Fri, 04 Dec 2015 14:06:38 +0100
6869327: Add new C2 flag to keep safepoints in counted loops.
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1998, 2014, 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 */
25 #include "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/superword.hpp"
40 //=============================================================================
41 //------------------------------is_loop_iv-------------------------------------
42 // Determine if a node is Counted loop induction variable.
43 // The method is declared in node.hpp.
44 const Node* Node::is_loop_iv() const {
45 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
46 this->as_Phi()->region()->is_CountedLoop() &&
47 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
48 return this;
49 } else {
50 return NULL;
51 }
52 }
54 //=============================================================================
55 //------------------------------dump_spec--------------------------------------
56 // Dump special per-node info
57 #ifndef PRODUCT
58 void LoopNode::dump_spec(outputStream *st) const {
59 if (is_inner_loop()) st->print( "inner " );
60 if (is_partial_peel_loop()) st->print( "partial_peel " );
61 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
62 }
63 #endif
65 //------------------------------is_valid_counted_loop-------------------------
66 bool LoopNode::is_valid_counted_loop() const {
67 if (is_CountedLoop()) {
68 CountedLoopNode* l = as_CountedLoop();
69 CountedLoopEndNode* le = l->loopexit();
70 if (le != NULL &&
71 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
72 Node* phi = l->phi();
73 Node* exit = le->proj_out(0 /* false */);
74 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
75 phi != NULL && phi->is_Phi() &&
76 phi->in(LoopNode::LoopBackControl) == l->incr() &&
77 le->loopnode() == l && le->stride_is_con()) {
78 return true;
79 }
80 }
81 }
82 return false;
83 }
85 //------------------------------get_early_ctrl---------------------------------
86 // Compute earliest legal control
87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
88 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
89 uint i;
90 Node *early;
91 if (n->in(0) && !n->is_expensive()) {
92 early = n->in(0);
93 if (!early->is_CFG()) // Might be a non-CFG multi-def
94 early = get_ctrl(early); // So treat input as a straight data input
95 i = 1;
96 } else {
97 early = get_ctrl(n->in(1));
98 i = 2;
99 }
100 uint e_d = dom_depth(early);
101 assert( early, "" );
102 for (; i < n->req(); i++) {
103 Node *cin = get_ctrl(n->in(i));
104 assert( cin, "" );
105 // Keep deepest dominator depth
106 uint c_d = dom_depth(cin);
107 if (c_d > e_d) { // Deeper guy?
108 early = cin; // Keep deepest found so far
109 e_d = c_d;
110 } else if (c_d == e_d && // Same depth?
111 early != cin) { // If not equal, must use slower algorithm
112 // If same depth but not equal, one _must_ dominate the other
113 // and we want the deeper (i.e., dominated) guy.
114 Node *n1 = early;
115 Node *n2 = cin;
116 while (1) {
117 n1 = idom(n1); // Walk up until break cycle
118 n2 = idom(n2);
119 if (n1 == cin || // Walked early up to cin
120 dom_depth(n2) < c_d)
121 break; // early is deeper; keep him
122 if (n2 == early || // Walked cin up to early
123 dom_depth(n1) < c_d) {
124 early = cin; // cin is deeper; keep him
125 break;
126 }
127 }
128 e_d = dom_depth(early); // Reset depth register cache
129 }
130 }
132 // Return earliest legal location
133 assert(early == find_non_split_ctrl(early), "unexpected early control");
135 if (n->is_expensive()) {
136 assert(n->in(0), "should have control input");
137 early = get_early_ctrl_for_expensive(n, early);
138 }
140 return early;
141 }
143 //------------------------------get_early_ctrl_for_expensive---------------------------------
144 // Move node up the dominator tree as high as legal while still beneficial
145 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
146 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
147 assert(OptimizeExpensiveOps, "optimization off?");
149 Node* ctl = n->in(0);
150 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
151 uint min_dom_depth = dom_depth(earliest);
152 #ifdef ASSERT
153 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
154 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
155 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
156 }
157 #endif
158 if (dom_depth(ctl) < min_dom_depth) {
159 return earliest;
160 }
162 while (1) {
163 Node *next = ctl;
164 // Moving the node out of a loop on the projection of a If
165 // confuses loop predication. So once we hit a Loop in a If branch
166 // that doesn't branch to an UNC, we stop. The code that process
167 // expensive nodes will notice the loop and skip over it to try to
168 // move the node further up.
169 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
170 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
171 break;
172 }
173 next = idom(ctl->in(1)->in(0));
174 } else if (ctl->is_Proj()) {
175 // We only move it up along a projection if the projection is
176 // the single control projection for its parent: same code path,
177 // if it's a If with UNC or fallthrough of a call.
178 Node* parent_ctl = ctl->in(0);
179 if (parent_ctl == NULL) {
180 break;
181 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
182 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
183 } else if (parent_ctl->is_If()) {
184 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
185 break;
186 }
187 assert(idom(ctl) == parent_ctl, "strange");
188 next = idom(parent_ctl);
189 } else if (ctl->is_CatchProj()) {
190 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
191 break;
192 }
193 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
194 next = parent_ctl->in(0)->in(0)->in(0);
195 } else {
196 // Check if parent control has a single projection (this
197 // control is the only possible successor of the parent
198 // control). If so, we can try to move the node above the
199 // parent control.
200 int nb_ctl_proj = 0;
201 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
202 Node *p = parent_ctl->fast_out(i);
203 if (p->is_Proj() && p->is_CFG()) {
204 nb_ctl_proj++;
205 if (nb_ctl_proj > 1) {
206 break;
207 }
208 }
209 }
211 if (nb_ctl_proj > 1) {
212 break;
213 }
214 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node");
215 assert(idom(ctl) == parent_ctl, "strange");
216 next = idom(parent_ctl);
217 }
218 } else {
219 next = idom(ctl);
220 }
221 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
222 break;
223 }
224 ctl = next;
225 }
227 if (ctl != n->in(0)) {
228 _igvn.hash_delete(n);
229 n->set_req(0, ctl);
230 _igvn.hash_insert(n);
231 }
233 return ctl;
234 }
237 //------------------------------set_early_ctrl---------------------------------
238 // Set earliest legal control
239 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
240 Node *early = get_early_ctrl(n);
242 // Record earliest legal location
243 set_ctrl(n, early);
244 }
246 //------------------------------set_subtree_ctrl-------------------------------
247 // set missing _ctrl entries on new nodes
248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
249 // Already set? Get out.
250 if( _nodes[n->_idx] ) return;
251 // Recursively set _nodes array to indicate where the Node goes
252 uint i;
253 for( i = 0; i < n->req(); ++i ) {
254 Node *m = n->in(i);
255 if( m && m != C->root() )
256 set_subtree_ctrl( m );
257 }
259 // Fixup self
260 set_early_ctrl( n );
261 }
263 //------------------------------is_counted_loop--------------------------------
264 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
265 PhaseGVN *gvn = &_igvn;
267 // Counted loop head must be a good RegionNode with only 3 not NULL
268 // control input edges: Self, Entry, LoopBack.
269 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
270 return false;
271 }
272 Node *init_control = x->in(LoopNode::EntryControl);
273 Node *back_control = x->in(LoopNode::LoopBackControl);
274 if (init_control == NULL || back_control == NULL) // Partially dead
275 return false;
276 // Must also check for TOP when looking for a dead loop
277 if (init_control->is_top() || back_control->is_top())
278 return false;
280 // Allow funny placement of Safepoint
281 if (back_control->Opcode() == Op_SafePoint)
282 back_control = back_control->in(TypeFunc::Control);
284 // Controlling test for loop
285 Node *iftrue = back_control;
286 uint iftrue_op = iftrue->Opcode();
287 if (iftrue_op != Op_IfTrue &&
288 iftrue_op != Op_IfFalse)
289 // I have a weird back-control. Probably the loop-exit test is in
290 // the middle of the loop and I am looking at some trailing control-flow
291 // merge point. To fix this I would have to partially peel the loop.
292 return false; // Obscure back-control
294 // Get boolean guarding loop-back test
295 Node *iff = iftrue->in(0);
296 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
297 return false;
298 BoolNode *test = iff->in(1)->as_Bool();
299 BoolTest::mask bt = test->_test._test;
300 float cl_prob = iff->as_If()->_prob;
301 if (iftrue_op == Op_IfFalse) {
302 bt = BoolTest(bt).negate();
303 cl_prob = 1.0 - cl_prob;
304 }
305 // Get backedge compare
306 Node *cmp = test->in(1);
307 int cmp_op = cmp->Opcode();
308 if (cmp_op != Op_CmpI)
309 return false; // Avoid pointer & float compares
311 // Find the trip-counter increment & limit. Limit must be loop invariant.
312 Node *incr = cmp->in(1);
313 Node *limit = cmp->in(2);
315 // ---------
316 // need 'loop()' test to tell if limit is loop invariant
317 // ---------
319 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
320 Node *tmp = incr; // Then reverse order into the CmpI
321 incr = limit;
322 limit = tmp;
323 bt = BoolTest(bt).commute(); // And commute the exit test
324 }
325 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
326 return false;
327 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
328 return false;
330 Node* phi_incr = NULL;
331 // Trip-counter increment must be commutative & associative.
332 if (incr->is_Phi()) {
333 if (incr->as_Phi()->region() != x || incr->req() != 3)
334 return false; // Not simple trip counter expression
335 phi_incr = incr;
336 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
337 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
338 return false;
339 }
341 Node* trunc1 = NULL;
342 Node* trunc2 = NULL;
343 const TypeInt* iv_trunc_t = NULL;
344 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
345 return false; // Funny increment opcode
346 }
347 assert(incr->Opcode() == Op_AddI, "wrong increment code");
349 // Get merge point
350 Node *xphi = incr->in(1);
351 Node *stride = incr->in(2);
352 if (!stride->is_Con()) { // Oops, swap these
353 if (!xphi->is_Con()) // Is the other guy a constant?
354 return false; // Nope, unknown stride, bail out
355 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
356 xphi = stride;
357 stride = tmp;
358 }
359 // Stride must be constant
360 int stride_con = stride->get_int();
361 if (stride_con == 0)
362 return false; // missed some peephole opt
364 if (!xphi->is_Phi())
365 return false; // Too much math on the trip counter
366 if (phi_incr != NULL && phi_incr != xphi)
367 return false;
368 PhiNode *phi = xphi->as_Phi();
370 // Phi must be of loop header; backedge must wrap to increment
371 if (phi->region() != x)
372 return false;
373 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
374 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
375 return false;
376 }
377 Node *init_trip = phi->in(LoopNode::EntryControl);
379 // If iv trunc type is smaller than int, check for possible wrap.
380 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
381 assert(trunc1 != NULL, "must have found some truncation");
383 // Get a better type for the phi (filtered thru if's)
384 const TypeInt* phi_ft = filtered_type(phi);
386 // Can iv take on a value that will wrap?
387 //
388 // Ensure iv's limit is not within "stride" of the wrap value.
389 //
390 // Example for "short" type
391 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
392 // If the stride is +10, then the last value of the induction
393 // variable before the increment (phi_ft->_hi) must be
394 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
395 // ensure no truncation occurs after the increment.
397 if (stride_con > 0) {
398 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
399 iv_trunc_t->_lo > phi_ft->_lo) {
400 return false; // truncation may occur
401 }
402 } else if (stride_con < 0) {
403 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
404 iv_trunc_t->_hi < phi_ft->_hi) {
405 return false; // truncation may occur
406 }
407 }
408 // No possibility of wrap so truncation can be discarded
409 // Promote iv type to Int
410 } else {
411 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
412 }
414 // If the condition is inverted and we will be rolling
415 // through MININT to MAXINT, then bail out.
416 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
417 // Odd stride
418 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
419 // Count down loop rolls through MAXINT
420 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
421 // Count up loop rolls through MININT
422 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
423 return false; // Bail out
424 }
426 const TypeInt* init_t = gvn->type(init_trip)->is_int();
427 const TypeInt* limit_t = gvn->type(limit)->is_int();
429 if (stride_con > 0) {
430 jlong init_p = (jlong)init_t->_lo + stride_con;
431 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
432 return false; // cyclic loop or this loop trips only once
433 } else {
434 jlong init_p = (jlong)init_t->_hi + stride_con;
435 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
436 return false; // cyclic loop or this loop trips only once
437 }
439 if (phi_incr != NULL) {
440 // check if there is a possiblity of IV overflowing after the first increment
441 if (stride_con > 0) {
442 if (init_t->_hi > max_jint - stride_con) {
443 return false;
444 }
445 } else {
446 if (init_t->_lo < min_jint - stride_con) {
447 return false;
448 }
449 }
450 }
452 // =================================================
453 // ---- SUCCESS! Found A Trip-Counted Loop! -----
454 //
455 assert(x->Opcode() == Op_Loop, "regular loops only");
456 C->print_method(PHASE_BEFORE_CLOOPS, 3);
458 Node *hook = new (C) Node(6);
460 if (LoopLimitCheck) {
462 // ===================================================
463 // Generate loop limit check to avoid integer overflow
464 // in cases like next (cyclic loops):
465 //
466 // for (i=0; i <= max_jint; i++) {}
467 // for (i=0; i < max_jint; i+=2) {}
468 //
469 //
470 // Limit check predicate depends on the loop test:
471 //
472 // for(;i != limit; i++) --> limit <= (max_jint)
473 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
474 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
475 //
477 // Check if limit is excluded to do more precise int overflow check.
478 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
479 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
481 // If compare points directly to the phi we need to adjust
482 // the compare so that it points to the incr. Limit have
483 // to be adjusted to keep trip count the same and the
484 // adjusted limit should be checked for int overflow.
485 if (phi_incr != NULL) {
486 stride_m += stride_con;
487 }
489 if (limit->is_Con()) {
490 int limit_con = limit->get_int();
491 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
492 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
493 // Bailout: it could be integer overflow.
494 return false;
495 }
496 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
497 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
498 // Limit's type may satisfy the condition, for example,
499 // when it is an array length.
500 } else {
501 // Generate loop's limit check.
502 // Loop limit check predicate should be near the loop.
503 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
504 if (!limit_check_proj) {
505 // The limit check predicate is not generated if this method trapped here before.
506 #ifdef ASSERT
507 if (TraceLoopLimitCheck) {
508 tty->print("missing loop limit check:");
509 loop->dump_head();
510 x->dump(1);
511 }
512 #endif
513 return false;
514 }
516 IfNode* check_iff = limit_check_proj->in(0)->as_If();
517 Node* cmp_limit;
518 Node* bol;
520 if (stride_con > 0) {
521 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
522 bol = new (C) BoolNode(cmp_limit, BoolTest::le);
523 } else {
524 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
525 bol = new (C) BoolNode(cmp_limit, BoolTest::ge);
526 }
527 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
528 bol = _igvn.register_new_node_with_optimizer(bol);
529 set_subtree_ctrl(bol);
531 // Replace condition in original predicate but preserve Opaque node
532 // so that previous predicates could be found.
533 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
534 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
535 Node* opq = check_iff->in(1)->in(1);
536 _igvn.hash_delete(opq);
537 opq->set_req(1, bol);
538 // Update ctrl.
539 set_ctrl(opq, check_iff->in(0));
540 set_ctrl(check_iff->in(1), check_iff->in(0));
542 #ifndef PRODUCT
543 // report that the loop predication has been actually performed
544 // for this loop
545 if (TraceLoopLimitCheck) {
546 tty->print_cr("Counted Loop Limit Check generated:");
547 debug_only( bol->dump(2); )
548 }
549 #endif
550 }
552 if (phi_incr != NULL) {
553 // If compare points directly to the phi we need to adjust
554 // the compare so that it points to the incr. Limit have
555 // to be adjusted to keep trip count the same and we
556 // should avoid int overflow.
557 //
558 // i = init; do {} while(i++ < limit);
559 // is converted to
560 // i = init; do {} while(++i < limit+1);
561 //
562 limit = gvn->transform(new (C) AddINode(limit, stride));
563 }
565 // Now we need to canonicalize loop condition.
566 if (bt == BoolTest::ne) {
567 assert(stride_con == 1 || stride_con == -1, "simple increment only");
568 // 'ne' can be replaced with 'lt' only when init < limit.
569 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
570 bt = BoolTest::lt;
571 // 'ne' can be replaced with 'gt' only when init > limit.
572 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
573 bt = BoolTest::gt;
574 }
576 if (incl_limit) {
577 // The limit check guaranties that 'limit <= (max_jint - stride)' so
578 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
579 //
580 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
581 limit = gvn->transform(new (C) AddINode(limit, one));
582 if (bt == BoolTest::le)
583 bt = BoolTest::lt;
584 else if (bt == BoolTest::ge)
585 bt = BoolTest::gt;
586 else
587 ShouldNotReachHere();
588 }
589 set_subtree_ctrl( limit );
591 } else { // LoopLimitCheck
593 // If compare points to incr, we are ok. Otherwise the compare
594 // can directly point to the phi; in this case adjust the compare so that
595 // it points to the incr by adjusting the limit.
596 if (cmp->in(1) == phi || cmp->in(2) == phi)
597 limit = gvn->transform(new (C) AddINode(limit,stride));
599 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
600 // Final value for iterator should be: trip_count * stride + init_trip.
601 Node *one_p = gvn->intcon( 1);
602 Node *one_m = gvn->intcon(-1);
604 Node *trip_count = NULL;
605 switch( bt ) {
606 case BoolTest::eq:
607 ShouldNotReachHere();
608 case BoolTest::ne: // Ahh, the case we desire
609 if (stride_con == 1)
610 trip_count = gvn->transform(new (C) SubINode(limit,init_trip));
611 else if (stride_con == -1)
612 trip_count = gvn->transform(new (C) SubINode(init_trip,limit));
613 else
614 ShouldNotReachHere();
615 set_subtree_ctrl(trip_count);
616 //_loop.map(trip_count->_idx,loop(limit));
617 break;
618 case BoolTest::le: // Maybe convert to '<' case
619 limit = gvn->transform(new (C) AddINode(limit,one_p));
620 set_subtree_ctrl( limit );
621 hook->init_req(4, limit);
623 bt = BoolTest::lt;
624 // Make the new limit be in the same loop nest as the old limit
625 //_loop.map(limit->_idx,limit_loop);
626 // Fall into next case
627 case BoolTest::lt: { // Maybe convert to '!=' case
628 if (stride_con < 0) // Count down loop rolls through MAXINT
629 ShouldNotReachHere();
630 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
631 set_subtree_ctrl( range );
632 hook->init_req(0, range);
634 Node *bias = gvn->transform(new (C) AddINode(range,stride));
635 set_subtree_ctrl( bias );
636 hook->init_req(1, bias);
638 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_m));
639 set_subtree_ctrl( bias1 );
640 hook->init_req(2, bias1);
642 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
643 set_subtree_ctrl( trip_count );
644 hook->init_req(3, trip_count);
645 break;
646 }
648 case BoolTest::ge: // Maybe convert to '>' case
649 limit = gvn->transform(new (C) AddINode(limit,one_m));
650 set_subtree_ctrl( limit );
651 hook->init_req(4 ,limit);
653 bt = BoolTest::gt;
654 // Make the new limit be in the same loop nest as the old limit
655 //_loop.map(limit->_idx,limit_loop);
656 // Fall into next case
657 case BoolTest::gt: { // Maybe convert to '!=' case
658 if (stride_con > 0) // count up loop rolls through MININT
659 ShouldNotReachHere();
660 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
661 set_subtree_ctrl( range );
662 hook->init_req(0, range);
664 Node *bias = gvn->transform(new (C) AddINode(range,stride));
665 set_subtree_ctrl( bias );
666 hook->init_req(1, bias);
668 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_p));
669 set_subtree_ctrl( bias1 );
670 hook->init_req(2, bias1);
672 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
673 set_subtree_ctrl( trip_count );
674 hook->init_req(3, trip_count);
675 break;
676 }
677 } // switch( bt )
679 Node *span = gvn->transform(new (C) MulINode(trip_count,stride));
680 set_subtree_ctrl( span );
681 hook->init_req(5, span);
683 limit = gvn->transform(new (C) AddINode(span,init_trip));
684 set_subtree_ctrl( limit );
686 } // LoopLimitCheck
688 if (!UseCountedLoopSafepoints) {
689 // Check for SafePoint on backedge and remove
690 Node *sfpt = x->in(LoopNode::LoopBackControl);
691 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
692 lazy_replace( sfpt, iftrue );
693 if (loop->_safepts != NULL) {
694 loop->_safepts->yank(sfpt);
695 }
696 loop->_tail = iftrue;
697 }
698 }
700 // Build a canonical trip test.
701 // Clone code, as old values may be in use.
702 incr = incr->clone();
703 incr->set_req(1,phi);
704 incr->set_req(2,stride);
705 incr = _igvn.register_new_node_with_optimizer(incr);
706 set_early_ctrl( incr );
707 _igvn.hash_delete(phi);
708 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
710 // If phi type is more restrictive than Int, raise to
711 // Int to prevent (almost) infinite recursion in igvn
712 // which can only handle integer types for constants or minint..maxint.
713 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
714 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
715 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
716 nphi = _igvn.register_new_node_with_optimizer(nphi);
717 set_ctrl(nphi, get_ctrl(phi));
718 _igvn.replace_node(phi, nphi);
719 phi = nphi->as_Phi();
720 }
721 cmp = cmp->clone();
722 cmp->set_req(1,incr);
723 cmp->set_req(2,limit);
724 cmp = _igvn.register_new_node_with_optimizer(cmp);
725 set_ctrl(cmp, iff->in(0));
727 test = test->clone()->as_Bool();
728 (*(BoolTest*)&test->_test)._test = bt;
729 test->set_req(1,cmp);
730 _igvn.register_new_node_with_optimizer(test);
731 set_ctrl(test, iff->in(0));
733 // Replace the old IfNode with a new LoopEndNode
734 Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
735 IfNode *le = lex->as_If();
736 uint dd = dom_depth(iff);
737 set_idom(le, le->in(0), dd); // Update dominance for loop exit
738 set_loop(le, loop);
740 // Get the loop-exit control
741 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
743 // Need to swap loop-exit and loop-back control?
744 if (iftrue_op == Op_IfFalse) {
745 Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le));
746 Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le));
748 loop->_tail = back_control = ift2;
749 set_loop(ift2, loop);
750 set_loop(iff2, get_loop(iffalse));
752 // Lazy update of 'get_ctrl' mechanism.
753 lazy_replace_proj( iffalse, iff2 );
754 lazy_replace_proj( iftrue, ift2 );
756 // Swap names
757 iffalse = iff2;
758 iftrue = ift2;
759 } else {
760 _igvn.hash_delete(iffalse);
761 _igvn.hash_delete(iftrue);
762 iffalse->set_req_X( 0, le, &_igvn );
763 iftrue ->set_req_X( 0, le, &_igvn );
764 }
766 set_idom(iftrue, le, dd+1);
767 set_idom(iffalse, le, dd+1);
768 assert(iff->outcnt() == 0, "should be dead now");
769 lazy_replace( iff, le ); // fix 'get_ctrl'
771 // Now setup a new CountedLoopNode to replace the existing LoopNode
772 CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control);
773 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
774 // The following assert is approximately true, and defines the intention
775 // of can_be_counted_loop. It fails, however, because phase->type
776 // is not yet initialized for this loop and its parts.
777 //assert(l->can_be_counted_loop(this), "sanity");
778 _igvn.register_new_node_with_optimizer(l);
779 set_loop(l, loop);
780 loop->_head = l;
781 // Fix all data nodes placed at the old loop head.
782 // Uses the lazy-update mechanism of 'get_ctrl'.
783 lazy_replace( x, l );
784 set_idom(l, init_control, dom_depth(x));
786 if (!UseCountedLoopSafepoints) {
787 // Check for immediately preceding SafePoint and remove
788 Node *sfpt2 = le->in(0);
789 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
790 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
791 if (loop->_safepts != NULL) {
792 loop->_safepts->yank(sfpt2);
793 }
794 }
795 }
797 // Free up intermediate goo
798 _igvn.remove_dead_node(hook);
800 #ifdef ASSERT
801 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
802 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
803 #endif
804 #ifndef PRODUCT
805 if (TraceLoopOpts) {
806 tty->print("Counted ");
807 loop->dump_head();
808 }
809 #endif
811 C->print_method(PHASE_AFTER_CLOOPS, 3);
813 return true;
814 }
816 //----------------------exact_limit-------------------------------------------
817 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
818 assert(loop->_head->is_CountedLoop(), "");
819 CountedLoopNode *cl = loop->_head->as_CountedLoop();
820 assert(cl->is_valid_counted_loop(), "");
822 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
823 cl->limit()->Opcode() == Op_LoopLimit) {
824 // Old code has exact limit (it could be incorrect in case of int overflow).
825 // Loop limit is exact with stride == 1. And loop may already have exact limit.
826 return cl->limit();
827 }
828 Node *limit = NULL;
829 #ifdef ASSERT
830 BoolTest::mask bt = cl->loopexit()->test_trip();
831 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
832 #endif
833 if (cl->has_exact_trip_count()) {
834 // Simple case: loop has constant boundaries.
835 // Use jlongs to avoid integer overflow.
836 int stride_con = cl->stride_con();
837 jlong init_con = cl->init_trip()->get_int();
838 jlong limit_con = cl->limit()->get_int();
839 julong trip_cnt = cl->trip_count();
840 jlong final_con = init_con + trip_cnt*stride_con;
841 int final_int = (int)final_con;
842 // The final value should be in integer range since the loop
843 // is counted and the limit was checked for overflow.
844 assert(final_con == (jlong)final_int, "final value should be integer");
845 limit = _igvn.intcon(final_int);
846 } else {
847 // Create new LoopLimit node to get exact limit (final iv value).
848 limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
849 register_new_node(limit, cl->in(LoopNode::EntryControl));
850 }
851 assert(limit != NULL, "sanity");
852 return limit;
853 }
855 //------------------------------Ideal------------------------------------------
856 // Return a node which is more "ideal" than the current node.
857 // Attempt to convert into a counted-loop.
858 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
859 if (!can_be_counted_loop(phase)) {
860 phase->C->set_major_progress();
861 }
862 return RegionNode::Ideal(phase, can_reshape);
863 }
866 //=============================================================================
867 //------------------------------Ideal------------------------------------------
868 // Return a node which is more "ideal" than the current node.
869 // Attempt to convert into a counted-loop.
870 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
871 return RegionNode::Ideal(phase, can_reshape);
872 }
874 //------------------------------dump_spec--------------------------------------
875 // Dump special per-node info
876 #ifndef PRODUCT
877 void CountedLoopNode::dump_spec(outputStream *st) const {
878 LoopNode::dump_spec(st);
879 if (stride_is_con()) {
880 st->print("stride: %d ",stride_con());
881 }
882 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
883 if (is_main_loop()) st->print("main of N%d", _idx);
884 if (is_post_loop()) st->print("post of N%d", _main_idx);
885 }
886 #endif
888 //=============================================================================
889 int CountedLoopEndNode::stride_con() const {
890 return stride()->bottom_type()->is_int()->get_con();
891 }
893 //=============================================================================
894 //------------------------------Value-----------------------------------------
895 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
896 const Type* init_t = phase->type(in(Init));
897 const Type* limit_t = phase->type(in(Limit));
898 const Type* stride_t = phase->type(in(Stride));
899 // Either input is TOP ==> the result is TOP
900 if (init_t == Type::TOP) return Type::TOP;
901 if (limit_t == Type::TOP) return Type::TOP;
902 if (stride_t == Type::TOP) return Type::TOP;
904 int stride_con = stride_t->is_int()->get_con();
905 if (stride_con == 1)
906 return NULL; // Identity
908 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
909 // Use jlongs to avoid integer overflow.
910 jlong init_con = init_t->is_int()->get_con();
911 jlong limit_con = limit_t->is_int()->get_con();
912 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
913 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
914 jlong final_con = init_con + stride_con*trip_count;
915 int final_int = (int)final_con;
916 // The final value should be in integer range since the loop
917 // is counted and the limit was checked for overflow.
918 assert(final_con == (jlong)final_int, "final value should be integer");
919 return TypeInt::make(final_int);
920 }
922 return bottom_type(); // TypeInt::INT
923 }
925 //------------------------------Ideal------------------------------------------
926 // Return a node which is more "ideal" than the current node.
927 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
928 if (phase->type(in(Init)) == Type::TOP ||
929 phase->type(in(Limit)) == Type::TOP ||
930 phase->type(in(Stride)) == Type::TOP)
931 return NULL; // Dead
933 int stride_con = phase->type(in(Stride))->is_int()->get_con();
934 if (stride_con == 1)
935 return NULL; // Identity
937 if (in(Init)->is_Con() && in(Limit)->is_Con())
938 return NULL; // Value
940 // Delay following optimizations until all loop optimizations
941 // done to keep Ideal graph simple.
942 if (!can_reshape || phase->C->major_progress())
943 return NULL;
945 const TypeInt* init_t = phase->type(in(Init) )->is_int();
946 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
947 int stride_p;
948 jlong lim, ini;
949 julong max;
950 if (stride_con > 0) {
951 stride_p = stride_con;
952 lim = limit_t->_hi;
953 ini = init_t->_lo;
954 max = (julong)max_jint;
955 } else {
956 stride_p = -stride_con;
957 lim = init_t->_hi;
958 ini = limit_t->_lo;
959 max = (julong)min_jint;
960 }
961 julong range = lim - ini + stride_p;
962 if (range <= max) {
963 // Convert to integer expression if it is not overflow.
964 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
965 Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init)));
966 Node *bias = phase->transform(new (phase->C) AddINode(range, stride_m));
967 Node *trip = phase->transform(new (phase->C) DivINode(0, bias, in(Stride)));
968 Node *span = phase->transform(new (phase->C) MulINode(trip, in(Stride)));
969 return new (phase->C) AddINode(span, in(Init)); // exact limit
970 }
972 if (is_power_of_2(stride_p) || // divisor is 2^n
973 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
974 // Convert to long expression to avoid integer overflow
975 // and let igvn optimizer convert this division.
976 //
977 Node* init = phase->transform( new (phase->C) ConvI2LNode(in(Init)));
978 Node* limit = phase->transform( new (phase->C) ConvI2LNode(in(Limit)));
979 Node* stride = phase->longcon(stride_con);
980 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
982 Node *range = phase->transform(new (phase->C) SubLNode(limit, init));
983 Node *bias = phase->transform(new (phase->C) AddLNode(range, stride_m));
984 Node *span;
985 if (stride_con > 0 && is_power_of_2(stride_p)) {
986 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
987 // and avoid generating rounding for division. Zero trip guard should
988 // guarantee that init < limit but sometimes the guard is missing and
989 // we can get situation when init > limit. Note, for the empty loop
990 // optimization zero trip guard is generated explicitly which leaves
991 // only RCE predicate where exact limit is used and the predicate
992 // will simply fail forcing recompilation.
993 Node* neg_stride = phase->longcon(-stride_con);
994 span = phase->transform(new (phase->C) AndLNode(bias, neg_stride));
995 } else {
996 Node *trip = phase->transform(new (phase->C) DivLNode(0, bias, stride));
997 span = phase->transform(new (phase->C) MulLNode(trip, stride));
998 }
999 // Convert back to int
1000 Node *span_int = phase->transform(new (phase->C) ConvL2INode(span));
1001 return new (phase->C) AddINode(span_int, in(Init)); // exact limit
1002 }
1004 return NULL; // No progress
1005 }
1007 //------------------------------Identity---------------------------------------
1008 // If stride == 1 return limit node.
1009 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
1010 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1011 if (stride_con == 1 || stride_con == -1)
1012 return in(Limit);
1013 return this;
1014 }
1016 //=============================================================================
1017 //----------------------match_incr_with_optional_truncation--------------------
1018 // Match increment with optional truncation:
1019 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1020 // Return NULL for failure. Success returns the increment node.
1021 Node* CountedLoopNode::match_incr_with_optional_truncation(
1022 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1023 // Quick cutouts:
1024 if (expr == NULL || expr->req() != 3) return NULL;
1026 Node *t1 = NULL;
1027 Node *t2 = NULL;
1028 const TypeInt* trunc_t = TypeInt::INT;
1029 Node* n1 = expr;
1030 int n1op = n1->Opcode();
1032 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1033 if (n1op == Op_AndI &&
1034 n1->in(2)->is_Con() &&
1035 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1036 // %%% This check should match any mask of 2**K-1.
1037 t1 = n1;
1038 n1 = t1->in(1);
1039 n1op = n1->Opcode();
1040 trunc_t = TypeInt::CHAR;
1041 } else if (n1op == Op_RShiftI &&
1042 n1->in(1) != NULL &&
1043 n1->in(1)->Opcode() == Op_LShiftI &&
1044 n1->in(2) == n1->in(1)->in(2) &&
1045 n1->in(2)->is_Con()) {
1046 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1047 // %%% This check should match any shift in [1..31].
1048 if (shift == 16 || shift == 8) {
1049 t1 = n1;
1050 t2 = t1->in(1);
1051 n1 = t2->in(1);
1052 n1op = n1->Opcode();
1053 if (shift == 16) {
1054 trunc_t = TypeInt::SHORT;
1055 } else if (shift == 8) {
1056 trunc_t = TypeInt::BYTE;
1057 }
1058 }
1059 }
1061 // If (maybe after stripping) it is an AddI, we won:
1062 if (n1op == Op_AddI) {
1063 *trunc1 = t1;
1064 *trunc2 = t2;
1065 *trunc_type = trunc_t;
1066 return n1;
1067 }
1069 // failed
1070 return NULL;
1071 }
1074 //------------------------------filtered_type--------------------------------
1075 // Return a type based on condition control flow
1076 // A successful return will be a type that is restricted due
1077 // to a series of dominating if-tests, such as:
1078 // if (i < 10) {
1079 // if (i > 0) {
1080 // here: "i" type is [1..10)
1081 // }
1082 // }
1083 // or a control flow merge
1084 // if (i < 10) {
1085 // do {
1086 // phi( , ) -- at top of loop type is [min_int..10)
1087 // i = ?
1088 // } while ( i < 10)
1089 //
1090 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1091 assert(n && n->bottom_type()->is_int(), "must be int");
1092 const TypeInt* filtered_t = NULL;
1093 if (!n->is_Phi()) {
1094 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1095 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1097 } else {
1098 Node* phi = n->as_Phi();
1099 Node* region = phi->in(0);
1100 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1101 if (region && region != C->top()) {
1102 for (uint i = 1; i < phi->req(); i++) {
1103 Node* val = phi->in(i);
1104 Node* use_c = region->in(i);
1105 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1106 if (val_t != NULL) {
1107 if (filtered_t == NULL) {
1108 filtered_t = val_t;
1109 } else {
1110 filtered_t = filtered_t->meet(val_t)->is_int();
1111 }
1112 }
1113 }
1114 }
1115 }
1116 const TypeInt* n_t = _igvn.type(n)->is_int();
1117 if (filtered_t != NULL) {
1118 n_t = n_t->join(filtered_t)->is_int();
1119 }
1120 return n_t;
1121 }
1124 //------------------------------filtered_type_from_dominators--------------------------------
1125 // Return a possibly more restrictive type for val based on condition control flow of dominators
1126 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1127 if (val->is_Con()) {
1128 return val->bottom_type()->is_int();
1129 }
1130 uint if_limit = 10; // Max number of dominating if's visited
1131 const TypeInt* rtn_t = NULL;
1133 if (use_ctrl && use_ctrl != C->top()) {
1134 Node* val_ctrl = get_ctrl(val);
1135 uint val_dom_depth = dom_depth(val_ctrl);
1136 Node* pred = use_ctrl;
1137 uint if_cnt = 0;
1138 while (if_cnt < if_limit) {
1139 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1140 if_cnt++;
1141 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1142 if (if_t != NULL) {
1143 if (rtn_t == NULL) {
1144 rtn_t = if_t;
1145 } else {
1146 rtn_t = rtn_t->join(if_t)->is_int();
1147 }
1148 }
1149 }
1150 pred = idom(pred);
1151 if (pred == NULL || pred == C->top()) {
1152 break;
1153 }
1154 // Stop if going beyond definition block of val
1155 if (dom_depth(pred) < val_dom_depth) {
1156 break;
1157 }
1158 }
1159 }
1160 return rtn_t;
1161 }
1164 //------------------------------dump_spec--------------------------------------
1165 // Dump special per-node info
1166 #ifndef PRODUCT
1167 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1168 if( in(TestValue)->is_Bool() ) {
1169 BoolTest bt( test_trip()); // Added this for g++.
1171 st->print("[");
1172 bt.dump_on(st);
1173 st->print("]");
1174 }
1175 st->print(" ");
1176 IfNode::dump_spec(st);
1177 }
1178 #endif
1180 //=============================================================================
1181 //------------------------------is_member--------------------------------------
1182 // Is 'l' a member of 'this'?
1183 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1184 while( l->_nest > _nest ) l = l->_parent;
1185 return l == this;
1186 }
1188 //------------------------------set_nest---------------------------------------
1189 // Set loop tree nesting depth. Accumulate _has_call bits.
1190 int IdealLoopTree::set_nest( uint depth ) {
1191 _nest = depth;
1192 int bits = _has_call;
1193 if( _child ) bits |= _child->set_nest(depth+1);
1194 if( bits ) _has_call = 1;
1195 if( _next ) bits |= _next ->set_nest(depth );
1196 return bits;
1197 }
1199 //------------------------------split_fall_in----------------------------------
1200 // Split out multiple fall-in edges from the loop header. Move them to a
1201 // private RegionNode before the loop. This becomes the loop landing pad.
1202 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1203 PhaseIterGVN &igvn = phase->_igvn;
1204 uint i;
1206 // Make a new RegionNode to be the landing pad.
1207 Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 );
1208 phase->set_loop(landing_pad,_parent);
1209 // Gather all the fall-in control paths into the landing pad
1210 uint icnt = fall_in_cnt;
1211 uint oreq = _head->req();
1212 for( i = oreq-1; i>0; i-- )
1213 if( !phase->is_member( this, _head->in(i) ) )
1214 landing_pad->set_req(icnt--,_head->in(i));
1216 // Peel off PhiNode edges as well
1217 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1218 Node *oj = _head->fast_out(j);
1219 if( oj->is_Phi() ) {
1220 PhiNode* old_phi = oj->as_Phi();
1221 assert( old_phi->region() == _head, "" );
1222 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1223 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1224 uint icnt = fall_in_cnt;
1225 for( i = oreq-1; i>0; i-- ) {
1226 if( !phase->is_member( this, _head->in(i) ) ) {
1227 p->init_req(icnt--, old_phi->in(i));
1228 // Go ahead and clean out old edges from old phi
1229 old_phi->del_req(i);
1230 }
1231 }
1232 // Search for CSE's here, because ZKM.jar does a lot of
1233 // loop hackery and we need to be a little incremental
1234 // with the CSE to avoid O(N^2) node blow-up.
1235 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1236 if( p2 ) { // Found CSE
1237 p->destruct(); // Recover useless new node
1238 p = p2; // Use old node
1239 } else {
1240 igvn.register_new_node_with_optimizer(p, old_phi);
1241 }
1242 // Make old Phi refer to new Phi.
1243 old_phi->add_req(p);
1244 // Check for the special case of making the old phi useless and
1245 // disappear it. In JavaGrande I have a case where this useless
1246 // Phi is the loop limit and prevents recognizing a CountedLoop
1247 // which in turn prevents removing an empty loop.
1248 Node *id_old_phi = old_phi->Identity( &igvn );
1249 if( id_old_phi != old_phi ) { // Found a simple identity?
1250 // Note that I cannot call 'replace_node' here, because
1251 // that will yank the edge from old_phi to the Region and
1252 // I'm mid-iteration over the Region's uses.
1253 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1254 Node* use = old_phi->last_out(i);
1255 igvn.rehash_node_delayed(use);
1256 uint uses_found = 0;
1257 for (uint j = 0; j < use->len(); j++) {
1258 if (use->in(j) == old_phi) {
1259 if (j < use->req()) use->set_req (j, id_old_phi);
1260 else use->set_prec(j, id_old_phi);
1261 uses_found++;
1262 }
1263 }
1264 i -= uses_found; // we deleted 1 or more copies of this edge
1265 }
1266 }
1267 igvn._worklist.push(old_phi);
1268 }
1269 }
1270 // Finally clean out the fall-in edges from the RegionNode
1271 for( i = oreq-1; i>0; i-- ) {
1272 if( !phase->is_member( this, _head->in(i) ) ) {
1273 _head->del_req(i);
1274 }
1275 }
1276 // Transform landing pad
1277 igvn.register_new_node_with_optimizer(landing_pad, _head);
1278 // Insert landing pad into the header
1279 _head->add_req(landing_pad);
1280 }
1282 //------------------------------split_outer_loop-------------------------------
1283 // Split out the outermost loop from this shared header.
1284 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1285 PhaseIterGVN &igvn = phase->_igvn;
1287 // Find index of outermost loop; it should also be my tail.
1288 uint outer_idx = 1;
1289 while( _head->in(outer_idx) != _tail ) outer_idx++;
1291 // Make a LoopNode for the outermost loop.
1292 Node *ctl = _head->in(LoopNode::EntryControl);
1293 Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) );
1294 outer = igvn.register_new_node_with_optimizer(outer, _head);
1295 phase->set_created_loop_node();
1297 // Outermost loop falls into '_head' loop
1298 _head->set_req(LoopNode::EntryControl, outer);
1299 _head->del_req(outer_idx);
1300 // Split all the Phis up between '_head' loop and 'outer' loop.
1301 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1302 Node *out = _head->fast_out(j);
1303 if( out->is_Phi() ) {
1304 PhiNode *old_phi = out->as_Phi();
1305 assert( old_phi->region() == _head, "" );
1306 Node *phi = PhiNode::make_blank(outer, old_phi);
1307 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1308 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1309 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1310 // Make old Phi point to new Phi on the fall-in path
1311 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1312 old_phi->del_req(outer_idx);
1313 }
1314 }
1316 // Use the new loop head instead of the old shared one
1317 _head = outer;
1318 phase->set_loop(_head, this);
1319 }
1321 //------------------------------fix_parent-------------------------------------
1322 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1323 loop->_parent = parent;
1324 if( loop->_child ) fix_parent( loop->_child, loop );
1325 if( loop->_next ) fix_parent( loop->_next , parent );
1326 }
1328 //------------------------------estimate_path_freq-----------------------------
1329 static float estimate_path_freq( Node *n ) {
1330 // Try to extract some path frequency info
1331 IfNode *iff;
1332 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1333 uint nop = n->Opcode();
1334 if( nop == Op_SafePoint ) { // Skip any safepoint
1335 n = n->in(0);
1336 continue;
1337 }
1338 if( nop == Op_CatchProj ) { // Get count from a prior call
1339 // Assume call does not always throw exceptions: means the call-site
1340 // count is also the frequency of the fall-through path.
1341 assert( n->is_CatchProj(), "" );
1342 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1343 return 0.0f; // Assume call exception path is rare
1344 Node *call = n->in(0)->in(0)->in(0);
1345 assert( call->is_Call(), "expect a call here" );
1346 const JVMState *jvms = ((CallNode*)call)->jvms();
1347 ciMethodData* methodData = jvms->method()->method_data();
1348 if (!methodData->is_mature()) return 0.0f; // No call-site data
1349 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1350 if ((data == NULL) || !data->is_CounterData()) {
1351 // no call profile available, try call's control input
1352 n = n->in(0);
1353 continue;
1354 }
1355 return data->as_CounterData()->count()/FreqCountInvocations;
1356 }
1357 // See if there's a gating IF test
1358 Node *n_c = n->in(0);
1359 if( !n_c->is_If() ) break; // No estimate available
1360 iff = n_c->as_If();
1361 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1362 // Compute how much count comes on this path
1363 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1364 // Have no count info. Skip dull uncommon-trap like branches.
1365 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1366 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1367 break;
1368 // Skip through never-taken branch; look for a real loop exit.
1369 n = iff->in(0);
1370 }
1371 return 0.0f; // No estimate available
1372 }
1374 //------------------------------merge_many_backedges---------------------------
1375 // Merge all the backedges from the shared header into a private Region.
1376 // Feed that region as the one backedge to this loop.
1377 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1378 uint i;
1380 // Scan for the top 2 hottest backedges
1381 float hotcnt = 0.0f;
1382 float warmcnt = 0.0f;
1383 uint hot_idx = 0;
1384 // Loop starts at 2 because slot 1 is the fall-in path
1385 for( i = 2; i < _head->req(); i++ ) {
1386 float cnt = estimate_path_freq(_head->in(i));
1387 if( cnt > hotcnt ) { // Grab hottest path
1388 warmcnt = hotcnt;
1389 hotcnt = cnt;
1390 hot_idx = i;
1391 } else if( cnt > warmcnt ) { // And 2nd hottest path
1392 warmcnt = cnt;
1393 }
1394 }
1396 // See if the hottest backedge is worthy of being an inner loop
1397 // by being much hotter than the next hottest backedge.
1398 if( hotcnt <= 0.0001 ||
1399 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1401 // Peel out the backedges into a private merge point; peel
1402 // them all except optionally hot_idx.
1403 PhaseIterGVN &igvn = phase->_igvn;
1405 Node *hot_tail = NULL;
1406 // Make a Region for the merge point
1407 Node *r = new (phase->C) RegionNode(1);
1408 for( i = 2; i < _head->req(); i++ ) {
1409 if( i != hot_idx )
1410 r->add_req( _head->in(i) );
1411 else hot_tail = _head->in(i);
1412 }
1413 igvn.register_new_node_with_optimizer(r, _head);
1414 // Plug region into end of loop _head, followed by hot_tail
1415 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1416 _head->set_req(2, r);
1417 if( hot_idx ) _head->add_req(hot_tail);
1419 // Split all the Phis up between '_head' loop and the Region 'r'
1420 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1421 Node *out = _head->fast_out(j);
1422 if( out->is_Phi() ) {
1423 PhiNode* n = out->as_Phi();
1424 igvn.hash_delete(n); // Delete from hash before hacking edges
1425 Node *hot_phi = NULL;
1426 Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type());
1427 // Check all inputs for the ones to peel out
1428 uint j = 1;
1429 for( uint i = 2; i < n->req(); i++ ) {
1430 if( i != hot_idx )
1431 phi->set_req( j++, n->in(i) );
1432 else hot_phi = n->in(i);
1433 }
1434 // Register the phi but do not transform until whole place transforms
1435 igvn.register_new_node_with_optimizer(phi, n);
1436 // Add the merge phi to the old Phi
1437 while( n->req() > 3 ) n->del_req( n->req()-1 );
1438 n->set_req(2, phi);
1439 if( hot_idx ) n->add_req(hot_phi);
1440 }
1441 }
1444 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1445 // of self loop tree. Turn self into a loop headed by _head and with
1446 // tail being the new merge point.
1447 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1448 phase->set_loop(_tail,ilt); // Adjust tail
1449 _tail = r; // Self's tail is new merge point
1450 phase->set_loop(r,this);
1451 ilt->_child = _child; // New guy has my children
1452 _child = ilt; // Self has new guy as only child
1453 ilt->_parent = this; // new guy has self for parent
1454 ilt->_nest = _nest; // Same nesting depth (for now)
1456 // Starting with 'ilt', look for child loop trees using the same shared
1457 // header. Flatten these out; they will no longer be loops in the end.
1458 IdealLoopTree **pilt = &_child;
1459 while( ilt ) {
1460 if( ilt->_head == _head ) {
1461 uint i;
1462 for( i = 2; i < _head->req(); i++ )
1463 if( _head->in(i) == ilt->_tail )
1464 break; // Still a loop
1465 if( i == _head->req() ) { // No longer a loop
1466 // Flatten ilt. Hang ilt's "_next" list from the end of
1467 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1468 IdealLoopTree **cp = &ilt->_child;
1469 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1470 *cp = ilt->_next; // Hang next list at end of child list
1471 *pilt = ilt->_child; // Move child up to replace ilt
1472 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1473 ilt = ilt->_child; // Repeat using new ilt
1474 continue; // do not advance over ilt->_child
1475 }
1476 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1477 phase->set_loop(_head,ilt);
1478 }
1479 pilt = &ilt->_child; // Advance to next
1480 ilt = *pilt;
1481 }
1483 if( _child ) fix_parent( _child, this );
1484 }
1486 //------------------------------beautify_loops---------------------------------
1487 // Split shared headers and insert loop landing pads.
1488 // Insert a LoopNode to replace the RegionNode.
1489 // Return TRUE if loop tree is structurally changed.
1490 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1491 bool result = false;
1492 // Cache parts in locals for easy
1493 PhaseIterGVN &igvn = phase->_igvn;
1495 igvn.hash_delete(_head); // Yank from hash before hacking edges
1497 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1498 int fall_in_cnt = 0;
1499 for( uint i = 1; i < _head->req(); i++ )
1500 if( !phase->is_member( this, _head->in(i) ) )
1501 fall_in_cnt++;
1502 assert( fall_in_cnt, "at least 1 fall-in path" );
1503 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1504 split_fall_in( phase, fall_in_cnt );
1506 // Swap inputs to the _head and all Phis to move the fall-in edge to
1507 // the left.
1508 fall_in_cnt = 1;
1509 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1510 fall_in_cnt++;
1511 if( fall_in_cnt > 1 ) {
1512 // Since I am just swapping inputs I do not need to update def-use info
1513 Node *tmp = _head->in(1);
1514 _head->set_req( 1, _head->in(fall_in_cnt) );
1515 _head->set_req( fall_in_cnt, tmp );
1516 // Swap also all Phis
1517 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1518 Node* phi = _head->fast_out(i);
1519 if( phi->is_Phi() ) {
1520 igvn.hash_delete(phi); // Yank from hash before hacking edges
1521 tmp = phi->in(1);
1522 phi->set_req( 1, phi->in(fall_in_cnt) );
1523 phi->set_req( fall_in_cnt, tmp );
1524 }
1525 }
1526 }
1527 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1528 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1530 // If I am a shared header (multiple backedges), peel off the many
1531 // backedges into a private merge point and use the merge point as
1532 // the one true backedge.
1533 if( _head->req() > 3 ) {
1534 // Merge the many backedges into a single backedge but leave
1535 // the hottest backedge as separate edge for the following peel.
1536 merge_many_backedges( phase );
1537 result = true;
1538 }
1540 // If I have one hot backedge, peel off myself loop.
1541 // I better be the outermost loop.
1542 if (_head->req() > 3 && !_irreducible) {
1543 split_outer_loop( phase );
1544 result = true;
1546 } else if (!_head->is_Loop() && !_irreducible) {
1547 // Make a new LoopNode to replace the old loop head
1548 Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) );
1549 l = igvn.register_new_node_with_optimizer(l, _head);
1550 phase->set_created_loop_node();
1551 // Go ahead and replace _head
1552 phase->_igvn.replace_node( _head, l );
1553 _head = l;
1554 phase->set_loop(_head, this);
1555 }
1557 // Now recursively beautify nested loops
1558 if( _child ) result |= _child->beautify_loops( phase );
1559 if( _next ) result |= _next ->beautify_loops( phase );
1560 return result;
1561 }
1563 //------------------------------allpaths_check_safepts----------------------------
1564 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1565 // encountered. Helper for check_safepts.
1566 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1567 assert(stack.size() == 0, "empty stack");
1568 stack.push(_tail);
1569 visited.Clear();
1570 visited.set(_tail->_idx);
1571 while (stack.size() > 0) {
1572 Node* n = stack.pop();
1573 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1574 // Terminate this path
1575 } else if (n->Opcode() == Op_SafePoint) {
1576 if (_phase->get_loop(n) != this) {
1577 if (_required_safept == NULL) _required_safept = new Node_List();
1578 _required_safept->push(n); // save the one closest to the tail
1579 }
1580 // Terminate this path
1581 } else {
1582 uint start = n->is_Region() ? 1 : 0;
1583 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1584 for (uint i = start; i < end; i++) {
1585 Node* in = n->in(i);
1586 assert(in->is_CFG(), "must be");
1587 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1588 stack.push(in);
1589 }
1590 }
1591 }
1592 }
1593 }
1595 //------------------------------check_safepts----------------------------
1596 // Given dominators, try to find loops with calls that must always be
1597 // executed (call dominates loop tail). These loops do not need non-call
1598 // safepoints (ncsfpt).
1599 //
1600 // A complication is that a safepoint in a inner loop may be needed
1601 // by an outer loop. In the following, the inner loop sees it has a
1602 // call (block 3) on every path from the head (block 2) to the
1603 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1604 // in block 2, _but_ this leaves the outer loop without a safepoint.
1605 //
1606 // entry 0
1607 // |
1608 // v
1609 // outer 1,2 +->1
1610 // | |
1611 // | v
1612 // | 2<---+ ncsfpt in 2
1613 // |_/|\ |
1614 // | v |
1615 // inner 2,3 / 3 | call in 3
1616 // / | |
1617 // v +--+
1618 // exit 4
1619 //
1620 //
1621 // This method creates a list (_required_safept) of ncsfpt nodes that must
1622 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1623 // is first looked for in the lists for the outer loops of the current loop.
1624 //
1625 // The insights into the problem:
1626 // A) counted loops are okay
1627 // B) innermost loops are okay (only an inner loop can delete
1628 // a ncsfpt needed by an outer loop)
1629 // C) a loop is immune from an inner loop deleting a safepoint
1630 // if the loop has a call on the idom-path
1631 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1632 // idom-path that is not in a nested loop
1633 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1634 // loop needs to be prevented from deletion by an inner loop
1635 //
1636 // There are two analyses:
1637 // 1) The first, and cheaper one, scans the loop body from
1638 // tail to head following the idom (immediate dominator)
1639 // chain, looking for the cases (C,D,E) above.
1640 // Since inner loops are scanned before outer loops, there is summary
1641 // information about inner loops. Inner loops can be skipped over
1642 // when the tail of an inner loop is encountered.
1643 //
1644 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1645 // the idom path (which is rare), scans all predecessor control paths
1646 // from the tail to the head, terminating a path when a call or sfpt
1647 // is encountered, to find the ncsfpt's that are closest to the tail.
1648 //
1649 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1650 // Bottom up traversal
1651 IdealLoopTree* ch = _child;
1652 if (_child) _child->check_safepts(visited, stack);
1653 if (_next) _next ->check_safepts(visited, stack);
1655 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1656 bool has_call = false; // call on dom-path
1657 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1658 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1659 // Scan the dom-path nodes from tail to head
1660 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1661 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1662 has_call = true;
1663 _has_sfpt = 1; // Then no need for a safept!
1664 break;
1665 } else if (n->Opcode() == Op_SafePoint) {
1666 if (_phase->get_loop(n) == this) {
1667 has_local_ncsfpt = true;
1668 break;
1669 }
1670 if (nonlocal_ncsfpt == NULL) {
1671 nonlocal_ncsfpt = n; // save the one closest to the tail
1672 }
1673 } else {
1674 IdealLoopTree* nlpt = _phase->get_loop(n);
1675 if (this != nlpt) {
1676 // If at an inner loop tail, see if the inner loop has already
1677 // recorded seeing a call on the dom-path (and stop.) If not,
1678 // jump to the head of the inner loop.
1679 assert(is_member(nlpt), "nested loop");
1680 Node* tail = nlpt->_tail;
1681 if (tail->in(0)->is_If()) tail = tail->in(0);
1682 if (n == tail) {
1683 // If inner loop has call on dom-path, so does outer loop
1684 if (nlpt->_has_sfpt) {
1685 has_call = true;
1686 _has_sfpt = 1;
1687 break;
1688 }
1689 // Skip to head of inner loop
1690 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1691 n = nlpt->_head;
1692 }
1693 }
1694 }
1695 }
1696 // Record safept's that this loop needs preserved when an
1697 // inner loop attempts to delete it's safepoints.
1698 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1699 if (nonlocal_ncsfpt != NULL) {
1700 if (_required_safept == NULL) _required_safept = new Node_List();
1701 _required_safept->push(nonlocal_ncsfpt);
1702 } else {
1703 // Failed to find a suitable safept on the dom-path. Now use
1704 // an all paths walk from tail to head, looking for safepoints to preserve.
1705 allpaths_check_safepts(visited, stack);
1706 }
1707 }
1708 }
1709 }
1711 //---------------------------is_deleteable_safept----------------------------
1712 // Is safept not required by an outer loop?
1713 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1714 assert(sfpt->Opcode() == Op_SafePoint, "");
1715 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1716 while (lp != NULL) {
1717 Node_List* sfpts = lp->_required_safept;
1718 if (sfpts != NULL) {
1719 for (uint i = 0; i < sfpts->size(); i++) {
1720 if (sfpt == sfpts->at(i))
1721 return false;
1722 }
1723 }
1724 lp = lp->_parent;
1725 }
1726 return true;
1727 }
1729 //---------------------------replace_parallel_iv-------------------------------
1730 // Replace parallel induction variable (parallel to trip counter)
1731 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1732 assert(loop->_head->is_CountedLoop(), "");
1733 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1734 if (!cl->is_valid_counted_loop())
1735 return; // skip malformed counted loop
1736 Node *incr = cl->incr();
1737 if (incr == NULL)
1738 return; // Dead loop?
1739 Node *init = cl->init_trip();
1740 Node *phi = cl->phi();
1741 int stride_con = cl->stride_con();
1743 // Visit all children, looking for Phis
1744 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1745 Node *out = cl->out(i);
1746 // Look for other phis (secondary IVs). Skip dead ones
1747 if (!out->is_Phi() || out == phi || !has_node(out))
1748 continue;
1749 PhiNode* phi2 = out->as_Phi();
1750 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1751 // Look for induction variables of the form: X += constant
1752 if (phi2->region() != loop->_head ||
1753 incr2->req() != 3 ||
1754 incr2->in(1) != phi2 ||
1755 incr2 == incr ||
1756 incr2->Opcode() != Op_AddI ||
1757 !incr2->in(2)->is_Con())
1758 continue;
1760 // Check for parallel induction variable (parallel to trip counter)
1761 // via an affine function. In particular, count-down loops with
1762 // count-up array indices are common. We only RCE references off
1763 // the trip-counter, so we need to convert all these to trip-counter
1764 // expressions.
1765 Node *init2 = phi2->in( LoopNode::EntryControl );
1766 int stride_con2 = incr2->in(2)->get_int();
1768 // The general case here gets a little tricky. We want to find the
1769 // GCD of all possible parallel IV's and make a new IV using this
1770 // GCD for the loop. Then all possible IVs are simple multiples of
1771 // the GCD. In practice, this will cover very few extra loops.
1772 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1773 // where +/-1 is the common case, but other integer multiples are
1774 // also easy to handle.
1775 int ratio_con = stride_con2/stride_con;
1777 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1778 #ifndef PRODUCT
1779 if (TraceLoopOpts) {
1780 tty->print("Parallel IV: %d ", phi2->_idx);
1781 loop->dump_head();
1782 }
1783 #endif
1784 // Convert to using the trip counter. The parallel induction
1785 // variable differs from the trip counter by a loop-invariant
1786 // amount, the difference between their respective initial values.
1787 // It is scaled by the 'ratio_con'.
1788 Node* ratio = _igvn.intcon(ratio_con);
1789 set_ctrl(ratio, C->root());
1790 Node* ratio_init = new (C) MulINode(init, ratio);
1791 _igvn.register_new_node_with_optimizer(ratio_init, init);
1792 set_early_ctrl(ratio_init);
1793 Node* diff = new (C) SubINode(init2, ratio_init);
1794 _igvn.register_new_node_with_optimizer(diff, init2);
1795 set_early_ctrl(diff);
1796 Node* ratio_idx = new (C) MulINode(phi, ratio);
1797 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1798 set_ctrl(ratio_idx, cl);
1799 Node* add = new (C) AddINode(ratio_idx, diff);
1800 _igvn.register_new_node_with_optimizer(add);
1801 set_ctrl(add, cl);
1802 _igvn.replace_node( phi2, add );
1803 // Sometimes an induction variable is unused
1804 if (add->outcnt() == 0) {
1805 _igvn.remove_dead_node(add);
1806 }
1807 --i; // deleted this phi; rescan starting with next position
1808 continue;
1809 }
1810 }
1811 }
1813 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
1814 // Look for a safepoint on the idom-path.
1815 Node* keep = NULL;
1816 if (keep_one) {
1817 // Keep one if possible
1818 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
1819 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
1820 keep = i;
1821 break; // Found one
1822 }
1823 }
1824 }
1826 // Delete other safepoints in this loop.
1827 Node_List* sfpts = _safepts;
1828 if (sfpts != NULL) {
1829 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
1830 for (uint i = 0; i < sfpts->size(); i++) {
1831 Node* n = sfpts->at(i);
1832 assert(phase->get_loop(n) == this, "");
1833 if (n != keep && phase->is_deleteable_safept(n)) {
1834 phase->lazy_replace(n, n->in(TypeFunc::Control));
1835 }
1836 }
1837 }
1838 }
1840 //------------------------------counted_loop-----------------------------------
1841 // Convert to counted loops where possible
1842 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1844 // For grins, set the inner-loop flag here
1845 if (!_child) {
1846 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1847 }
1849 if (_head->is_CountedLoop() ||
1850 phase->is_counted_loop(_head, this)) {
1852 if (!UseCountedLoopSafepoints) {
1853 // Indicate we do not need a safepoint here
1854 _has_sfpt = 1;
1855 }
1857 // Remove safepoints
1858 bool keep_one_sfpt = !(_has_call || _has_sfpt);
1859 remove_safepoints(phase, keep_one_sfpt);
1861 // Look for induction variables
1862 phase->replace_parallel_iv(this);
1864 } else if (_parent != NULL && !_irreducible) {
1865 // Not a counted loop. Keep one safepoint.
1866 bool keep_one_sfpt = true;
1867 remove_safepoints(phase, keep_one_sfpt);
1868 }
1870 // Recursively
1871 if (_child) _child->counted_loop( phase );
1872 if (_next) _next ->counted_loop( phase );
1873 }
1875 #ifndef PRODUCT
1876 //------------------------------dump_head--------------------------------------
1877 // Dump 1 liner for loop header info
1878 void IdealLoopTree::dump_head( ) const {
1879 for (uint i=0; i<_nest; i++)
1880 tty->print(" ");
1881 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1882 if (_irreducible) tty->print(" IRREDUCIBLE");
1883 Node* entry = _head->in(LoopNode::EntryControl);
1884 if (LoopLimitCheck) {
1885 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1886 if (predicate != NULL ) {
1887 tty->print(" limit_check");
1888 entry = entry->in(0)->in(0);
1889 }
1890 }
1891 if (UseLoopPredicate) {
1892 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1893 if (entry != NULL) {
1894 tty->print(" predicated");
1895 }
1896 }
1897 if (_head->is_CountedLoop()) {
1898 CountedLoopNode *cl = _head->as_CountedLoop();
1899 tty->print(" counted");
1901 Node* init_n = cl->init_trip();
1902 if (init_n != NULL && init_n->is_Con())
1903 tty->print(" [%d,", cl->init_trip()->get_int());
1904 else
1905 tty->print(" [int,");
1906 Node* limit_n = cl->limit();
1907 if (limit_n != NULL && limit_n->is_Con())
1908 tty->print("%d),", cl->limit()->get_int());
1909 else
1910 tty->print("int),");
1911 int stride_con = cl->stride_con();
1912 if (stride_con > 0) tty->print("+");
1913 tty->print("%d", stride_con);
1915 tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
1917 if (cl->is_pre_loop ()) tty->print(" pre" );
1918 if (cl->is_main_loop()) tty->print(" main");
1919 if (cl->is_post_loop()) tty->print(" post");
1920 }
1921 tty->cr();
1922 }
1924 //------------------------------dump-------------------------------------------
1925 // Dump loops by loop tree
1926 void IdealLoopTree::dump( ) const {
1927 dump_head();
1928 if (_child) _child->dump();
1929 if (_next) _next ->dump();
1930 }
1932 #endif
1934 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1935 if (loop == root) {
1936 if (loop->_child != NULL) {
1937 log->begin_head("loop_tree");
1938 log->end_head();
1939 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1940 log->tail("loop_tree");
1941 assert(loop->_next == NULL, "what?");
1942 }
1943 } else {
1944 Node* head = loop->_head;
1945 log->begin_head("loop");
1946 log->print(" idx='%d' ", head->_idx);
1947 if (loop->_irreducible) log->print("irreducible='1' ");
1948 if (head->is_Loop()) {
1949 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1950 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1951 }
1952 if (head->is_CountedLoop()) {
1953 CountedLoopNode* cl = head->as_CountedLoop();
1954 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1955 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1956 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1957 }
1958 log->end_head();
1959 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1960 log->tail("loop");
1961 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1962 }
1963 }
1965 //---------------------collect_potentially_useful_predicates-----------------------
1966 // Helper function to collect potentially useful predicates to prevent them from
1967 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1968 void PhaseIdealLoop::collect_potentially_useful_predicates(
1969 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1970 if (loop->_child) { // child
1971 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1972 }
1974 // self (only loops that we can apply loop predication may use their predicates)
1975 if (loop->_head->is_Loop() &&
1976 !loop->_irreducible &&
1977 !loop->tail()->is_top()) {
1978 LoopNode* lpn = loop->_head->as_Loop();
1979 Node* entry = lpn->in(LoopNode::EntryControl);
1980 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1981 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1982 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1983 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1984 entry = entry->in(0)->in(0);
1985 }
1986 predicate_proj = find_predicate(entry); // Predicate
1987 if (predicate_proj != NULL ) {
1988 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1989 }
1990 }
1992 if (loop->_next) { // sibling
1993 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1994 }
1995 }
1997 //------------------------eliminate_useless_predicates-----------------------------
1998 // Eliminate all inserted predicates if they could not be used by loop predication.
1999 // Note: it will also eliminates loop limits check predicate since it also uses
2000 // Opaque1 node (see Parse::add_predicate()).
2001 void PhaseIdealLoop::eliminate_useless_predicates() {
2002 if (C->predicate_count() == 0)
2003 return; // no predicate left
2005 Unique_Node_List useful_predicates; // to store useful predicates
2006 if (C->has_loops()) {
2007 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2008 }
2010 for (int i = C->predicate_count(); i > 0; i--) {
2011 Node * n = C->predicate_opaque1_node(i-1);
2012 assert(n->Opcode() == Op_Opaque1, "must be");
2013 if (!useful_predicates.member(n)) { // not in the useful list
2014 _igvn.replace_node(n, n->in(1));
2015 }
2016 }
2017 }
2019 //------------------------process_expensive_nodes-----------------------------
2020 // Expensive nodes have their control input set to prevent the GVN
2021 // from commoning them and as a result forcing the resulting node to
2022 // be in a more frequent path. Use CFG information here, to change the
2023 // control inputs so that some expensive nodes can be commoned while
2024 // not executed more frequently.
2025 bool PhaseIdealLoop::process_expensive_nodes() {
2026 assert(OptimizeExpensiveOps, "optimization off?");
2028 // Sort nodes to bring similar nodes together
2029 C->sort_expensive_nodes();
2031 bool progress = false;
2033 for (int i = 0; i < C->expensive_count(); ) {
2034 Node* n = C->expensive_node(i);
2035 int start = i;
2036 // Find nodes similar to n
2037 i++;
2038 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2039 int end = i;
2040 // And compare them two by two
2041 for (int j = start; j < end; j++) {
2042 Node* n1 = C->expensive_node(j);
2043 if (is_node_unreachable(n1)) {
2044 continue;
2045 }
2046 for (int k = j+1; k < end; k++) {
2047 Node* n2 = C->expensive_node(k);
2048 if (is_node_unreachable(n2)) {
2049 continue;
2050 }
2052 assert(n1 != n2, "should be pair of nodes");
2054 Node* c1 = n1->in(0);
2055 Node* c2 = n2->in(0);
2057 Node* parent_c1 = c1;
2058 Node* parent_c2 = c2;
2060 // The call to get_early_ctrl_for_expensive() moves the
2061 // expensive nodes up but stops at loops that are in a if
2062 // branch. See whether we can exit the loop and move above the
2063 // If.
2064 if (c1->is_Loop()) {
2065 parent_c1 = c1->in(1);
2066 }
2067 if (c2->is_Loop()) {
2068 parent_c2 = c2->in(1);
2069 }
2071 if (parent_c1 == parent_c2) {
2072 _igvn._worklist.push(n1);
2073 _igvn._worklist.push(n2);
2074 continue;
2075 }
2077 // Look for identical expensive node up the dominator chain.
2078 if (is_dominator(c1, c2)) {
2079 c2 = c1;
2080 } else if (is_dominator(c2, c1)) {
2081 c1 = c2;
2082 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2083 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2084 // Both branches have the same expensive node so move it up
2085 // before the if.
2086 c1 = c2 = idom(parent_c1->in(0));
2087 }
2088 // Do the actual moves
2089 if (n1->in(0) != c1) {
2090 _igvn.hash_delete(n1);
2091 n1->set_req(0, c1);
2092 _igvn.hash_insert(n1);
2093 _igvn._worklist.push(n1);
2094 progress = true;
2095 }
2096 if (n2->in(0) != c2) {
2097 _igvn.hash_delete(n2);
2098 n2->set_req(0, c2);
2099 _igvn.hash_insert(n2);
2100 _igvn._worklist.push(n2);
2101 progress = true;
2102 }
2103 }
2104 }
2105 }
2107 return progress;
2108 }
2111 //=============================================================================
2112 //----------------------------build_and_optimize-------------------------------
2113 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2114 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2115 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
2116 ResourceMark rm;
2118 int old_progress = C->major_progress();
2119 uint orig_worklist_size = _igvn._worklist.size();
2121 // Reset major-progress flag for the driver's heuristics
2122 C->clear_major_progress();
2124 #ifndef PRODUCT
2125 // Capture for later assert
2126 uint unique = C->unique();
2127 _loop_invokes++;
2128 _loop_work += unique;
2129 #endif
2131 // True if the method has at least 1 irreducible loop
2132 _has_irreducible_loops = false;
2134 _created_loop_node = false;
2136 Arena *a = Thread::current()->resource_area();
2137 VectorSet visited(a);
2138 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2139 _nodes.map(C->unique(), NULL);
2140 memset(_nodes.adr(), 0, wordSize * C->unique());
2142 // Pre-build the top-level outermost loop tree entry
2143 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2144 // Do not need a safepoint at the top level
2145 _ltree_root->_has_sfpt = 1;
2147 // Initialize Dominators.
2148 // Checked in clone_loop_predicate() during beautify_loops().
2149 _idom_size = 0;
2150 _idom = NULL;
2151 _dom_depth = NULL;
2152 _dom_stk = NULL;
2154 // Empty pre-order array
2155 allocate_preorders();
2157 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2158 // IdealLoopTree entries. Data nodes are NOT walked.
2159 build_loop_tree();
2160 // Check for bailout, and return
2161 if (C->failing()) {
2162 return;
2163 }
2165 // No loops after all
2166 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2168 // There should always be an outer loop containing the Root and Return nodes.
2169 // If not, we have a degenerate empty program. Bail out in this case.
2170 if (!has_node(C->root())) {
2171 if (!_verify_only) {
2172 C->clear_major_progress();
2173 C->record_method_not_compilable("empty program detected during loop optimization");
2174 }
2175 return;
2176 }
2178 // Nothing to do, so get out
2179 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
2180 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2181 if (stop_early && !do_expensive_nodes) {
2182 _igvn.optimize(); // Cleanup NeverBranches
2183 return;
2184 }
2186 // Set loop nesting depth
2187 _ltree_root->set_nest( 0 );
2189 // Split shared headers and insert loop landing pads.
2190 // Do not bother doing this on the Root loop of course.
2191 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2192 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2193 if( _ltree_root->_child->beautify_loops( this ) ) {
2194 // Re-build loop tree!
2195 _ltree_root->_child = NULL;
2196 _nodes.clear();
2197 reallocate_preorders();
2198 build_loop_tree();
2199 // Check for bailout, and return
2200 if (C->failing()) {
2201 return;
2202 }
2203 // Reset loop nesting depth
2204 _ltree_root->set_nest( 0 );
2206 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2207 }
2208 }
2210 // Build Dominators for elision of NULL checks & loop finding.
2211 // Since nodes do not have a slot for immediate dominator, make
2212 // a persistent side array for that info indexed on node->_idx.
2213 _idom_size = C->unique();
2214 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2215 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2216 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2217 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2219 Dominators();
2221 if (!_verify_only) {
2222 // As a side effect, Dominators removed any unreachable CFG paths
2223 // into RegionNodes. It doesn't do this test against Root, so
2224 // we do it here.
2225 for( uint i = 1; i < C->root()->req(); i++ ) {
2226 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2227 _igvn.delete_input_of(C->root(), i);
2228 i--; // Rerun same iteration on compressed edges
2229 }
2230 }
2232 // Given dominators, try to find inner loops with calls that must
2233 // always be executed (call dominates loop tail). These loops do
2234 // not need a separate safepoint.
2235 Node_List cisstack(a);
2236 _ltree_root->check_safepts(visited, cisstack);
2237 }
2239 // Walk the DATA nodes and place into loops. Find earliest control
2240 // node. For CFG nodes, the _nodes array starts out and remains
2241 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2242 // _nodes array holds the earliest legal controlling CFG node.
2244 // Allocate stack with enough space to avoid frequent realloc
2245 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2246 Node_Stack nstack( a, stack_size );
2248 visited.Clear();
2249 Node_List worklist(a);
2250 // Don't need C->root() on worklist since
2251 // it will be processed among C->top() inputs
2252 worklist.push( C->top() );
2253 visited.set( C->top()->_idx ); // Set C->top() as visited now
2254 build_loop_early( visited, worklist, nstack );
2256 // Given early legal placement, try finding counted loops. This placement
2257 // is good enough to discover most loop invariants.
2258 if( !_verify_me && !_verify_only )
2259 _ltree_root->counted_loop( this );
2261 // Find latest loop placement. Find ideal loop placement.
2262 visited.Clear();
2263 init_dom_lca_tags();
2264 // Need C->root() on worklist when processing outs
2265 worklist.push( C->root() );
2266 NOT_PRODUCT( C->verify_graph_edges(); )
2267 worklist.push( C->top() );
2268 build_loop_late( visited, worklist, nstack );
2270 if (_verify_only) {
2271 // restore major progress flag
2272 for (int i = 0; i < old_progress; i++)
2273 C->set_major_progress();
2274 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2275 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2276 return;
2277 }
2279 // clear out the dead code after build_loop_late
2280 while (_deadlist.size()) {
2281 _igvn.remove_globally_dead_node(_deadlist.pop());
2282 }
2284 if (stop_early) {
2285 assert(do_expensive_nodes, "why are we here?");
2286 if (process_expensive_nodes()) {
2287 // If we made some progress when processing expensive nodes then
2288 // the IGVN may modify the graph in a way that will allow us to
2289 // make some more progress: we need to try processing expensive
2290 // nodes again.
2291 C->set_major_progress();
2292 }
2293 _igvn.optimize();
2294 return;
2295 }
2297 // Some parser-inserted loop predicates could never be used by loop
2298 // predication or they were moved away from loop during some optimizations.
2299 // For example, peeling. Eliminate them before next loop optimizations.
2300 if (UseLoopPredicate || LoopLimitCheck) {
2301 eliminate_useless_predicates();
2302 }
2304 #ifndef PRODUCT
2305 C->verify_graph_edges();
2306 if (_verify_me) { // Nested verify pass?
2307 // Check to see if the verify mode is broken
2308 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2309 return;
2310 }
2311 if(VerifyLoopOptimizations) verify();
2312 if(TraceLoopOpts && C->has_loops()) {
2313 _ltree_root->dump();
2314 }
2315 #endif
2317 if (skip_loop_opts) {
2318 // Cleanup any modified bits
2319 _igvn.optimize();
2321 if (C->log() != NULL) {
2322 log_loop_tree(_ltree_root, _ltree_root, C->log());
2323 }
2324 return;
2325 }
2327 if (ReassociateInvariants) {
2328 // Reassociate invariants and prep for split_thru_phi
2329 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2330 IdealLoopTree* lpt = iter.current();
2331 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2333 lpt->reassociate_invariants(this);
2335 // Because RCE opportunities can be masked by split_thru_phi,
2336 // look for RCE candidates and inhibit split_thru_phi
2337 // on just their loop-phi's for this pass of loop opts
2338 if (SplitIfBlocks && do_split_ifs) {
2339 if (lpt->policy_range_check(this)) {
2340 lpt->_rce_candidate = 1; // = true
2341 }
2342 }
2343 }
2344 }
2346 // Check for aggressive application of split-if and other transforms
2347 // that require basic-block info (like cloning through Phi's)
2348 if( SplitIfBlocks && do_split_ifs ) {
2349 visited.Clear();
2350 split_if_with_blocks( visited, nstack );
2351 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2352 }
2354 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2355 C->set_major_progress();
2356 }
2358 // Perform loop predication before iteration splitting
2359 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2360 _ltree_root->_child->loop_predication(this);
2361 }
2363 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2364 if (do_intrinsify_fill()) {
2365 C->set_major_progress();
2366 }
2367 }
2369 // Perform iteration-splitting on inner loops. Split iterations to avoid
2370 // range checks or one-shot null checks.
2372 // If split-if's didn't hack the graph too bad (no CFG changes)
2373 // then do loop opts.
2374 if (C->has_loops() && !C->major_progress()) {
2375 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2376 _ltree_root->_child->iteration_split( this, worklist );
2377 // No verify after peeling! GCM has hoisted code out of the loop.
2378 // After peeling, the hoisted code could sink inside the peeled area.
2379 // The peeling code does not try to recompute the best location for
2380 // all the code before the peeled area, so the verify pass will always
2381 // complain about it.
2382 }
2383 // Do verify graph edges in any case
2384 NOT_PRODUCT( C->verify_graph_edges(); );
2386 if (!do_split_ifs) {
2387 // We saw major progress in Split-If to get here. We forced a
2388 // pass with unrolling and not split-if, however more split-if's
2389 // might make progress. If the unrolling didn't make progress
2390 // then the major-progress flag got cleared and we won't try
2391 // another round of Split-If. In particular the ever-common
2392 // instance-of/check-cast pattern requires at least 2 rounds of
2393 // Split-If to clear out.
2394 C->set_major_progress();
2395 }
2397 // Repeat loop optimizations if new loops were seen
2398 if (created_loop_node()) {
2399 C->set_major_progress();
2400 }
2402 // Keep loop predicates and perform optimizations with them
2403 // until no more loop optimizations could be done.
2404 // After that switch predicates off and do more loop optimizations.
2405 if (!C->major_progress() && (C->predicate_count() > 0)) {
2406 C->cleanup_loop_predicates(_igvn);
2407 #ifndef PRODUCT
2408 if (TraceLoopOpts) {
2409 tty->print_cr("PredicatesOff");
2410 }
2411 #endif
2412 C->set_major_progress();
2413 }
2415 // Convert scalar to superword operations at the end of all loop opts.
2416 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2417 // SuperWord transform
2418 SuperWord sw(this);
2419 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2420 IdealLoopTree* lpt = iter.current();
2421 if (lpt->is_counted()) {
2422 sw.transform_loop(lpt);
2423 }
2424 }
2425 }
2427 // Cleanup any modified bits
2428 _igvn.optimize();
2430 // disable assert until issue with split_flow_path is resolved (6742111)
2431 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2432 // "shouldn't introduce irreducible loops");
2434 if (C->log() != NULL) {
2435 log_loop_tree(_ltree_root, _ltree_root, C->log());
2436 }
2437 }
2439 #ifndef PRODUCT
2440 //------------------------------print_statistics-------------------------------
2441 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2442 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2443 void PhaseIdealLoop::print_statistics() {
2444 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2445 }
2447 //------------------------------verify-----------------------------------------
2448 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2449 static int fail; // debug only, so its multi-thread dont care
2450 void PhaseIdealLoop::verify() const {
2451 int old_progress = C->major_progress();
2452 ResourceMark rm;
2453 PhaseIdealLoop loop_verify( _igvn, this );
2454 VectorSet visited(Thread::current()->resource_area());
2456 fail = 0;
2457 verify_compare( C->root(), &loop_verify, visited );
2458 assert( fail == 0, "verify loops failed" );
2459 // Verify loop structure is the same
2460 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2461 // Reset major-progress. It was cleared by creating a verify version of
2462 // PhaseIdealLoop.
2463 for( int i=0; i<old_progress; i++ )
2464 C->set_major_progress();
2465 }
2467 //------------------------------verify_compare---------------------------------
2468 // Make sure me and the given PhaseIdealLoop agree on key data structures
2469 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2470 if( !n ) return;
2471 if( visited.test_set( n->_idx ) ) return;
2472 if( !_nodes[n->_idx] ) { // Unreachable
2473 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2474 return;
2475 }
2477 uint i;
2478 for( i = 0; i < n->req(); i++ )
2479 verify_compare( n->in(i), loop_verify, visited );
2481 // Check the '_nodes' block/loop structure
2482 i = n->_idx;
2483 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2484 if( _nodes[i] != loop_verify->_nodes[i] &&
2485 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2486 tty->print("Mismatched control setting for: ");
2487 n->dump();
2488 if( fail++ > 10 ) return;
2489 Node *c = get_ctrl_no_update(n);
2490 tty->print("We have it as: ");
2491 if( c->in(0) ) c->dump();
2492 else tty->print_cr("N%d",c->_idx);
2493 tty->print("Verify thinks: ");
2494 if( loop_verify->has_ctrl(n) )
2495 loop_verify->get_ctrl_no_update(n)->dump();
2496 else
2497 loop_verify->get_loop_idx(n)->dump();
2498 tty->cr();
2499 }
2500 } else { // We have a loop
2501 IdealLoopTree *us = get_loop_idx(n);
2502 if( loop_verify->has_ctrl(n) ) {
2503 tty->print("Mismatched loop setting for: ");
2504 n->dump();
2505 if( fail++ > 10 ) return;
2506 tty->print("We have it as: ");
2507 us->dump();
2508 tty->print("Verify thinks: ");
2509 loop_verify->get_ctrl_no_update(n)->dump();
2510 tty->cr();
2511 } else if (!C->major_progress()) {
2512 // Loop selection can be messed up if we did a major progress
2513 // operation, like split-if. Do not verify in that case.
2514 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2515 if( us->_head != them->_head || us->_tail != them->_tail ) {
2516 tty->print("Unequals loops for: ");
2517 n->dump();
2518 if( fail++ > 10 ) return;
2519 tty->print("We have it as: ");
2520 us->dump();
2521 tty->print("Verify thinks: ");
2522 them->dump();
2523 tty->cr();
2524 }
2525 }
2526 }
2528 // Check for immediate dominators being equal
2529 if( i >= _idom_size ) {
2530 if( !n->is_CFG() ) return;
2531 tty->print("CFG Node with no idom: ");
2532 n->dump();
2533 return;
2534 }
2535 if( !n->is_CFG() ) return;
2536 if( n == C->root() ) return; // No IDOM here
2538 assert(n->_idx == i, "sanity");
2539 Node *id = idom_no_update(n);
2540 if( id != loop_verify->idom_no_update(n) ) {
2541 tty->print("Unequals idoms for: ");
2542 n->dump();
2543 if( fail++ > 10 ) return;
2544 tty->print("We have it as: ");
2545 id->dump();
2546 tty->print("Verify thinks: ");
2547 loop_verify->idom_no_update(n)->dump();
2548 tty->cr();
2549 }
2551 }
2553 //------------------------------verify_tree------------------------------------
2554 // Verify that tree structures match. Because the CFG can change, siblings
2555 // within the loop tree can be reordered. We attempt to deal with that by
2556 // reordering the verify's loop tree if possible.
2557 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2558 assert( _parent == parent, "Badly formed loop tree" );
2560 // Siblings not in same order? Attempt to re-order.
2561 if( _head != loop->_head ) {
2562 // Find _next pointer to update
2563 IdealLoopTree **pp = &loop->_parent->_child;
2564 while( *pp != loop )
2565 pp = &((*pp)->_next);
2566 // Find proper sibling to be next
2567 IdealLoopTree **nn = &loop->_next;
2568 while( (*nn) && (*nn)->_head != _head )
2569 nn = &((*nn)->_next);
2571 // Check for no match.
2572 if( !(*nn) ) {
2573 // Annoyingly, irreducible loops can pick different headers
2574 // after a major_progress operation, so the rest of the loop
2575 // tree cannot be matched.
2576 if (_irreducible && Compile::current()->major_progress()) return;
2577 assert( 0, "failed to match loop tree" );
2578 }
2580 // Move (*nn) to (*pp)
2581 IdealLoopTree *hit = *nn;
2582 *nn = hit->_next;
2583 hit->_next = loop;
2584 *pp = loop;
2585 loop = hit;
2586 // Now try again to verify
2587 }
2589 assert( _head == loop->_head , "mismatched loop head" );
2590 Node *tail = _tail; // Inline a non-updating version of
2591 while( !tail->in(0) ) // the 'tail()' call.
2592 tail = tail->in(1);
2593 assert( tail == loop->_tail, "mismatched loop tail" );
2595 // Counted loops that are guarded should be able to find their guards
2596 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2597 CountedLoopNode *cl = _head->as_CountedLoop();
2598 Node *init = cl->init_trip();
2599 Node *ctrl = cl->in(LoopNode::EntryControl);
2600 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2601 Node *iff = ctrl->in(0);
2602 assert( iff->Opcode() == Op_If, "" );
2603 Node *bol = iff->in(1);
2604 assert( bol->Opcode() == Op_Bool, "" );
2605 Node *cmp = bol->in(1);
2606 assert( cmp->Opcode() == Op_CmpI, "" );
2607 Node *add = cmp->in(1);
2608 Node *opaq;
2609 if( add->Opcode() == Op_Opaque1 ) {
2610 opaq = add;
2611 } else {
2612 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2613 assert( add == init, "" );
2614 opaq = cmp->in(2);
2615 }
2616 assert( opaq->Opcode() == Op_Opaque1, "" );
2618 }
2620 if (_child != NULL) _child->verify_tree(loop->_child, this);
2621 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2622 // Innermost loops need to verify loop bodies,
2623 // but only if no 'major_progress'
2624 int fail = 0;
2625 if (!Compile::current()->major_progress() && _child == NULL) {
2626 for( uint i = 0; i < _body.size(); i++ ) {
2627 Node *n = _body.at(i);
2628 if (n->outcnt() == 0) continue; // Ignore dead
2629 uint j;
2630 for( j = 0; j < loop->_body.size(); j++ )
2631 if( loop->_body.at(j) == n )
2632 break;
2633 if( j == loop->_body.size() ) { // Not found in loop body
2634 // Last ditch effort to avoid assertion: Its possible that we
2635 // have some users (so outcnt not zero) but are still dead.
2636 // Try to find from root.
2637 if (Compile::current()->root()->find(n->_idx)) {
2638 fail++;
2639 tty->print("We have that verify does not: ");
2640 n->dump();
2641 }
2642 }
2643 }
2644 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2645 Node *n = loop->_body.at(i2);
2646 if (n->outcnt() == 0) continue; // Ignore dead
2647 uint j;
2648 for( j = 0; j < _body.size(); j++ )
2649 if( _body.at(j) == n )
2650 break;
2651 if( j == _body.size() ) { // Not found in loop body
2652 // Last ditch effort to avoid assertion: Its possible that we
2653 // have some users (so outcnt not zero) but are still dead.
2654 // Try to find from root.
2655 if (Compile::current()->root()->find(n->_idx)) {
2656 fail++;
2657 tty->print("Verify has that we do not: ");
2658 n->dump();
2659 }
2660 }
2661 }
2662 assert( !fail, "loop body mismatch" );
2663 }
2664 }
2666 #endif
2668 //------------------------------set_idom---------------------------------------
2669 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2670 uint idx = d->_idx;
2671 if (idx >= _idom_size) {
2672 uint newsize = _idom_size<<1;
2673 while( idx >= newsize ) {
2674 newsize <<= 1;
2675 }
2676 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2677 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2678 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2679 _idom_size = newsize;
2680 }
2681 _idom[idx] = n;
2682 _dom_depth[idx] = dom_depth;
2683 }
2685 //------------------------------recompute_dom_depth---------------------------------------
2686 // The dominator tree is constructed with only parent pointers.
2687 // This recomputes the depth in the tree by first tagging all
2688 // nodes as "no depth yet" marker. The next pass then runs up
2689 // the dom tree from each node marked "no depth yet", and computes
2690 // the depth on the way back down.
2691 void PhaseIdealLoop::recompute_dom_depth() {
2692 uint no_depth_marker = C->unique();
2693 uint i;
2694 // Initialize depth to "no depth yet"
2695 for (i = 0; i < _idom_size; i++) {
2696 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2697 _dom_depth[i] = no_depth_marker;
2698 }
2699 }
2700 if (_dom_stk == NULL) {
2701 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
2702 if (init_size < 10) init_size = 10;
2703 _dom_stk = new GrowableArray<uint>(init_size);
2704 }
2705 // Compute new depth for each node.
2706 for (i = 0; i < _idom_size; i++) {
2707 uint j = i;
2708 // Run up the dom tree to find a node with a depth
2709 while (_dom_depth[j] == no_depth_marker) {
2710 _dom_stk->push(j);
2711 j = _idom[j]->_idx;
2712 }
2713 // Compute the depth on the way back down this tree branch
2714 uint dd = _dom_depth[j] + 1;
2715 while (_dom_stk->length() > 0) {
2716 uint j = _dom_stk->pop();
2717 _dom_depth[j] = dd;
2718 dd++;
2719 }
2720 }
2721 }
2723 //------------------------------sort-------------------------------------------
2724 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2725 // loop tree, not the root.
2726 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2727 if( !innermost ) return loop; // New innermost loop
2729 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2730 assert( loop_preorder, "not yet post-walked loop" );
2731 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2732 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2734 // Insert at start of list
2735 while( l ) { // Insertion sort based on pre-order
2736 if( l == loop ) return innermost; // Already on list!
2737 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2738 assert( l_preorder, "not yet post-walked l" );
2739 // Check header pre-order number to figure proper nesting
2740 if( loop_preorder > l_preorder )
2741 break; // End of insertion
2742 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2743 // Since I split shared headers, you'd think this could not happen.
2744 // BUT: I must first do the preorder numbering before I can discover I
2745 // have shared headers, so the split headers all get the same preorder
2746 // number as the RegionNode they split from.
2747 if( loop_preorder == l_preorder &&
2748 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2749 break; // Also check for shared headers (same pre#)
2750 pp = &l->_parent; // Chain up list
2751 l = *pp;
2752 }
2753 // Link into list
2754 // Point predecessor to me
2755 *pp = loop;
2756 // Point me to successor
2757 IdealLoopTree *p = loop->_parent;
2758 loop->_parent = l; // Point me to successor
2759 if( p ) sort( p, innermost ); // Insert my parents into list as well
2760 return innermost;
2761 }
2763 //------------------------------build_loop_tree--------------------------------
2764 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2765 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2766 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2767 // tightest enclosing IdealLoopTree for post-walked.
2768 //
2769 // During my forward walk I do a short 1-layer lookahead to see if I can find
2770 // a loop backedge with that doesn't have any work on the backedge. This
2771 // helps me construct nested loops with shared headers better.
2772 //
2773 // Once I've done the forward recursion, I do the post-work. For each child
2774 // I check to see if there is a backedge. Backedges define a loop! I
2775 // insert an IdealLoopTree at the target of the backedge.
2776 //
2777 // During the post-work I also check to see if I have several children
2778 // belonging to different loops. If so, then this Node is a decision point
2779 // where control flow can choose to change loop nests. It is at this
2780 // decision point where I can figure out how loops are nested. At this
2781 // time I can properly order the different loop nests from my children.
2782 // Note that there may not be any backedges at the decision point!
2783 //
2784 // Since the decision point can be far removed from the backedges, I can't
2785 // order my loops at the time I discover them. Thus at the decision point
2786 // I need to inspect loop header pre-order numbers to properly nest my
2787 // loops. This means I need to sort my childrens' loops by pre-order.
2788 // The sort is of size number-of-control-children, which generally limits
2789 // it to size 2 (i.e., I just choose between my 2 target loops).
2790 void PhaseIdealLoop::build_loop_tree() {
2791 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
2792 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
2793 Node *n = C->root();
2794 bltstack.push(n);
2795 int pre_order = 1;
2796 int stack_size;
2798 while ( ( stack_size = bltstack.length() ) != 0 ) {
2799 n = bltstack.top(); // Leave node on stack
2800 if ( !is_visited(n) ) {
2801 // ---- Pre-pass Work ----
2802 // Pre-walked but not post-walked nodes need a pre_order number.
2804 set_preorder_visited( n, pre_order ); // set as visited
2806 // ---- Scan over children ----
2807 // Scan first over control projections that lead to loop headers.
2808 // This helps us find inner-to-outer loops with shared headers better.
2810 // Scan children's children for loop headers.
2811 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2812 Node* m = n->raw_out(i); // Child
2813 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2814 // Scan over children's children to find loop
2815 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2816 Node* l = m->fast_out(j);
2817 if( is_visited(l) && // Been visited?
2818 !is_postvisited(l) && // But not post-visited
2819 get_preorder(l) < pre_order ) { // And smaller pre-order
2820 // Found! Scan the DFS down this path before doing other paths
2821 bltstack.push(m);
2822 break;
2823 }
2824 }
2825 }
2826 }
2827 pre_order++;
2828 }
2829 else if ( !is_postvisited(n) ) {
2830 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2831 // such as com.sun.rsasign.am::a.
2832 // For non-recursive version, first, process current children.
2833 // On next iteration, check if additional children were added.
2834 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2835 Node* u = n->raw_out(k);
2836 if ( u->is_CFG() && !is_visited(u) ) {
2837 bltstack.push(u);
2838 }
2839 }
2840 if ( bltstack.length() == stack_size ) {
2841 // There were no additional children, post visit node now
2842 (void)bltstack.pop(); // Remove node from stack
2843 pre_order = build_loop_tree_impl( n, pre_order );
2844 // Check for bailout
2845 if (C->failing()) {
2846 return;
2847 }
2848 // Check to grow _preorders[] array for the case when
2849 // build_loop_tree_impl() adds new nodes.
2850 check_grow_preorders();
2851 }
2852 }
2853 else {
2854 (void)bltstack.pop(); // Remove post-visited node from stack
2855 }
2856 }
2857 }
2859 //------------------------------build_loop_tree_impl---------------------------
2860 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2861 // ---- Post-pass Work ----
2862 // Pre-walked but not post-walked nodes need a pre_order number.
2864 // Tightest enclosing loop for this Node
2865 IdealLoopTree *innermost = NULL;
2867 // For all children, see if any edge is a backedge. If so, make a loop
2868 // for it. Then find the tightest enclosing loop for the self Node.
2869 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2870 Node* m = n->fast_out(i); // Child
2871 if( n == m ) continue; // Ignore control self-cycles
2872 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2874 IdealLoopTree *l; // Child's loop
2875 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2876 // Found a backedge
2877 assert( get_preorder(m) < pre_order, "should be backedge" );
2878 // Check for the RootNode, which is already a LoopNode and is allowed
2879 // to have multiple "backedges".
2880 if( m == C->root()) { // Found the root?
2881 l = _ltree_root; // Root is the outermost LoopNode
2882 } else { // Else found a nested loop
2883 // Insert a LoopNode to mark this loop.
2884 l = new IdealLoopTree(this, m, n);
2885 } // End of Else found a nested loop
2886 if( !has_loop(m) ) // If 'm' does not already have a loop set
2887 set_loop(m, l); // Set loop header to loop now
2889 } else { // Else not a nested loop
2890 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2891 l = get_loop(m); // Get previously determined loop
2892 // If successor is header of a loop (nest), move up-loop till it
2893 // is a member of some outer enclosing loop. Since there are no
2894 // shared headers (I've split them already) I only need to go up
2895 // at most 1 level.
2896 while( l && l->_head == m ) // Successor heads loop?
2897 l = l->_parent; // Move up 1 for me
2898 // If this loop is not properly parented, then this loop
2899 // has no exit path out, i.e. its an infinite loop.
2900 if( !l ) {
2901 // Make loop "reachable" from root so the CFG is reachable. Basically
2902 // insert a bogus loop exit that is never taken. 'm', the loop head,
2903 // points to 'n', one (of possibly many) fall-in paths. There may be
2904 // many backedges as well.
2906 // Here I set the loop to be the root loop. I could have, after
2907 // inserting a bogus loop exit, restarted the recursion and found my
2908 // new loop exit. This would make the infinite loop a first-class
2909 // loop and it would then get properly optimized. What's the use of
2910 // optimizing an infinite loop?
2911 l = _ltree_root; // Oops, found infinite loop
2913 if (!_verify_only) {
2914 // Insert the NeverBranch between 'm' and it's control user.
2915 NeverBranchNode *iff = new (C) NeverBranchNode( m );
2916 _igvn.register_new_node_with_optimizer(iff);
2917 set_loop(iff, l);
2918 Node *if_t = new (C) CProjNode( iff, 0 );
2919 _igvn.register_new_node_with_optimizer(if_t);
2920 set_loop(if_t, l);
2922 Node* cfg = NULL; // Find the One True Control User of m
2923 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2924 Node* x = m->fast_out(j);
2925 if (x->is_CFG() && x != m && x != iff)
2926 { cfg = x; break; }
2927 }
2928 assert(cfg != NULL, "must find the control user of m");
2929 uint k = 0; // Probably cfg->in(0)
2930 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2931 cfg->set_req( k, if_t ); // Now point to NeverBranch
2933 // Now create the never-taken loop exit
2934 Node *if_f = new (C) CProjNode( iff, 1 );
2935 _igvn.register_new_node_with_optimizer(if_f);
2936 set_loop(if_f, l);
2937 // Find frame ptr for Halt. Relies on the optimizer
2938 // V-N'ing. Easier and quicker than searching through
2939 // the program structure.
2940 Node *frame = new (C) ParmNode( C->start(), TypeFunc::FramePtr );
2941 _igvn.register_new_node_with_optimizer(frame);
2942 // Halt & Catch Fire
2943 Node *halt = new (C) HaltNode( if_f, frame );
2944 _igvn.register_new_node_with_optimizer(halt);
2945 set_loop(halt, l);
2946 C->root()->add_req(halt);
2947 }
2948 set_loop(C->root(), _ltree_root);
2949 }
2950 }
2951 // Weeny check for irreducible. This child was already visited (this
2952 // IS the post-work phase). Is this child's loop header post-visited
2953 // as well? If so, then I found another entry into the loop.
2954 if (!_verify_only) {
2955 while( is_postvisited(l->_head) ) {
2956 // found irreducible
2957 l->_irreducible = 1; // = true
2958 l = l->_parent;
2959 _has_irreducible_loops = true;
2960 // Check for bad CFG here to prevent crash, and bailout of compile
2961 if (l == NULL) {
2962 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2963 return pre_order;
2964 }
2965 }
2966 C->set_has_irreducible_loop(_has_irreducible_loops);
2967 }
2969 // This Node might be a decision point for loops. It is only if
2970 // it's children belong to several different loops. The sort call
2971 // does a trivial amount of work if there is only 1 child or all
2972 // children belong to the same loop. If however, the children
2973 // belong to different loops, the sort call will properly set the
2974 // _parent pointers to show how the loops nest.
2975 //
2976 // In any case, it returns the tightest enclosing loop.
2977 innermost = sort( l, innermost );
2978 }
2980 // Def-use info will have some dead stuff; dead stuff will have no
2981 // loop decided on.
2983 // Am I a loop header? If so fix up my parent's child and next ptrs.
2984 if( innermost && innermost->_head == n ) {
2985 assert( get_loop(n) == innermost, "" );
2986 IdealLoopTree *p = innermost->_parent;
2987 IdealLoopTree *l = innermost;
2988 while( p && l->_head == n ) {
2989 l->_next = p->_child; // Put self on parents 'next child'
2990 p->_child = l; // Make self as first child of parent
2991 l = p; // Now walk up the parent chain
2992 p = l->_parent;
2993 }
2994 } else {
2995 // Note that it is possible for a LoopNode to reach here, if the
2996 // backedge has been made unreachable (hence the LoopNode no longer
2997 // denotes a Loop, and will eventually be removed).
2999 // Record tightest enclosing loop for self. Mark as post-visited.
3000 set_loop(n, innermost);
3001 // Also record has_call flag early on
3002 if( innermost ) {
3003 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3004 // Do not count uncommon calls
3005 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3006 Node *iff = n->in(0)->in(0);
3007 // No any calls for vectorized loops.
3008 if( UseSuperWord || !iff->is_If() ||
3009 (n->in(0)->Opcode() == Op_IfFalse &&
3010 (1.0 - iff->as_If()->_prob) >= 0.01) ||
3011 (iff->as_If()->_prob >= 0.01) )
3012 innermost->_has_call = 1;
3013 }
3014 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3015 // Disable loop optimizations if the loop has a scalar replaceable
3016 // allocation. This disabling may cause a potential performance lost
3017 // if the allocation is not eliminated for some reason.
3018 innermost->_allow_optimizations = false;
3019 innermost->_has_call = 1; // = true
3020 } else if (n->Opcode() == Op_SafePoint) {
3021 // Record all safepoints in this loop.
3022 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3023 innermost->_safepts->push(n);
3024 }
3025 }
3026 }
3028 // Flag as post-visited now
3029 set_postvisited(n);
3030 return pre_order;
3031 }
3034 //------------------------------build_loop_early-------------------------------
3035 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3036 // First pass computes the earliest controlling node possible. This is the
3037 // controlling input with the deepest dominating depth.
3038 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3039 while (worklist.size() != 0) {
3040 // Use local variables nstack_top_n & nstack_top_i to cache values
3041 // on nstack's top.
3042 Node *nstack_top_n = worklist.pop();
3043 uint nstack_top_i = 0;
3044 //while_nstack_nonempty:
3045 while (true) {
3046 // Get parent node and next input's index from stack's top.
3047 Node *n = nstack_top_n;
3048 uint i = nstack_top_i;
3049 uint cnt = n->req(); // Count of inputs
3050 if (i == 0) { // Pre-process the node.
3051 if( has_node(n) && // Have either loop or control already?
3052 !has_ctrl(n) ) { // Have loop picked out already?
3053 // During "merge_many_backedges" we fold up several nested loops
3054 // into a single loop. This makes the members of the original
3055 // loop bodies pointing to dead loops; they need to move up
3056 // to the new UNION'd larger loop. I set the _head field of these
3057 // dead loops to NULL and the _parent field points to the owning
3058 // loop. Shades of UNION-FIND algorithm.
3059 IdealLoopTree *ilt;
3060 while( !(ilt = get_loop(n))->_head ) {
3061 // Normally I would use a set_loop here. But in this one special
3062 // case, it is legal (and expected) to change what loop a Node
3063 // belongs to.
3064 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3065 }
3066 // Remove safepoints ONLY if I've already seen I don't need one.
3067 // (the old code here would yank a 2nd safepoint after seeing a
3068 // first one, even though the 1st did not dominate in the loop body
3069 // and thus could be avoided indefinitely)
3070 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3071 is_deleteable_safept(n)) {
3072 Node *in = n->in(TypeFunc::Control);
3073 lazy_replace(n,in); // Pull safepoint now
3074 if (ilt->_safepts != NULL) {
3075 ilt->_safepts->yank(n);
3076 }
3077 // Carry on with the recursion "as if" we are walking
3078 // only the control input
3079 if( !visited.test_set( in->_idx ) ) {
3080 worklist.push(in); // Visit this guy later, using worklist
3081 }
3082 // Get next node from nstack:
3083 // - skip n's inputs processing by setting i > cnt;
3084 // - we also will not call set_early_ctrl(n) since
3085 // has_node(n) == true (see the condition above).
3086 i = cnt + 1;
3087 }
3088 }
3089 } // if (i == 0)
3091 // Visit all inputs
3092 bool done = true; // Assume all n's inputs will be processed
3093 while (i < cnt) {
3094 Node *in = n->in(i);
3095 ++i;
3096 if (in == NULL) continue;
3097 if (in->pinned() && !in->is_CFG())
3098 set_ctrl(in, in->in(0));
3099 int is_visited = visited.test_set( in->_idx );
3100 if (!has_node(in)) { // No controlling input yet?
3101 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3102 assert( !is_visited, "visit only once" );
3103 nstack.push(n, i); // Save parent node and next input's index.
3104 nstack_top_n = in; // Process current input now.
3105 nstack_top_i = 0;
3106 done = false; // Not all n's inputs processed.
3107 break; // continue while_nstack_nonempty;
3108 } else if (!is_visited) {
3109 // This guy has a location picked out for him, but has not yet
3110 // been visited. Happens to all CFG nodes, for instance.
3111 // Visit him using the worklist instead of recursion, to break
3112 // cycles. Since he has a location already we do not need to
3113 // find his location before proceeding with the current Node.
3114 worklist.push(in); // Visit this guy later, using worklist
3115 }
3116 }
3117 if (done) {
3118 // All of n's inputs have been processed, complete post-processing.
3120 // Compute earliest point this Node can go.
3121 // CFG, Phi, pinned nodes already know their controlling input.
3122 if (!has_node(n)) {
3123 // Record earliest legal location
3124 set_early_ctrl( n );
3125 }
3126 if (nstack.is_empty()) {
3127 // Finished all nodes on stack.
3128 // Process next node on the worklist.
3129 break;
3130 }
3131 // Get saved parent node and next input's index.
3132 nstack_top_n = nstack.node();
3133 nstack_top_i = nstack.index();
3134 nstack.pop();
3135 }
3136 } // while (true)
3137 }
3138 }
3140 //------------------------------dom_lca_internal--------------------------------
3141 // Pair-wise LCA
3142 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3143 if( !n1 ) return n2; // Handle NULL original LCA
3144 assert( n1->is_CFG(), "" );
3145 assert( n2->is_CFG(), "" );
3146 // find LCA of all uses
3147 uint d1 = dom_depth(n1);
3148 uint d2 = dom_depth(n2);
3149 while (n1 != n2) {
3150 if (d1 > d2) {
3151 n1 = idom(n1);
3152 d1 = dom_depth(n1);
3153 } else if (d1 < d2) {
3154 n2 = idom(n2);
3155 d2 = dom_depth(n2);
3156 } else {
3157 // Here d1 == d2. Due to edits of the dominator-tree, sections
3158 // of the tree might have the same depth. These sections have
3159 // to be searched more carefully.
3161 // Scan up all the n1's with equal depth, looking for n2.
3162 Node *t1 = idom(n1);
3163 while (dom_depth(t1) == d1) {
3164 if (t1 == n2) return n2;
3165 t1 = idom(t1);
3166 }
3167 // Scan up all the n2's with equal depth, looking for n1.
3168 Node *t2 = idom(n2);
3169 while (dom_depth(t2) == d2) {
3170 if (t2 == n1) return n1;
3171 t2 = idom(t2);
3172 }
3173 // Move up to a new dominator-depth value as well as up the dom-tree.
3174 n1 = t1;
3175 n2 = t2;
3176 d1 = dom_depth(n1);
3177 d2 = dom_depth(n2);
3178 }
3179 }
3180 return n1;
3181 }
3183 //------------------------------compute_idom-----------------------------------
3184 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3185 // IDOMs are correct.
3186 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3187 assert( region->is_Region(), "" );
3188 Node *LCA = NULL;
3189 for( uint i = 1; i < region->req(); i++ ) {
3190 if( region->in(i) != C->top() )
3191 LCA = dom_lca( LCA, region->in(i) );
3192 }
3193 return LCA;
3194 }
3196 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3197 bool had_error = false;
3198 #ifdef ASSERT
3199 if (early != C->root()) {
3200 // Make sure that there's a dominance path from LCA to early
3201 Node* d = LCA;
3202 while (d != early) {
3203 if (d == C->root()) {
3204 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3205 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3206 had_error = true;
3207 break;
3208 }
3209 d = idom(d);
3210 }
3211 }
3212 #endif
3213 return had_error;
3214 }
3217 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3218 // Compute LCA over list of uses
3219 bool had_error = false;
3220 Node *LCA = NULL;
3221 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3222 Node* c = n->fast_out(i);
3223 if (_nodes[c->_idx] == NULL)
3224 continue; // Skip the occasional dead node
3225 if( c->is_Phi() ) { // For Phis, we must land above on the path
3226 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3227 if( c->in(j) == n ) { // Found matching input?
3228 Node *use = c->in(0)->in(j);
3229 if (_verify_only && use->is_top()) continue;
3230 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3231 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3232 }
3233 }
3234 } else {
3235 // For CFG data-users, use is in the block just prior
3236 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3237 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3238 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3239 }
3240 }
3241 assert(!had_error, "bad dominance");
3242 return LCA;
3243 }
3245 //------------------------------get_late_ctrl----------------------------------
3246 // Compute latest legal control.
3247 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3248 assert(early != NULL, "early control should not be NULL");
3250 Node* LCA = compute_lca_of_uses(n, early);
3251 #ifdef ASSERT
3252 if (LCA == C->root() && LCA != early) {
3253 // def doesn't dominate uses so print some useful debugging output
3254 compute_lca_of_uses(n, early, true);
3255 }
3256 #endif
3258 // if this is a load, check for anti-dependent stores
3259 // We use a conservative algorithm to identify potential interfering
3260 // instructions and for rescheduling the load. The users of the memory
3261 // input of this load are examined. Any use which is not a load and is
3262 // dominated by early is considered a potentially interfering store.
3263 // This can produce false positives.
3264 if (n->is_Load() && LCA != early) {
3265 Node_List worklist;
3267 Node *mem = n->in(MemNode::Memory);
3268 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3269 Node* s = mem->fast_out(i);
3270 worklist.push(s);
3271 }
3272 while(worklist.size() != 0 && LCA != early) {
3273 Node* s = worklist.pop();
3274 if (s->is_Load()) {
3275 continue;
3276 } else if (s->is_MergeMem()) {
3277 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3278 Node* s1 = s->fast_out(i);
3279 worklist.push(s1);
3280 }
3281 } else {
3282 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3283 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3284 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3285 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3286 }
3287 }
3288 }
3289 }
3291 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3292 return LCA;
3293 }
3295 // true if CFG node d dominates CFG node n
3296 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3297 if (d == n)
3298 return true;
3299 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3300 uint dd = dom_depth(d);
3301 while (dom_depth(n) >= dd) {
3302 if (n == d)
3303 return true;
3304 n = idom(n);
3305 }
3306 return false;
3307 }
3309 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3310 // Pair-wise LCA with tags.
3311 // Tag each index with the node 'tag' currently being processed
3312 // before advancing up the dominator chain using idom().
3313 // Later calls that find a match to 'tag' know that this path has already
3314 // been considered in the current LCA (which is input 'n1' by convention).
3315 // Since get_late_ctrl() is only called once for each node, the tag array
3316 // does not need to be cleared between calls to get_late_ctrl().
3317 // Algorithm trades a larger constant factor for better asymptotic behavior
3318 //
3319 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3320 uint d1 = dom_depth(n1);
3321 uint d2 = dom_depth(n2);
3323 do {
3324 if (d1 > d2) {
3325 // current lca is deeper than n2
3326 _dom_lca_tags.map(n1->_idx, tag);
3327 n1 = idom(n1);
3328 d1 = dom_depth(n1);
3329 } else if (d1 < d2) {
3330 // n2 is deeper than current lca
3331 Node *memo = _dom_lca_tags[n2->_idx];
3332 if( memo == tag ) {
3333 return n1; // Return the current LCA
3334 }
3335 _dom_lca_tags.map(n2->_idx, tag);
3336 n2 = idom(n2);
3337 d2 = dom_depth(n2);
3338 } else {
3339 // Here d1 == d2. Due to edits of the dominator-tree, sections
3340 // of the tree might have the same depth. These sections have
3341 // to be searched more carefully.
3343 // Scan up all the n1's with equal depth, looking for n2.
3344 _dom_lca_tags.map(n1->_idx, tag);
3345 Node *t1 = idom(n1);
3346 while (dom_depth(t1) == d1) {
3347 if (t1 == n2) return n2;
3348 _dom_lca_tags.map(t1->_idx, tag);
3349 t1 = idom(t1);
3350 }
3351 // Scan up all the n2's with equal depth, looking for n1.
3352 _dom_lca_tags.map(n2->_idx, tag);
3353 Node *t2 = idom(n2);
3354 while (dom_depth(t2) == d2) {
3355 if (t2 == n1) return n1;
3356 _dom_lca_tags.map(t2->_idx, tag);
3357 t2 = idom(t2);
3358 }
3359 // Move up to a new dominator-depth value as well as up the dom-tree.
3360 n1 = t1;
3361 n2 = t2;
3362 d1 = dom_depth(n1);
3363 d2 = dom_depth(n2);
3364 }
3365 } while (n1 != n2);
3366 return n1;
3367 }
3369 //------------------------------init_dom_lca_tags------------------------------
3370 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3371 // Intended use does not involve any growth for the array, so it could
3372 // be of fixed size.
3373 void PhaseIdealLoop::init_dom_lca_tags() {
3374 uint limit = C->unique() + 1;
3375 _dom_lca_tags.map( limit, NULL );
3376 #ifdef ASSERT
3377 for( uint i = 0; i < limit; ++i ) {
3378 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3379 }
3380 #endif // ASSERT
3381 }
3383 //------------------------------clear_dom_lca_tags------------------------------
3384 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3385 // Intended use does not involve any growth for the array, so it could
3386 // be of fixed size.
3387 void PhaseIdealLoop::clear_dom_lca_tags() {
3388 uint limit = C->unique() + 1;
3389 _dom_lca_tags.map( limit, NULL );
3390 _dom_lca_tags.clear();
3391 #ifdef ASSERT
3392 for( uint i = 0; i < limit; ++i ) {
3393 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3394 }
3395 #endif // ASSERT
3396 }
3398 //------------------------------build_loop_late--------------------------------
3399 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3400 // Second pass finds latest legal placement, and ideal loop placement.
3401 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3402 while (worklist.size() != 0) {
3403 Node *n = worklist.pop();
3404 // Only visit once
3405 if (visited.test_set(n->_idx)) continue;
3406 uint cnt = n->outcnt();
3407 uint i = 0;
3408 while (true) {
3409 assert( _nodes[n->_idx], "no dead nodes" );
3410 // Visit all children
3411 if (i < cnt) {
3412 Node* use = n->raw_out(i);
3413 ++i;
3414 // Check for dead uses. Aggressively prune such junk. It might be
3415 // dead in the global sense, but still have local uses so I cannot
3416 // easily call 'remove_dead_node'.
3417 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3418 // Due to cycles, we might not hit the same fixed point in the verify
3419 // pass as we do in the regular pass. Instead, visit such phis as
3420 // simple uses of the loop head.
3421 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3422 if( !visited.test(use->_idx) )
3423 worklist.push(use);
3424 } else if( !visited.test_set(use->_idx) ) {
3425 nstack.push(n, i); // Save parent and next use's index.
3426 n = use; // Process all children of current use.
3427 cnt = use->outcnt();
3428 i = 0;
3429 }
3430 } else {
3431 // Do not visit around the backedge of loops via data edges.
3432 // push dead code onto a worklist
3433 _deadlist.push(use);
3434 }
3435 } else {
3436 // All of n's children have been processed, complete post-processing.
3437 build_loop_late_post(n);
3438 if (nstack.is_empty()) {
3439 // Finished all nodes on stack.
3440 // Process next node on the worklist.
3441 break;
3442 }
3443 // Get saved parent node and next use's index. Visit the rest of uses.
3444 n = nstack.node();
3445 cnt = n->outcnt();
3446 i = nstack.index();
3447 nstack.pop();
3448 }
3449 }
3450 }
3451 }
3453 //------------------------------build_loop_late_post---------------------------
3454 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3455 // Second pass finds latest legal placement, and ideal loop placement.
3456 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3458 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3459 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3460 }
3462 #ifdef ASSERT
3463 if (_verify_only && !n->is_CFG()) {
3464 // Check def-use domination.
3465 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
3466 }
3467 #endif
3469 // CFG and pinned nodes already handled
3470 if( n->in(0) ) {
3471 if( n->in(0)->is_top() ) return; // Dead?
3473 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3474 // _must_ be pinned (they have to observe their control edge of course).
3475 // Unlike Stores (which modify an unallocable resource, the memory
3476 // state), Mods/Loads can float around. So free them up.
3477 bool pinned = true;
3478 switch( n->Opcode() ) {
3479 case Op_DivI:
3480 case Op_DivF:
3481 case Op_DivD:
3482 case Op_ModI:
3483 case Op_ModF:
3484 case Op_ModD:
3485 case Op_LoadB: // Same with Loads; they can sink
3486 case Op_LoadUB: // during loop optimizations.
3487 case Op_LoadUS:
3488 case Op_LoadD:
3489 case Op_LoadF:
3490 case Op_LoadI:
3491 case Op_LoadKlass:
3492 case Op_LoadNKlass:
3493 case Op_LoadL:
3494 case Op_LoadS:
3495 case Op_LoadP:
3496 case Op_LoadN:
3497 case Op_LoadRange:
3498 case Op_LoadD_unaligned:
3499 case Op_LoadL_unaligned:
3500 case Op_StrComp: // Does a bunch of load-like effects
3501 case Op_StrEquals:
3502 case Op_StrIndexOf:
3503 case Op_AryEq:
3504 pinned = false;
3505 }
3506 if( pinned ) {
3507 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3508 if( !chosen_loop->_child ) // Inner loop?
3509 chosen_loop->_body.push(n); // Collect inner loops
3510 return;
3511 }
3512 } else { // No slot zero
3513 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3514 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3515 return;
3516 }
3517 assert(!n->is_CFG() || n->outcnt() == 0, "");
3518 }
3520 // Do I have a "safe range" I can select over?
3521 Node *early = get_ctrl(n);// Early location already computed
3523 // Compute latest point this Node can go
3524 Node *LCA = get_late_ctrl( n, early );
3525 // LCA is NULL due to uses being dead
3526 if( LCA == NULL ) {
3527 #ifdef ASSERT
3528 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3529 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3530 }
3531 #endif
3532 _nodes.map(n->_idx, 0); // This node is useless
3533 _deadlist.push(n);
3534 return;
3535 }
3536 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3538 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3539 Node *least = legal; // Best legal position so far
3540 while( early != legal ) { // While not at earliest legal
3541 #ifdef ASSERT
3542 if (legal->is_Start() && !early->is_Root()) {
3543 // Bad graph. Print idom path and fail.
3544 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
3545 assert(false, "Bad graph detected in build_loop_late");
3546 }
3547 #endif
3548 // Find least loop nesting depth
3549 legal = idom(legal); // Bump up the IDOM tree
3550 // Check for lower nesting depth
3551 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3552 least = legal;
3553 }
3554 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3556 // Try not to place code on a loop entry projection
3557 // which can inhibit range check elimination.
3558 if (least != early) {
3559 Node* ctrl_out = least->unique_ctrl_out();
3560 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3561 least == ctrl_out->in(LoopNode::EntryControl)) {
3562 Node* least_dom = idom(least);
3563 if (get_loop(least_dom)->is_member(get_loop(least))) {
3564 least = least_dom;
3565 }
3566 }
3567 }
3569 #ifdef ASSERT
3570 // If verifying, verify that 'verify_me' has a legal location
3571 // and choose it as our location.
3572 if( _verify_me ) {
3573 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3574 Node *legal = LCA;
3575 while( early != legal ) { // While not at earliest legal
3576 if( legal == v_ctrl ) break; // Check for prior good location
3577 legal = idom(legal) ;// Bump up the IDOM tree
3578 }
3579 // Check for prior good location
3580 if( legal == v_ctrl ) least = legal; // Keep prior if found
3581 }
3582 #endif
3584 // Assign discovered "here or above" point
3585 least = find_non_split_ctrl(least);
3586 set_ctrl(n, least);
3588 // Collect inner loop bodies
3589 IdealLoopTree *chosen_loop = get_loop(least);
3590 if( !chosen_loop->_child ) // Inner loop?
3591 chosen_loop->_body.push(n);// Collect inner loops
3592 }
3594 #ifdef ASSERT
3595 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
3596 tty->print_cr("%s", msg);
3597 tty->print("n: "); n->dump();
3598 tty->print("early(n): "); early->dump();
3599 if (n->in(0) != NULL && !n->in(0)->is_top() &&
3600 n->in(0) != early && !n->in(0)->is_Root()) {
3601 tty->print("n->in(0): "); n->in(0)->dump();
3602 }
3603 for (uint i = 1; i < n->req(); i++) {
3604 Node* in1 = n->in(i);
3605 if (in1 != NULL && in1 != n && !in1->is_top()) {
3606 tty->print("n->in(%d): ", i); in1->dump();
3607 Node* in1_early = get_ctrl(in1);
3608 tty->print("early(n->in(%d)): ", i); in1_early->dump();
3609 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
3610 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3611 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3612 }
3613 for (uint j = 1; j < in1->req(); j++) {
3614 Node* in2 = in1->in(j);
3615 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3616 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3617 Node* in2_early = get_ctrl(in2);
3618 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3619 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
3620 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3621 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3622 }
3623 }
3624 }
3625 }
3626 }
3627 tty->cr();
3628 tty->print("LCA(n): "); LCA->dump();
3629 for (uint i = 0; i < n->outcnt(); i++) {
3630 Node* u1 = n->raw_out(i);
3631 if (u1 == n)
3632 continue;
3633 tty->print("n->out(%d): ", i); u1->dump();
3634 if (u1->is_CFG()) {
3635 for (uint j = 0; j < u1->outcnt(); j++) {
3636 Node* u2 = u1->raw_out(j);
3637 if (u2 != u1 && u2 != n && u2->is_CFG()) {
3638 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3639 }
3640 }
3641 } else {
3642 Node* u1_later = get_ctrl(u1);
3643 tty->print("later(n->out(%d)): ", i); u1_later->dump();
3644 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
3645 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3646 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3647 }
3648 for (uint j = 0; j < u1->outcnt(); j++) {
3649 Node* u2 = u1->raw_out(j);
3650 if (u2 == n || u2 == u1)
3651 continue;
3652 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3653 if (!u2->is_CFG()) {
3654 Node* u2_later = get_ctrl(u2);
3655 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3656 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
3657 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3658 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3659 }
3660 }
3661 }
3662 }
3663 }
3664 tty->cr();
3665 int ct = 0;
3666 Node *dbg_legal = LCA;
3667 while(!dbg_legal->is_Start() && ct < 100) {
3668 tty->print("idom[%d] ",ct); dbg_legal->dump();
3669 ct++;
3670 dbg_legal = idom(dbg_legal);
3671 }
3672 tty->cr();
3673 }
3674 #endif
3676 #ifndef PRODUCT
3677 //------------------------------dump-------------------------------------------
3678 void PhaseIdealLoop::dump( ) const {
3679 ResourceMark rm;
3680 Arena* arena = Thread::current()->resource_area();
3681 Node_Stack stack(arena, C->live_nodes() >> 2);
3682 Node_List rpo_list;
3683 VectorSet visited(arena);
3684 visited.set(C->top()->_idx);
3685 rpo( C->root(), stack, visited, rpo_list );
3686 // Dump root loop indexed by last element in PO order
3687 dump( _ltree_root, rpo_list.size(), rpo_list );
3688 }
3690 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3691 loop->dump_head();
3693 // Now scan for CFG nodes in the same loop
3694 for( uint j=idx; j > 0; j-- ) {
3695 Node *n = rpo_list[j-1];
3696 if( !_nodes[n->_idx] ) // Skip dead nodes
3697 continue;
3698 if( get_loop(n) != loop ) { // Wrong loop nest
3699 if( get_loop(n)->_head == n && // Found nested loop?
3700 get_loop(n)->_parent == loop )
3701 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3702 continue;
3703 }
3705 // Dump controlling node
3706 for( uint x = 0; x < loop->_nest; x++ )
3707 tty->print(" ");
3708 tty->print("C");
3709 if( n == C->root() ) {
3710 n->dump();
3711 } else {
3712 Node* cached_idom = idom_no_update(n);
3713 Node *computed_idom = n->in(0);
3714 if( n->is_Region() ) {
3715 computed_idom = compute_idom(n);
3716 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3717 // any MultiBranch ctrl node), so apply a similar transform to
3718 // the cached idom returned from idom_no_update.
3719 cached_idom = find_non_split_ctrl(cached_idom);
3720 }
3721 tty->print(" ID:%d",computed_idom->_idx);
3722 n->dump();
3723 if( cached_idom != computed_idom ) {
3724 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3725 computed_idom->_idx, cached_idom->_idx);
3726 }
3727 }
3728 // Dump nodes it controls
3729 for( uint k = 0; k < _nodes.Size(); k++ ) {
3730 // (k < C->unique() && get_ctrl(find(k)) == n)
3731 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3732 Node *m = C->root()->find(k);
3733 if( m && m->outcnt() > 0 ) {
3734 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3735 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3736 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3737 }
3738 for( uint j = 0; j < loop->_nest; j++ )
3739 tty->print(" ");
3740 tty->print(" ");
3741 m->dump();
3742 }
3743 }
3744 }
3745 }
3746 }
3748 // Collect a R-P-O for the whole CFG.
3749 // Result list is in post-order (scan backwards for RPO)
3750 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3751 stk.push(start, 0);
3752 visited.set(start->_idx);
3754 while (stk.is_nonempty()) {
3755 Node* m = stk.node();
3756 uint idx = stk.index();
3757 if (idx < m->outcnt()) {
3758 stk.set_index(idx + 1);
3759 Node* n = m->raw_out(idx);
3760 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3761 stk.push(n, 0);
3762 }
3763 } else {
3764 rpo_list.push(m);
3765 stk.pop();
3766 }
3767 }
3768 }
3769 #endif
3772 //=============================================================================
3773 //------------------------------LoopTreeIterator-----------------------------------
3775 // Advance to next loop tree using a preorder, left-to-right traversal.
3776 void LoopTreeIterator::next() {
3777 assert(!done(), "must not be done.");
3778 if (_curnt->_child != NULL) {
3779 _curnt = _curnt->_child;
3780 } else if (_curnt->_next != NULL) {
3781 _curnt = _curnt->_next;
3782 } else {
3783 while (_curnt != _root && _curnt->_next == NULL) {
3784 _curnt = _curnt->_parent;
3785 }
3786 if (_curnt == _root) {
3787 _curnt = NULL;
3788 assert(done(), "must be done.");
3789 } else {
3790 assert(_curnt->_next != NULL, "must be more to do");
3791 _curnt = _curnt->_next;
3792 }
3793 }
3794 }