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