) {
3362 if (n == d)
3363 return true;
3364 n = idom(n);
3365 }
3366 return false;
3367 }
3368
3369 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3370 // Pair-wise LCA with tags.
3371 // Tag each index with the node 'tag' currently being processed
3372 // before advancing up the dominator chain using idom().
3373 // Later calls that find a match to 'tag' know that this path has already
3374 // been considered in the current LCA (which is input 'n1' by convention).
3375 // Since get_late_ctrl() is only called once for each node, the tag array
3376 // does not need to be cleared between calls to get_late_ctrl().
3377 // Algorithm trades a larger constant factor for better asymptotic behavior
3378 //
3379 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3380 uint d1 = dom_depth(n1);
3381 uint d2 = dom_depth(n2);
3382
3383 do {
3384 if (d1 > d2) {
3385 // current lca is deeper than n2
3386 _dom_lca_tags.map(n1->_idx, tag);
3387 n1 = idom(n1);
3388 d1 = dom_depth(n1);
3389 } else if (d1 < d2) {
3390 // n2 is deeper than current lca
3391 Node *memo = _dom_lca_tags[n2->_idx];
3392 if( memo == tag ) {
3393 return n1; // Return the current LCA
3394 }
3395 _dom_lca_tags.map(n2->_idx, tag);
3396 n2 = idom(n2);
3397 d2 = dom_depth(n2);
3398 } else {
3399 // Here d1 == d2. Due to edits of the dominator-tree, sections
3400 // of the tree might have the same depth. These sections have
3401 // to be searched more carefully.
3402
3403 // Scan up all the n1's with equal depth, looking for n2.
3404 _dom_lca_tags.map(n1->_idx, tag);
3405 Node *t1 = idom(n1);
3406 while (dom_depth(t1) == d1) {
3407 if (t1 == n2) return n2;
3408 _dom_lca_tags.map(t1->_idx, tag);
3409 t1 = idom(t1);
3410 }
3411 // Scan up all the n2's with equal depth, looking for n1.
3412 _dom_lca_tags.map(n2->_idx, tag);
3413 Node *t2 = idom(n2);
3414 while (dom_depth(t2) == d2) {
3415 if (t2 == n1) return n1;
3416 _dom_lca_tags.map(t2->_idx, tag);
3417 t2 = idom(t2);
3418 }
3419 // Move up to a new dominator-depth value as well as up the dom-tree.
3420 n1 = t1;
3421 n2 = t2;
3422 d1 = dom_depth(n1);
3423 d2 = dom_depth(n2);
3424 }
3425 } while (n1 != n2);
3426 return n1;
3427 }
3428
3429 //------------------------------init_dom_lca_tags------------------------------
3430 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3431 // Intended use does not involve any growth for the array, so it could
3432 // be of fixed size.
3433 void PhaseIdealLoop::init_dom_lca_tags() {
3434 uint limit = C->unique() + 1;
3435 _dom_lca_tags.map( limit, NULL );
3436 #ifdef ASSERT
3437 for( uint i = 0; i < limit; ++i ) {
3438 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3439 }
3440 #endif // ASSERT
3441 }
3442
3443 //------------------------------clear_dom_lca_tags------------------------------
3444 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3445 // Intended use does not involve any growth for the array, so it could
3446 // be of fixed size.
3447 void PhaseIdealLoop::clear_dom_lca_tags() {
3448 uint limit = C->unique() + 1;
3449 _dom_lca_tags.map( limit, NULL );
3450 _dom_lca_tags.clear();
3451 #ifdef ASSERT
3452 for( uint i = 0; i < limit; ++i ) {
3453 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3454 }
3455 #endif // ASSERT
3456 }
3457
3458 //------------------------------build_loop_late--------------------------------
3459 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3460 // Second pass finds latest legal placement, and ideal loop placement.
3461 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3462 while (worklist.size() != 0) {
3463 Node *n = worklist.pop();
3464 // Only visit once
3465 if (visited.test_set(n->_idx)) continue;
3466 uint cnt = n->outcnt();
3467 uint i = 0;
3468 while (true) {
3469 assert( _nodes[n->_idx], "no dead nodes" );
3470 // Visit all children
3471 if (i < cnt) {
3472 Node* use = n->raw_out(i);
3473 ++i;
3474 // Check for dead uses. Aggressively prune such junk. It might be
3475 // dead in the global sense, but still have local uses so I cannot
3476 // easily call 'remove_dead_node'.
3477 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3478 // Due to cycles, we might not hit the same fixed point in the verify
3479 // pass as we do in the regular pass. Instead, visit such phis as
3480 // simple uses of the loop head.
3481 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3482 if( !visited.test(use->_idx) )
3483 worklist.push(use);
3484 } else if( !visited.test_set(use->_idx) ) {
3485 nstack.push(n, i); // Save parent and next use's index.
3486 n = use; // Process all children of current use.
3487 cnt = use->outcnt();
3488 i = 0;
3489 }
3490 } else {
3491 // Do not visit around the backedge of loops via data edges.
3492 // push dead code onto a worklist
3493 _deadlist.push(use);
3494 }
3495 } else {
3496 // All of n's children have been processed, complete post-processing.
3497 build_loop_late_post(n);
3498 if (nstack.is_empty()) {
3499 // Finished all nodes on stack.
3500 // Process next node on the worklist.
3501 break;
3502 }
3503 // Get saved parent node and next use's index. Visit the rest of uses.
3504 n = nstack.node();
3505 cnt = n->outcnt();
3506 i = nstack.index();
3507 nstack.pop();
3508 }
3509 }
3510 }
3511 }
3512
3513 //------------------------------build_loop_late_post---------------------------
3514 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3515 // Second pass finds latest legal placement, and ideal loop placement.
3516 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3517
3518 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
3519 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3520 }
3521
3522 #ifdef ASSERT
3523 if (_verify_only && !n->is_CFG()) {
3524 // Check def-use domination.
3525 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
3526 }
3527 #endif
3528
3529 // CFG and pinned nodes already handled
3530 if( n->in(0) ) {
3531 if( n->in(0)->is_top() ) return; // Dead?
3532
3533 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3534 // _must_ be pinned (they have to observe their control edge of course).
3535 // Unlike Stores (which modify an unallocable resource, the memory
3536 // state), Mods/Loads can float around. So free them up.
3537 bool pinned = true;
3538 switch( n->Opcode() ) {
3539 case Op_DivI:
3540 case Op_DivF:
3541 case Op_DivD:
3542 case Op_ModI:
3543 case Op_ModF:
3544 case Op_ModD:
3545 case Op_LoadB: // Same with Loads; they can sink
3546 case Op_LoadUB: // during loop optimizations.
3547 case Op_LoadUS:
3548 case Op_LoadD:
3549 case Op_LoadF:
3550 case Op_LoadI:
3551 case Op_LoadKlass:
3552 case Op_LoadNKlass:
3553 case Op_LoadL:
3554 case Op_LoadS:
3555 case Op_LoadP:
3556 case Op_LoadN:
3557 case Op_LoadRange:
3558 case Op_LoadD_unaligned:
3559 case Op_LoadL_unaligned:
3560 case Op_StrComp: // Does a bunch of load-like effects
3561 case Op_StrEquals:
3562 case Op_StrIndexOf:
3563 case Op_StrIndexOfChar:
3564 case Op_AryEq:
3565 case Op_HasNegatives:
3566 pinned = false;
3567 }
3568 if( pinned ) {
3569 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3570 if( !chosen_loop->_child ) // Inner loop?
3571 chosen_loop->_body.push(n); // Collect inner loops
3572 return;
3573 }
3574 } else { // No slot zero
3575 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3576 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3577 return;
3578 }
3579 assert(!n->is_CFG() || n->outcnt() == 0, "");
3580 }
3581
3582 // Do I have a "safe range" I can select over?
3583 Node *early = get_ctrl(n);// Early location already computed
3584
3585 // Compute latest point this Node can go
3586 Node *LCA = get_late_ctrl( n, early );
3587 // LCA is NULL due to uses being dead
3588 if( LCA == NULL ) {
3589 #ifdef ASSERT
3590 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3591 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3592 }
3593 #endif
3594 _nodes.map(n->_idx, 0); // This node is useless
3595 _deadlist.push(n);
3596 return;
3597 }
3598 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3599
3600 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3601 Node *least = legal; // Best legal position so far
3602 while( early != legal ) { // While not at earliest legal
3603 #ifdef ASSERT
3604 if (legal->is_Start() && !early->is_Root()) {
3605 // Bad graph. Print idom path and fail.
3606 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
3607 assert(false, "Bad graph detected in build_loop_late");
3608 }
3609 #endif
3610 // Find least loop nesting depth
3611 legal = idom(legal); // Bump up the IDOM tree
3612 // Check for lower nesting depth
3613 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3614 least = legal;
3615 }
3616 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3617
3618 // Try not to place code on a loop entry projection
3619 // which can inhibit range check elimination.
3620 if (least != early) {
3621 Node* ctrl_out = least->unique_ctrl_out();
3622 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3623 least == ctrl_out->in(LoopNode::EntryControl)) {
3624 Node* least_dom = idom(least);
3625 if (get_loop(least_dom)->is_member(get_loop(least))) {
3626 least = least_dom;
3627 }
3628 }
3629 }
3630
3631 #ifdef ASSERT
3632 // If verifying, verify that 'verify_me' has a legal location
3633 // and choose it as our location.
3634 if( _verify_me ) {
3635 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3636 Node *legal = LCA;
3637 while( early != legal ) { // While not at earliest legal
3638 if( legal == v_ctrl ) break; // Check for prior good location
3639 legal = idom(legal) ;// Bump up the IDOM tree
3640 }
3641 // Check for prior good location
3642 if( legal == v_ctrl ) least = legal; // Keep prior if found
3643 }
3644 #endif
3645
3646 // Assign discovered "here or above" point
3647 least = find_non_split_ctrl(least);
3648 set_ctrl(n, least);
3649
3650 // Collect inner loop bodies
3651 IdealLoopTree *chosen_loop = get_loop(least);
3652 if( !chosen_loop->_child ) // Inner loop?
3653 chosen_loop->_body.push(n);// Collect inner loops
3654 }
3655
3656 #ifdef ASSERT
3657 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
3658 tty->print_cr("%s", msg);
3659 tty->print("n: "); n->dump();
3660 tty->print("early(n): "); early->dump();
3661 if (n->in(0) != NULL && !n->in(0)->is_top() &&
3662 n->in(0) != early && !n->in(0)->is_Root()) {
3663 tty->print("n->in(0): "); n->in(0)->dump();
3664 }
3665 for (uint i = 1; i < n->req(); i++) {
3666 Node* in1 = n->in(i);
3667 if (in1 != NULL && in1 != n && !in1->is_top()) {
3668 tty->print("n->in(%d): ", i); in1->dump();
3669 Node* in1_early = get_ctrl(in1);
3670 tty->print("early(n->in(%d)): ", i); in1_early->dump();
3671 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
3672 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3673 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3674 }
3675 for (uint j = 1; j < in1->req(); j++) {
3676 Node* in2 = in1->in(j);
3677 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3678 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3679 Node* in2_early = get_ctrl(in2);
3680 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3681 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
3682 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3683 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3684 }
3685 }
3686 }
3687 }
3688 }
3689 tty->cr();
3690 tty->print("LCA(n): "); LCA->dump();
3691 for (uint i = 0; i < n->outcnt(); i++) {
3692 Node* u1 = n->raw_out(i);
3693 if (u1 == n)
3694 continue;
3695 tty->print("n->out(%d): ", i); u1->dump();
3696 if (u1->is_CFG()) {
3697 for (uint j = 0; j < u1->outcnt(); j++) {
3698 Node* u2 = u1->raw_out(j);
3699 if (u2 != u1 && u2 != n && u2->is_CFG()) {
3700 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3701 }
3702 }
3703 } else {
3704 Node* u1_later = get_ctrl(u1);
3705 tty->print("later(n->out(%d)): ", i); u1_later->dump();
3706 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
3707 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3708 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3709 }
3710 for (uint j = 0; j < u1->outcnt(); j++) {
3711 Node* u2 = u1->raw_out(j);
3712 if (u2 == n || u2 == u1)
3713 continue;
3714 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3715 if (!u2->is_CFG()) {
3716 Node* u2_later = get_ctrl(u2);
3717 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3718 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
3719 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3720 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3721 }
3722 }
3723 }
3724 }
3725 }
3726 tty->cr();
3727 int ct = 0;
3728 Node *dbg_legal = LCA;
3729 while(!dbg_legal->is_Start() && ct < 100) {
3730 tty->print("idom[%d] ",ct); dbg_legal->dump();
3731 ct++;
3732 dbg_legal = idom(dbg_legal);
3733 }
3734 tty->cr();
3735 }
3736 #endif
3737
3738 #ifndef PRODUCT
3739 //------------------------------dump-------------------------------------------
3740 void PhaseIdealLoop::dump( ) const {
3741 ResourceMark rm;
3742 Arena* arena = Thread::current()->resource_area();
3743 Node_Stack stack(arena, C->live_nodes() >> 2);
3744 Node_List rpo_list;
3745 VectorSet visited(arena);
3746 visited.set(C->top()->_idx);
3747 rpo( C->root(), stack, visited, rpo_list );
3748 // Dump root loop indexed by last element in PO order
3749 dump( _ltree_root, rpo_list.size(), rpo_list );
3750 }
3751
3752 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3753 loop->dump_head();
3754
3755 // Now scan for CFG nodes in the same loop
3756 for( uint j=idx; j > 0; j-- ) {
3757 Node *n = rpo_list[j-1];
3758 if( !_nodes[n->_idx] ) // Skip dead nodes
3759 continue;
3760 if( get_loop(n) != loop ) { // Wrong loop nest
3761 if( get_loop(n)->_head == n && // Found nested loop?
3762 get_loop(n)->_parent == loop )
3763 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3764 continue;
3765 }
3766
3767 // Dump controlling node
3768 for( uint x = 0; x < loop->_nest; x++ )
3769 tty->print(" ");
3770 tty->print("C");
3771 if( n == C->root() ) {
3772 n->dump();
3773 } else {
3774 Node* cached_idom = idom_no_update(n);
3775 Node *computed_idom = n->in(0);
3776 if( n->is_Region() ) {
3777 computed_idom = compute_idom(n);
3778 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3779 // any MultiBranch ctrl node), so apply a similar transform to
3780 // the cached idom returned from idom_no_update.
3781 cached_idom = find_non_split_ctrl(cached_idom);
3782 }
3783 tty->print(" ID:%d",computed_idom->_idx);
3784 n->dump();
3785 if( cached_idom != computed_idom ) {
3786 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3787 computed_idom->_idx, cached_idom->_idx);
3788 }
3789 }
3790 // Dump nodes it controls
3791 for( uint k = 0; k < _nodes.Size(); k++ ) {
3792 // (k < C->unique() && get_ctrl(find(k)) == n)
3793 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3794 Node *m = C->root()->find(k);
3795 if( m && m->outcnt() > 0 ) {
3796 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3797 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3798 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3799 }
3800 for( uint j = 0; j < loop->_nest; j++ )
3801 tty->print(" ");
3802 tty->print(" ");
3803 m->dump();
3804 }
3805 }
3806 }
3807 }
3808 }
3809
3810 // Collect a R-P-O for the whole CFG.
3811 // Result list is in post-order (scan backwards for RPO)
3812 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3813 stk.push(start, 0);
3814 visited.set(start->_idx);
3815
3816 while (stk.is_nonempty()) {
3817 Node* m = stk.node();
3818 uint idx = stk.index();
3819 if (idx < m->outcnt()) {
3820 stk.set_index(idx + 1);
3821 Node* n = m->raw_out(idx);
3822 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3823 stk.push(n, 0);
3824 }
3825 } else {
3826 rpo_list.push(m);
3827 stk.pop();
3828 }
3829 }
3830 }
3831 #endif
3832
3833
3834 //=============================================================================
3835 //------------------------------LoopTreeIterator-----------------------------------
3836
3837 // Advance to next loop tree using a preorder, left-to-right traversal.
3838 void LoopTreeIterator::next() {
3839 assert(!done(), "must not be done.");
3840 if (_curnt->_child != NULL) {
3841 _curnt = _curnt->_child;
3842 } else if (_curnt->_next != NULL) {
3843 _curnt = _curnt->_next;
3844 } else {
3845 while (_curnt != _root && _curnt->_next == NULL) {
3846 _curnt = _curnt->_parent;
3847 }
3848 if (_curnt == _root) {
3849 _curnt = NULL;
3850 assert(done(), "must be done.");
3851 } else {
3852 assert(_curnt->_next != NULL, "must be more to do");
3853 _curnt = _curnt->_next;
3854 }
3855 }
3856 }