Fri, 15 Jan 2016 22:33:15 +0000
8132051: Better byte behavior
Reviewed-by: coleenp, roland
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 // Check for SafePoint on backedge and remove
689 Node *sfpt = x->in(LoopNode::LoopBackControl);
690 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
691 lazy_replace( sfpt, iftrue );
692 if (loop->_safepts != NULL) {
693 loop->_safepts->yank(sfpt);
694 }
695 loop->_tail = iftrue;
696 }
698 // Build a canonical trip test.
699 // Clone code, as old values may be in use.
700 incr = incr->clone();
701 incr->set_req(1,phi);
702 incr->set_req(2,stride);
703 incr = _igvn.register_new_node_with_optimizer(incr);
704 set_early_ctrl( incr );
705 _igvn.hash_delete(phi);
706 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
708 // If phi type is more restrictive than Int, raise to
709 // Int to prevent (almost) infinite recursion in igvn
710 // which can only handle integer types for constants or minint..maxint.
711 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
712 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
713 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
714 nphi = _igvn.register_new_node_with_optimizer(nphi);
715 set_ctrl(nphi, get_ctrl(phi));
716 _igvn.replace_node(phi, nphi);
717 phi = nphi->as_Phi();
718 }
719 cmp = cmp->clone();
720 cmp->set_req(1,incr);
721 cmp->set_req(2,limit);
722 cmp = _igvn.register_new_node_with_optimizer(cmp);
723 set_ctrl(cmp, iff->in(0));
725 test = test->clone()->as_Bool();
726 (*(BoolTest*)&test->_test)._test = bt;
727 test->set_req(1,cmp);
728 _igvn.register_new_node_with_optimizer(test);
729 set_ctrl(test, iff->in(0));
731 // Replace the old IfNode with a new LoopEndNode
732 Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
733 IfNode *le = lex->as_If();
734 uint dd = dom_depth(iff);
735 set_idom(le, le->in(0), dd); // Update dominance for loop exit
736 set_loop(le, loop);
738 // Get the loop-exit control
739 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
741 // Need to swap loop-exit and loop-back control?
742 if (iftrue_op == Op_IfFalse) {
743 Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le));
744 Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le));
746 loop->_tail = back_control = ift2;
747 set_loop(ift2, loop);
748 set_loop(iff2, get_loop(iffalse));
750 // Lazy update of 'get_ctrl' mechanism.
751 lazy_replace_proj( iffalse, iff2 );
752 lazy_replace_proj( iftrue, ift2 );
754 // Swap names
755 iffalse = iff2;
756 iftrue = ift2;
757 } else {
758 _igvn.hash_delete(iffalse);
759 _igvn.hash_delete(iftrue);
760 iffalse->set_req_X( 0, le, &_igvn );
761 iftrue ->set_req_X( 0, le, &_igvn );
762 }
764 set_idom(iftrue, le, dd+1);
765 set_idom(iffalse, le, dd+1);
766 assert(iff->outcnt() == 0, "should be dead now");
767 lazy_replace( iff, le ); // fix 'get_ctrl'
769 // Now setup a new CountedLoopNode to replace the existing LoopNode
770 CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control);
771 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
772 // The following assert is approximately true, and defines the intention
773 // of can_be_counted_loop. It fails, however, because phase->type
774 // is not yet initialized for this loop and its parts.
775 //assert(l->can_be_counted_loop(this), "sanity");
776 _igvn.register_new_node_with_optimizer(l);
777 set_loop(l, loop);
778 loop->_head = l;
779 // Fix all data nodes placed at the old loop head.
780 // Uses the lazy-update mechanism of 'get_ctrl'.
781 lazy_replace( x, l );
782 set_idom(l, init_control, dom_depth(x));
784 // Check for immediately preceding SafePoint and remove
785 Node *sfpt2 = le->in(0);
786 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
787 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
788 if (loop->_safepts != NULL) {
789 loop->_safepts->yank(sfpt2);
790 }
791 }
793 // Free up intermediate goo
794 _igvn.remove_dead_node(hook);
796 #ifdef ASSERT
797 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
798 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
799 #endif
800 #ifndef PRODUCT
801 if (TraceLoopOpts) {
802 tty->print("Counted ");
803 loop->dump_head();
804 }
805 #endif
807 C->print_method(PHASE_AFTER_CLOOPS, 3);
809 return true;
810 }
812 //----------------------exact_limit-------------------------------------------
813 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
814 assert(loop->_head->is_CountedLoop(), "");
815 CountedLoopNode *cl = loop->_head->as_CountedLoop();
816 assert(cl->is_valid_counted_loop(), "");
818 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
819 cl->limit()->Opcode() == Op_LoopLimit) {
820 // Old code has exact limit (it could be incorrect in case of int overflow).
821 // Loop limit is exact with stride == 1. And loop may already have exact limit.
822 return cl->limit();
823 }
824 Node *limit = NULL;
825 #ifdef ASSERT
826 BoolTest::mask bt = cl->loopexit()->test_trip();
827 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
828 #endif
829 if (cl->has_exact_trip_count()) {
830 // Simple case: loop has constant boundaries.
831 // Use jlongs to avoid integer overflow.
832 int stride_con = cl->stride_con();
833 jlong init_con = cl->init_trip()->get_int();
834 jlong limit_con = cl->limit()->get_int();
835 julong trip_cnt = cl->trip_count();
836 jlong final_con = init_con + trip_cnt*stride_con;
837 int final_int = (int)final_con;
838 // The final value should be in integer range since the loop
839 // is counted and the limit was checked for overflow.
840 assert(final_con == (jlong)final_int, "final value should be integer");
841 limit = _igvn.intcon(final_int);
842 } else {
843 // Create new LoopLimit node to get exact limit (final iv value).
844 limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
845 register_new_node(limit, cl->in(LoopNode::EntryControl));
846 }
847 assert(limit != NULL, "sanity");
848 return limit;
849 }
851 //------------------------------Ideal------------------------------------------
852 // Return a node which is more "ideal" than the current node.
853 // Attempt to convert into a counted-loop.
854 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
855 if (!can_be_counted_loop(phase)) {
856 phase->C->set_major_progress();
857 }
858 return RegionNode::Ideal(phase, can_reshape);
859 }
862 //=============================================================================
863 //------------------------------Ideal------------------------------------------
864 // Return a node which is more "ideal" than the current node.
865 // Attempt to convert into a counted-loop.
866 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
867 return RegionNode::Ideal(phase, can_reshape);
868 }
870 //------------------------------dump_spec--------------------------------------
871 // Dump special per-node info
872 #ifndef PRODUCT
873 void CountedLoopNode::dump_spec(outputStream *st) const {
874 LoopNode::dump_spec(st);
875 if (stride_is_con()) {
876 st->print("stride: %d ",stride_con());
877 }
878 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
879 if (is_main_loop()) st->print("main of N%d", _idx);
880 if (is_post_loop()) st->print("post of N%d", _main_idx);
881 }
882 #endif
884 //=============================================================================
885 int CountedLoopEndNode::stride_con() const {
886 return stride()->bottom_type()->is_int()->get_con();
887 }
889 //=============================================================================
890 //------------------------------Value-----------------------------------------
891 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
892 const Type* init_t = phase->type(in(Init));
893 const Type* limit_t = phase->type(in(Limit));
894 const Type* stride_t = phase->type(in(Stride));
895 // Either input is TOP ==> the result is TOP
896 if (init_t == Type::TOP) return Type::TOP;
897 if (limit_t == Type::TOP) return Type::TOP;
898 if (stride_t == Type::TOP) return Type::TOP;
900 int stride_con = stride_t->is_int()->get_con();
901 if (stride_con == 1)
902 return NULL; // Identity
904 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
905 // Use jlongs to avoid integer overflow.
906 jlong init_con = init_t->is_int()->get_con();
907 jlong limit_con = limit_t->is_int()->get_con();
908 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
909 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
910 jlong final_con = init_con + stride_con*trip_count;
911 int final_int = (int)final_con;
912 // The final value should be in integer range since the loop
913 // is counted and the limit was checked for overflow.
914 assert(final_con == (jlong)final_int, "final value should be integer");
915 return TypeInt::make(final_int);
916 }
918 return bottom_type(); // TypeInt::INT
919 }
921 //------------------------------Ideal------------------------------------------
922 // Return a node which is more "ideal" than the current node.
923 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
924 if (phase->type(in(Init)) == Type::TOP ||
925 phase->type(in(Limit)) == Type::TOP ||
926 phase->type(in(Stride)) == Type::TOP)
927 return NULL; // Dead
929 int stride_con = phase->type(in(Stride))->is_int()->get_con();
930 if (stride_con == 1)
931 return NULL; // Identity
933 if (in(Init)->is_Con() && in(Limit)->is_Con())
934 return NULL; // Value
936 // Delay following optimizations until all loop optimizations
937 // done to keep Ideal graph simple.
938 if (!can_reshape || phase->C->major_progress())
939 return NULL;
941 const TypeInt* init_t = phase->type(in(Init) )->is_int();
942 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
943 int stride_p;
944 jlong lim, ini;
945 julong max;
946 if (stride_con > 0) {
947 stride_p = stride_con;
948 lim = limit_t->_hi;
949 ini = init_t->_lo;
950 max = (julong)max_jint;
951 } else {
952 stride_p = -stride_con;
953 lim = init_t->_hi;
954 ini = limit_t->_lo;
955 max = (julong)min_jint;
956 }
957 julong range = lim - ini + stride_p;
958 if (range <= max) {
959 // Convert to integer expression if it is not overflow.
960 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
961 Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init)));
962 Node *bias = phase->transform(new (phase->C) AddINode(range, stride_m));
963 Node *trip = phase->transform(new (phase->C) DivINode(0, bias, in(Stride)));
964 Node *span = phase->transform(new (phase->C) MulINode(trip, in(Stride)));
965 return new (phase->C) AddINode(span, in(Init)); // exact limit
966 }
968 if (is_power_of_2(stride_p) || // divisor is 2^n
969 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
970 // Convert to long expression to avoid integer overflow
971 // and let igvn optimizer convert this division.
972 //
973 Node* init = phase->transform( new (phase->C) ConvI2LNode(in(Init)));
974 Node* limit = phase->transform( new (phase->C) ConvI2LNode(in(Limit)));
975 Node* stride = phase->longcon(stride_con);
976 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
978 Node *range = phase->transform(new (phase->C) SubLNode(limit, init));
979 Node *bias = phase->transform(new (phase->C) AddLNode(range, stride_m));
980 Node *span;
981 if (stride_con > 0 && is_power_of_2(stride_p)) {
982 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
983 // and avoid generating rounding for division. Zero trip guard should
984 // guarantee that init < limit but sometimes the guard is missing and
985 // we can get situation when init > limit. Note, for the empty loop
986 // optimization zero trip guard is generated explicitly which leaves
987 // only RCE predicate where exact limit is used and the predicate
988 // will simply fail forcing recompilation.
989 Node* neg_stride = phase->longcon(-stride_con);
990 span = phase->transform(new (phase->C) AndLNode(bias, neg_stride));
991 } else {
992 Node *trip = phase->transform(new (phase->C) DivLNode(0, bias, stride));
993 span = phase->transform(new (phase->C) MulLNode(trip, stride));
994 }
995 // Convert back to int
996 Node *span_int = phase->transform(new (phase->C) ConvL2INode(span));
997 return new (phase->C) AddINode(span_int, in(Init)); // exact limit
998 }
1000 return NULL; // No progress
1001 }
1003 //------------------------------Identity---------------------------------------
1004 // If stride == 1 return limit node.
1005 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
1006 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1007 if (stride_con == 1 || stride_con == -1)
1008 return in(Limit);
1009 return this;
1010 }
1012 //=============================================================================
1013 //----------------------match_incr_with_optional_truncation--------------------
1014 // Match increment with optional truncation:
1015 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1016 // Return NULL for failure. Success returns the increment node.
1017 Node* CountedLoopNode::match_incr_with_optional_truncation(
1018 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1019 // Quick cutouts:
1020 if (expr == NULL || expr->req() != 3) return NULL;
1022 Node *t1 = NULL;
1023 Node *t2 = NULL;
1024 const TypeInt* trunc_t = TypeInt::INT;
1025 Node* n1 = expr;
1026 int n1op = n1->Opcode();
1028 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1029 if (n1op == Op_AndI &&
1030 n1->in(2)->is_Con() &&
1031 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1032 // %%% This check should match any mask of 2**K-1.
1033 t1 = n1;
1034 n1 = t1->in(1);
1035 n1op = n1->Opcode();
1036 trunc_t = TypeInt::CHAR;
1037 } else if (n1op == Op_RShiftI &&
1038 n1->in(1) != NULL &&
1039 n1->in(1)->Opcode() == Op_LShiftI &&
1040 n1->in(2) == n1->in(1)->in(2) &&
1041 n1->in(2)->is_Con()) {
1042 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1043 // %%% This check should match any shift in [1..31].
1044 if (shift == 16 || shift == 8) {
1045 t1 = n1;
1046 t2 = t1->in(1);
1047 n1 = t2->in(1);
1048 n1op = n1->Opcode();
1049 if (shift == 16) {
1050 trunc_t = TypeInt::SHORT;
1051 } else if (shift == 8) {
1052 trunc_t = TypeInt::BYTE;
1053 }
1054 }
1055 }
1057 // If (maybe after stripping) it is an AddI, we won:
1058 if (n1op == Op_AddI) {
1059 *trunc1 = t1;
1060 *trunc2 = t2;
1061 *trunc_type = trunc_t;
1062 return n1;
1063 }
1065 // failed
1066 return NULL;
1067 }
1070 //------------------------------filtered_type--------------------------------
1071 // Return a type based on condition control flow
1072 // A successful return will be a type that is restricted due
1073 // to a series of dominating if-tests, such as:
1074 // if (i < 10) {
1075 // if (i > 0) {
1076 // here: "i" type is [1..10)
1077 // }
1078 // }
1079 // or a control flow merge
1080 // if (i < 10) {
1081 // do {
1082 // phi( , ) -- at top of loop type is [min_int..10)
1083 // i = ?
1084 // } while ( i < 10)
1085 //
1086 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1087 assert(n && n->bottom_type()->is_int(), "must be int");
1088 const TypeInt* filtered_t = NULL;
1089 if (!n->is_Phi()) {
1090 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1091 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1093 } else {
1094 Node* phi = n->as_Phi();
1095 Node* region = phi->in(0);
1096 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1097 if (region && region != C->top()) {
1098 for (uint i = 1; i < phi->req(); i++) {
1099 Node* val = phi->in(i);
1100 Node* use_c = region->in(i);
1101 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1102 if (val_t != NULL) {
1103 if (filtered_t == NULL) {
1104 filtered_t = val_t;
1105 } else {
1106 filtered_t = filtered_t->meet(val_t)->is_int();
1107 }
1108 }
1109 }
1110 }
1111 }
1112 const TypeInt* n_t = _igvn.type(n)->is_int();
1113 if (filtered_t != NULL) {
1114 n_t = n_t->join(filtered_t)->is_int();
1115 }
1116 return n_t;
1117 }
1120 //------------------------------filtered_type_from_dominators--------------------------------
1121 // Return a possibly more restrictive type for val based on condition control flow of dominators
1122 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1123 if (val->is_Con()) {
1124 return val->bottom_type()->is_int();
1125 }
1126 uint if_limit = 10; // Max number of dominating if's visited
1127 const TypeInt* rtn_t = NULL;
1129 if (use_ctrl && use_ctrl != C->top()) {
1130 Node* val_ctrl = get_ctrl(val);
1131 uint val_dom_depth = dom_depth(val_ctrl);
1132 Node* pred = use_ctrl;
1133 uint if_cnt = 0;
1134 while (if_cnt < if_limit) {
1135 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1136 if_cnt++;
1137 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1138 if (if_t != NULL) {
1139 if (rtn_t == NULL) {
1140 rtn_t = if_t;
1141 } else {
1142 rtn_t = rtn_t->join(if_t)->is_int();
1143 }
1144 }
1145 }
1146 pred = idom(pred);
1147 if (pred == NULL || pred == C->top()) {
1148 break;
1149 }
1150 // Stop if going beyond definition block of val
1151 if (dom_depth(pred) < val_dom_depth) {
1152 break;
1153 }
1154 }
1155 }
1156 return rtn_t;
1157 }
1160 //------------------------------dump_spec--------------------------------------
1161 // Dump special per-node info
1162 #ifndef PRODUCT
1163 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1164 if( in(TestValue)->is_Bool() ) {
1165 BoolTest bt( test_trip()); // Added this for g++.
1167 st->print("[");
1168 bt.dump_on(st);
1169 st->print("]");
1170 }
1171 st->print(" ");
1172 IfNode::dump_spec(st);
1173 }
1174 #endif
1176 //=============================================================================
1177 //------------------------------is_member--------------------------------------
1178 // Is 'l' a member of 'this'?
1179 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1180 while( l->_nest > _nest ) l = l->_parent;
1181 return l == this;
1182 }
1184 //------------------------------set_nest---------------------------------------
1185 // Set loop tree nesting depth. Accumulate _has_call bits.
1186 int IdealLoopTree::set_nest( uint depth ) {
1187 _nest = depth;
1188 int bits = _has_call;
1189 if( _child ) bits |= _child->set_nest(depth+1);
1190 if( bits ) _has_call = 1;
1191 if( _next ) bits |= _next ->set_nest(depth );
1192 return bits;
1193 }
1195 //------------------------------split_fall_in----------------------------------
1196 // Split out multiple fall-in edges from the loop header. Move them to a
1197 // private RegionNode before the loop. This becomes the loop landing pad.
1198 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1199 PhaseIterGVN &igvn = phase->_igvn;
1200 uint i;
1202 // Make a new RegionNode to be the landing pad.
1203 Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 );
1204 phase->set_loop(landing_pad,_parent);
1205 // Gather all the fall-in control paths into the landing pad
1206 uint icnt = fall_in_cnt;
1207 uint oreq = _head->req();
1208 for( i = oreq-1; i>0; i-- )
1209 if( !phase->is_member( this, _head->in(i) ) )
1210 landing_pad->set_req(icnt--,_head->in(i));
1212 // Peel off PhiNode edges as well
1213 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1214 Node *oj = _head->fast_out(j);
1215 if( oj->is_Phi() ) {
1216 PhiNode* old_phi = oj->as_Phi();
1217 assert( old_phi->region() == _head, "" );
1218 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1219 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1220 uint icnt = fall_in_cnt;
1221 for( i = oreq-1; i>0; i-- ) {
1222 if( !phase->is_member( this, _head->in(i) ) ) {
1223 p->init_req(icnt--, old_phi->in(i));
1224 // Go ahead and clean out old edges from old phi
1225 old_phi->del_req(i);
1226 }
1227 }
1228 // Search for CSE's here, because ZKM.jar does a lot of
1229 // loop hackery and we need to be a little incremental
1230 // with the CSE to avoid O(N^2) node blow-up.
1231 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1232 if( p2 ) { // Found CSE
1233 p->destruct(); // Recover useless new node
1234 p = p2; // Use old node
1235 } else {
1236 igvn.register_new_node_with_optimizer(p, old_phi);
1237 }
1238 // Make old Phi refer to new Phi.
1239 old_phi->add_req(p);
1240 // Check for the special case of making the old phi useless and
1241 // disappear it. In JavaGrande I have a case where this useless
1242 // Phi is the loop limit and prevents recognizing a CountedLoop
1243 // which in turn prevents removing an empty loop.
1244 Node *id_old_phi = old_phi->Identity( &igvn );
1245 if( id_old_phi != old_phi ) { // Found a simple identity?
1246 // Note that I cannot call 'replace_node' here, because
1247 // that will yank the edge from old_phi to the Region and
1248 // I'm mid-iteration over the Region's uses.
1249 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1250 Node* use = old_phi->last_out(i);
1251 igvn.rehash_node_delayed(use);
1252 uint uses_found = 0;
1253 for (uint j = 0; j < use->len(); j++) {
1254 if (use->in(j) == old_phi) {
1255 if (j < use->req()) use->set_req (j, id_old_phi);
1256 else use->set_prec(j, id_old_phi);
1257 uses_found++;
1258 }
1259 }
1260 i -= uses_found; // we deleted 1 or more copies of this edge
1261 }
1262 }
1263 igvn._worklist.push(old_phi);
1264 }
1265 }
1266 // Finally clean out the fall-in edges from the RegionNode
1267 for( i = oreq-1; i>0; i-- ) {
1268 if( !phase->is_member( this, _head->in(i) ) ) {
1269 _head->del_req(i);
1270 }
1271 }
1272 // Transform landing pad
1273 igvn.register_new_node_with_optimizer(landing_pad, _head);
1274 // Insert landing pad into the header
1275 _head->add_req(landing_pad);
1276 }
1278 //------------------------------split_outer_loop-------------------------------
1279 // Split out the outermost loop from this shared header.
1280 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1281 PhaseIterGVN &igvn = phase->_igvn;
1283 // Find index of outermost loop; it should also be my tail.
1284 uint outer_idx = 1;
1285 while( _head->in(outer_idx) != _tail ) outer_idx++;
1287 // Make a LoopNode for the outermost loop.
1288 Node *ctl = _head->in(LoopNode::EntryControl);
1289 Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) );
1290 outer = igvn.register_new_node_with_optimizer(outer, _head);
1291 phase->set_created_loop_node();
1293 // Outermost loop falls into '_head' loop
1294 _head->set_req(LoopNode::EntryControl, outer);
1295 _head->del_req(outer_idx);
1296 // Split all the Phis up between '_head' loop and 'outer' loop.
1297 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1298 Node *out = _head->fast_out(j);
1299 if( out->is_Phi() ) {
1300 PhiNode *old_phi = out->as_Phi();
1301 assert( old_phi->region() == _head, "" );
1302 Node *phi = PhiNode::make_blank(outer, old_phi);
1303 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1304 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1305 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1306 // Make old Phi point to new Phi on the fall-in path
1307 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1308 old_phi->del_req(outer_idx);
1309 }
1310 }
1312 // Use the new loop head instead of the old shared one
1313 _head = outer;
1314 phase->set_loop(_head, this);
1315 }
1317 //------------------------------fix_parent-------------------------------------
1318 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1319 loop->_parent = parent;
1320 if( loop->_child ) fix_parent( loop->_child, loop );
1321 if( loop->_next ) fix_parent( loop->_next , parent );
1322 }
1324 //------------------------------estimate_path_freq-----------------------------
1325 static float estimate_path_freq( Node *n ) {
1326 // Try to extract some path frequency info
1327 IfNode *iff;
1328 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1329 uint nop = n->Opcode();
1330 if( nop == Op_SafePoint ) { // Skip any safepoint
1331 n = n->in(0);
1332 continue;
1333 }
1334 if( nop == Op_CatchProj ) { // Get count from a prior call
1335 // Assume call does not always throw exceptions: means the call-site
1336 // count is also the frequency of the fall-through path.
1337 assert( n->is_CatchProj(), "" );
1338 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1339 return 0.0f; // Assume call exception path is rare
1340 Node *call = n->in(0)->in(0)->in(0);
1341 assert( call->is_Call(), "expect a call here" );
1342 const JVMState *jvms = ((CallNode*)call)->jvms();
1343 ciMethodData* methodData = jvms->method()->method_data();
1344 if (!methodData->is_mature()) return 0.0f; // No call-site data
1345 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1346 if ((data == NULL) || !data->is_CounterData()) {
1347 // no call profile available, try call's control input
1348 n = n->in(0);
1349 continue;
1350 }
1351 return data->as_CounterData()->count()/FreqCountInvocations;
1352 }
1353 // See if there's a gating IF test
1354 Node *n_c = n->in(0);
1355 if( !n_c->is_If() ) break; // No estimate available
1356 iff = n_c->as_If();
1357 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1358 // Compute how much count comes on this path
1359 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1360 // Have no count info. Skip dull uncommon-trap like branches.
1361 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1362 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1363 break;
1364 // Skip through never-taken branch; look for a real loop exit.
1365 n = iff->in(0);
1366 }
1367 return 0.0f; // No estimate available
1368 }
1370 //------------------------------merge_many_backedges---------------------------
1371 // Merge all the backedges from the shared header into a private Region.
1372 // Feed that region as the one backedge to this loop.
1373 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1374 uint i;
1376 // Scan for the top 2 hottest backedges
1377 float hotcnt = 0.0f;
1378 float warmcnt = 0.0f;
1379 uint hot_idx = 0;
1380 // Loop starts at 2 because slot 1 is the fall-in path
1381 for( i = 2; i < _head->req(); i++ ) {
1382 float cnt = estimate_path_freq(_head->in(i));
1383 if( cnt > hotcnt ) { // Grab hottest path
1384 warmcnt = hotcnt;
1385 hotcnt = cnt;
1386 hot_idx = i;
1387 } else if( cnt > warmcnt ) { // And 2nd hottest path
1388 warmcnt = cnt;
1389 }
1390 }
1392 // See if the hottest backedge is worthy of being an inner loop
1393 // by being much hotter than the next hottest backedge.
1394 if( hotcnt <= 0.0001 ||
1395 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1397 // Peel out the backedges into a private merge point; peel
1398 // them all except optionally hot_idx.
1399 PhaseIterGVN &igvn = phase->_igvn;
1401 Node *hot_tail = NULL;
1402 // Make a Region for the merge point
1403 Node *r = new (phase->C) RegionNode(1);
1404 for( i = 2; i < _head->req(); i++ ) {
1405 if( i != hot_idx )
1406 r->add_req( _head->in(i) );
1407 else hot_tail = _head->in(i);
1408 }
1409 igvn.register_new_node_with_optimizer(r, _head);
1410 // Plug region into end of loop _head, followed by hot_tail
1411 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1412 _head->set_req(2, r);
1413 if( hot_idx ) _head->add_req(hot_tail);
1415 // Split all the Phis up between '_head' loop and the Region 'r'
1416 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1417 Node *out = _head->fast_out(j);
1418 if( out->is_Phi() ) {
1419 PhiNode* n = out->as_Phi();
1420 igvn.hash_delete(n); // Delete from hash before hacking edges
1421 Node *hot_phi = NULL;
1422 Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type());
1423 // Check all inputs for the ones to peel out
1424 uint j = 1;
1425 for( uint i = 2; i < n->req(); i++ ) {
1426 if( i != hot_idx )
1427 phi->set_req( j++, n->in(i) );
1428 else hot_phi = n->in(i);
1429 }
1430 // Register the phi but do not transform until whole place transforms
1431 igvn.register_new_node_with_optimizer(phi, n);
1432 // Add the merge phi to the old Phi
1433 while( n->req() > 3 ) n->del_req( n->req()-1 );
1434 n->set_req(2, phi);
1435 if( hot_idx ) n->add_req(hot_phi);
1436 }
1437 }
1440 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1441 // of self loop tree. Turn self into a loop headed by _head and with
1442 // tail being the new merge point.
1443 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1444 phase->set_loop(_tail,ilt); // Adjust tail
1445 _tail = r; // Self's tail is new merge point
1446 phase->set_loop(r,this);
1447 ilt->_child = _child; // New guy has my children
1448 _child = ilt; // Self has new guy as only child
1449 ilt->_parent = this; // new guy has self for parent
1450 ilt->_nest = _nest; // Same nesting depth (for now)
1452 // Starting with 'ilt', look for child loop trees using the same shared
1453 // header. Flatten these out; they will no longer be loops in the end.
1454 IdealLoopTree **pilt = &_child;
1455 while( ilt ) {
1456 if( ilt->_head == _head ) {
1457 uint i;
1458 for( i = 2; i < _head->req(); i++ )
1459 if( _head->in(i) == ilt->_tail )
1460 break; // Still a loop
1461 if( i == _head->req() ) { // No longer a loop
1462 // Flatten ilt. Hang ilt's "_next" list from the end of
1463 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1464 IdealLoopTree **cp = &ilt->_child;
1465 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1466 *cp = ilt->_next; // Hang next list at end of child list
1467 *pilt = ilt->_child; // Move child up to replace ilt
1468 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1469 ilt = ilt->_child; // Repeat using new ilt
1470 continue; // do not advance over ilt->_child
1471 }
1472 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1473 phase->set_loop(_head,ilt);
1474 }
1475 pilt = &ilt->_child; // Advance to next
1476 ilt = *pilt;
1477 }
1479 if( _child ) fix_parent( _child, this );
1480 }
1482 //------------------------------beautify_loops---------------------------------
1483 // Split shared headers and insert loop landing pads.
1484 // Insert a LoopNode to replace the RegionNode.
1485 // Return TRUE if loop tree is structurally changed.
1486 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1487 bool result = false;
1488 // Cache parts in locals for easy
1489 PhaseIterGVN &igvn = phase->_igvn;
1491 igvn.hash_delete(_head); // Yank from hash before hacking edges
1493 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1494 int fall_in_cnt = 0;
1495 for( uint i = 1; i < _head->req(); i++ )
1496 if( !phase->is_member( this, _head->in(i) ) )
1497 fall_in_cnt++;
1498 assert( fall_in_cnt, "at least 1 fall-in path" );
1499 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1500 split_fall_in( phase, fall_in_cnt );
1502 // Swap inputs to the _head and all Phis to move the fall-in edge to
1503 // the left.
1504 fall_in_cnt = 1;
1505 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1506 fall_in_cnt++;
1507 if( fall_in_cnt > 1 ) {
1508 // Since I am just swapping inputs I do not need to update def-use info
1509 Node *tmp = _head->in(1);
1510 _head->set_req( 1, _head->in(fall_in_cnt) );
1511 _head->set_req( fall_in_cnt, tmp );
1512 // Swap also all Phis
1513 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1514 Node* phi = _head->fast_out(i);
1515 if( phi->is_Phi() ) {
1516 igvn.hash_delete(phi); // Yank from hash before hacking edges
1517 tmp = phi->in(1);
1518 phi->set_req( 1, phi->in(fall_in_cnt) );
1519 phi->set_req( fall_in_cnt, tmp );
1520 }
1521 }
1522 }
1523 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1524 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1526 // If I am a shared header (multiple backedges), peel off the many
1527 // backedges into a private merge point and use the merge point as
1528 // the one true backedge.
1529 if( _head->req() > 3 ) {
1530 // Merge the many backedges into a single backedge but leave
1531 // the hottest backedge as separate edge for the following peel.
1532 merge_many_backedges( phase );
1533 result = true;
1534 }
1536 // If I have one hot backedge, peel off myself loop.
1537 // I better be the outermost loop.
1538 if (_head->req() > 3 && !_irreducible) {
1539 split_outer_loop( phase );
1540 result = true;
1542 } else if (!_head->is_Loop() && !_irreducible) {
1543 // Make a new LoopNode to replace the old loop head
1544 Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) );
1545 l = igvn.register_new_node_with_optimizer(l, _head);
1546 phase->set_created_loop_node();
1547 // Go ahead and replace _head
1548 phase->_igvn.replace_node( _head, l );
1549 _head = l;
1550 phase->set_loop(_head, this);
1551 }
1553 // Now recursively beautify nested loops
1554 if( _child ) result |= _child->beautify_loops( phase );
1555 if( _next ) result |= _next ->beautify_loops( phase );
1556 return result;
1557 }
1559 //------------------------------allpaths_check_safepts----------------------------
1560 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1561 // encountered. Helper for check_safepts.
1562 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1563 assert(stack.size() == 0, "empty stack");
1564 stack.push(_tail);
1565 visited.Clear();
1566 visited.set(_tail->_idx);
1567 while (stack.size() > 0) {
1568 Node* n = stack.pop();
1569 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1570 // Terminate this path
1571 } else if (n->Opcode() == Op_SafePoint) {
1572 if (_phase->get_loop(n) != this) {
1573 if (_required_safept == NULL) _required_safept = new Node_List();
1574 _required_safept->push(n); // save the one closest to the tail
1575 }
1576 // Terminate this path
1577 } else {
1578 uint start = n->is_Region() ? 1 : 0;
1579 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1580 for (uint i = start; i < end; i++) {
1581 Node* in = n->in(i);
1582 assert(in->is_CFG(), "must be");
1583 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1584 stack.push(in);
1585 }
1586 }
1587 }
1588 }
1589 }
1591 //------------------------------check_safepts----------------------------
1592 // Given dominators, try to find loops with calls that must always be
1593 // executed (call dominates loop tail). These loops do not need non-call
1594 // safepoints (ncsfpt).
1595 //
1596 // A complication is that a safepoint in a inner loop may be needed
1597 // by an outer loop. In the following, the inner loop sees it has a
1598 // call (block 3) on every path from the head (block 2) to the
1599 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1600 // in block 2, _but_ this leaves the outer loop without a safepoint.
1601 //
1602 // entry 0
1603 // |
1604 // v
1605 // outer 1,2 +->1
1606 // | |
1607 // | v
1608 // | 2<---+ ncsfpt in 2
1609 // |_/|\ |
1610 // | v |
1611 // inner 2,3 / 3 | call in 3
1612 // / | |
1613 // v +--+
1614 // exit 4
1615 //
1616 //
1617 // This method creates a list (_required_safept) of ncsfpt nodes that must
1618 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1619 // is first looked for in the lists for the outer loops of the current loop.
1620 //
1621 // The insights into the problem:
1622 // A) counted loops are okay
1623 // B) innermost loops are okay (only an inner loop can delete
1624 // a ncsfpt needed by an outer loop)
1625 // C) a loop is immune from an inner loop deleting a safepoint
1626 // if the loop has a call on the idom-path
1627 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1628 // idom-path that is not in a nested loop
1629 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1630 // loop needs to be prevented from deletion by an inner loop
1631 //
1632 // There are two analyses:
1633 // 1) The first, and cheaper one, scans the loop body from
1634 // tail to head following the idom (immediate dominator)
1635 // chain, looking for the cases (C,D,E) above.
1636 // Since inner loops are scanned before outer loops, there is summary
1637 // information about inner loops. Inner loops can be skipped over
1638 // when the tail of an inner loop is encountered.
1639 //
1640 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1641 // the idom path (which is rare), scans all predecessor control paths
1642 // from the tail to the head, terminating a path when a call or sfpt
1643 // is encountered, to find the ncsfpt's that are closest to the tail.
1644 //
1645 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1646 // Bottom up traversal
1647 IdealLoopTree* ch = _child;
1648 if (_child) _child->check_safepts(visited, stack);
1649 if (_next) _next ->check_safepts(visited, stack);
1651 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1652 bool has_call = false; // call on dom-path
1653 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1654 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1655 // Scan the dom-path nodes from tail to head
1656 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1657 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1658 has_call = true;
1659 _has_sfpt = 1; // Then no need for a safept!
1660 break;
1661 } else if (n->Opcode() == Op_SafePoint) {
1662 if (_phase->get_loop(n) == this) {
1663 has_local_ncsfpt = true;
1664 break;
1665 }
1666 if (nonlocal_ncsfpt == NULL) {
1667 nonlocal_ncsfpt = n; // save the one closest to the tail
1668 }
1669 } else {
1670 IdealLoopTree* nlpt = _phase->get_loop(n);
1671 if (this != nlpt) {
1672 // If at an inner loop tail, see if the inner loop has already
1673 // recorded seeing a call on the dom-path (and stop.) If not,
1674 // jump to the head of the inner loop.
1675 assert(is_member(nlpt), "nested loop");
1676 Node* tail = nlpt->_tail;
1677 if (tail->in(0)->is_If()) tail = tail->in(0);
1678 if (n == tail) {
1679 // If inner loop has call on dom-path, so does outer loop
1680 if (nlpt->_has_sfpt) {
1681 has_call = true;
1682 _has_sfpt = 1;
1683 break;
1684 }
1685 // Skip to head of inner loop
1686 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1687 n = nlpt->_head;
1688 }
1689 }
1690 }
1691 }
1692 // Record safept's that this loop needs preserved when an
1693 // inner loop attempts to delete it's safepoints.
1694 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1695 if (nonlocal_ncsfpt != NULL) {
1696 if (_required_safept == NULL) _required_safept = new Node_List();
1697 _required_safept->push(nonlocal_ncsfpt);
1698 } else {
1699 // Failed to find a suitable safept on the dom-path. Now use
1700 // an all paths walk from tail to head, looking for safepoints to preserve.
1701 allpaths_check_safepts(visited, stack);
1702 }
1703 }
1704 }
1705 }
1707 //---------------------------is_deleteable_safept----------------------------
1708 // Is safept not required by an outer loop?
1709 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1710 assert(sfpt->Opcode() == Op_SafePoint, "");
1711 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1712 while (lp != NULL) {
1713 Node_List* sfpts = lp->_required_safept;
1714 if (sfpts != NULL) {
1715 for (uint i = 0; i < sfpts->size(); i++) {
1716 if (sfpt == sfpts->at(i))
1717 return false;
1718 }
1719 }
1720 lp = lp->_parent;
1721 }
1722 return true;
1723 }
1725 //---------------------------replace_parallel_iv-------------------------------
1726 // Replace parallel induction variable (parallel to trip counter)
1727 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1728 assert(loop->_head->is_CountedLoop(), "");
1729 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1730 if (!cl->is_valid_counted_loop())
1731 return; // skip malformed counted loop
1732 Node *incr = cl->incr();
1733 if (incr == NULL)
1734 return; // Dead loop?
1735 Node *init = cl->init_trip();
1736 Node *phi = cl->phi();
1737 int stride_con = cl->stride_con();
1739 // Visit all children, looking for Phis
1740 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1741 Node *out = cl->out(i);
1742 // Look for other phis (secondary IVs). Skip dead ones
1743 if (!out->is_Phi() || out == phi || !has_node(out))
1744 continue;
1745 PhiNode* phi2 = out->as_Phi();
1746 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1747 // Look for induction variables of the form: X += constant
1748 if (phi2->region() != loop->_head ||
1749 incr2->req() != 3 ||
1750 incr2->in(1) != phi2 ||
1751 incr2 == incr ||
1752 incr2->Opcode() != Op_AddI ||
1753 !incr2->in(2)->is_Con())
1754 continue;
1756 // Check for parallel induction variable (parallel to trip counter)
1757 // via an affine function. In particular, count-down loops with
1758 // count-up array indices are common. We only RCE references off
1759 // the trip-counter, so we need to convert all these to trip-counter
1760 // expressions.
1761 Node *init2 = phi2->in( LoopNode::EntryControl );
1762 int stride_con2 = incr2->in(2)->get_int();
1764 // The general case here gets a little tricky. We want to find the
1765 // GCD of all possible parallel IV's and make a new IV using this
1766 // GCD for the loop. Then all possible IVs are simple multiples of
1767 // the GCD. In practice, this will cover very few extra loops.
1768 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1769 // where +/-1 is the common case, but other integer multiples are
1770 // also easy to handle.
1771 int ratio_con = stride_con2/stride_con;
1773 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1774 #ifndef PRODUCT
1775 if (TraceLoopOpts) {
1776 tty->print("Parallel IV: %d ", phi2->_idx);
1777 loop->dump_head();
1778 }
1779 #endif
1780 // Convert to using the trip counter. The parallel induction
1781 // variable differs from the trip counter by a loop-invariant
1782 // amount, the difference between their respective initial values.
1783 // It is scaled by the 'ratio_con'.
1784 Node* ratio = _igvn.intcon(ratio_con);
1785 set_ctrl(ratio, C->root());
1786 Node* ratio_init = new (C) MulINode(init, ratio);
1787 _igvn.register_new_node_with_optimizer(ratio_init, init);
1788 set_early_ctrl(ratio_init);
1789 Node* diff = new (C) SubINode(init2, ratio_init);
1790 _igvn.register_new_node_with_optimizer(diff, init2);
1791 set_early_ctrl(diff);
1792 Node* ratio_idx = new (C) MulINode(phi, ratio);
1793 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1794 set_ctrl(ratio_idx, cl);
1795 Node* add = new (C) AddINode(ratio_idx, diff);
1796 _igvn.register_new_node_with_optimizer(add);
1797 set_ctrl(add, cl);
1798 _igvn.replace_node( phi2, add );
1799 // Sometimes an induction variable is unused
1800 if (add->outcnt() == 0) {
1801 _igvn.remove_dead_node(add);
1802 }
1803 --i; // deleted this phi; rescan starting with next position
1804 continue;
1805 }
1806 }
1807 }
1809 //------------------------------counted_loop-----------------------------------
1810 // Convert to counted loops where possible
1811 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1813 // For grins, set the inner-loop flag here
1814 if (!_child) {
1815 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1816 }
1818 if (_head->is_CountedLoop() ||
1819 phase->is_counted_loop(_head, this)) {
1820 _has_sfpt = 1; // Indicate we do not need a safepoint here
1822 // Look for safepoints to remove.
1823 Node_List* sfpts = _safepts;
1824 if (sfpts != NULL) {
1825 for (uint i = 0; i < sfpts->size(); i++) {
1826 Node* n = sfpts->at(i);
1827 assert(phase->get_loop(n) == this, "");
1828 if (phase->is_deleteable_safept(n)) {
1829 phase->lazy_replace(n, n->in(TypeFunc::Control));
1830 }
1831 }
1832 }
1834 // Look for induction variables
1835 phase->replace_parallel_iv(this);
1837 } else if (_parent != NULL && !_irreducible) {
1838 // Not a counted loop.
1839 // Look for a safepoint on the idom-path.
1840 Node* sfpt = tail();
1841 for (; sfpt != _head; sfpt = phase->idom(sfpt)) {
1842 if (sfpt->Opcode() == Op_SafePoint && phase->get_loop(sfpt) == this)
1843 break; // Found one
1844 }
1845 // Delete other safepoints in this loop.
1846 Node_List* sfpts = _safepts;
1847 if (sfpts != NULL && sfpt != _head && sfpt->Opcode() == Op_SafePoint) {
1848 for (uint i = 0; i < sfpts->size(); i++) {
1849 Node* n = sfpts->at(i);
1850 assert(phase->get_loop(n) == this, "");
1851 if (n != sfpt && phase->is_deleteable_safept(n)) {
1852 phase->lazy_replace(n, n->in(TypeFunc::Control));
1853 }
1854 }
1855 }
1856 }
1858 // Recursively
1859 if (_child) _child->counted_loop( phase );
1860 if (_next) _next ->counted_loop( phase );
1861 }
1863 #ifndef PRODUCT
1864 //------------------------------dump_head--------------------------------------
1865 // Dump 1 liner for loop header info
1866 void IdealLoopTree::dump_head( ) const {
1867 for (uint i=0; i<_nest; i++)
1868 tty->print(" ");
1869 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1870 if (_irreducible) tty->print(" IRREDUCIBLE");
1871 Node* entry = _head->in(LoopNode::EntryControl);
1872 if (LoopLimitCheck) {
1873 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1874 if (predicate != NULL ) {
1875 tty->print(" limit_check");
1876 entry = entry->in(0)->in(0);
1877 }
1878 }
1879 if (UseLoopPredicate) {
1880 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1881 if (entry != NULL) {
1882 tty->print(" predicated");
1883 }
1884 }
1885 if (_head->is_CountedLoop()) {
1886 CountedLoopNode *cl = _head->as_CountedLoop();
1887 tty->print(" counted");
1889 Node* init_n = cl->init_trip();
1890 if (init_n != NULL && init_n->is_Con())
1891 tty->print(" [%d,", cl->init_trip()->get_int());
1892 else
1893 tty->print(" [int,");
1894 Node* limit_n = cl->limit();
1895 if (limit_n != NULL && limit_n->is_Con())
1896 tty->print("%d),", cl->limit()->get_int());
1897 else
1898 tty->print("int),");
1899 int stride_con = cl->stride_con();
1900 if (stride_con > 0) tty->print("+");
1901 tty->print("%d", stride_con);
1903 tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
1905 if (cl->is_pre_loop ()) tty->print(" pre" );
1906 if (cl->is_main_loop()) tty->print(" main");
1907 if (cl->is_post_loop()) tty->print(" post");
1908 }
1909 tty->cr();
1910 }
1912 //------------------------------dump-------------------------------------------
1913 // Dump loops by loop tree
1914 void IdealLoopTree::dump( ) const {
1915 dump_head();
1916 if (_child) _child->dump();
1917 if (_next) _next ->dump();
1918 }
1920 #endif
1922 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1923 if (loop == root) {
1924 if (loop->_child != NULL) {
1925 log->begin_head("loop_tree");
1926 log->end_head();
1927 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1928 log->tail("loop_tree");
1929 assert(loop->_next == NULL, "what?");
1930 }
1931 } else {
1932 Node* head = loop->_head;
1933 log->begin_head("loop");
1934 log->print(" idx='%d' ", head->_idx);
1935 if (loop->_irreducible) log->print("irreducible='1' ");
1936 if (head->is_Loop()) {
1937 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1938 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1939 }
1940 if (head->is_CountedLoop()) {
1941 CountedLoopNode* cl = head->as_CountedLoop();
1942 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1943 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1944 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1945 }
1946 log->end_head();
1947 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1948 log->tail("loop");
1949 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1950 }
1951 }
1953 //---------------------collect_potentially_useful_predicates-----------------------
1954 // Helper function to collect potentially useful predicates to prevent them from
1955 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1956 void PhaseIdealLoop::collect_potentially_useful_predicates(
1957 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1958 if (loop->_child) { // child
1959 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1960 }
1962 // self (only loops that we can apply loop predication may use their predicates)
1963 if (loop->_head->is_Loop() &&
1964 !loop->_irreducible &&
1965 !loop->tail()->is_top()) {
1966 LoopNode* lpn = loop->_head->as_Loop();
1967 Node* entry = lpn->in(LoopNode::EntryControl);
1968 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1969 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1970 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1971 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1972 entry = entry->in(0)->in(0);
1973 }
1974 predicate_proj = find_predicate(entry); // Predicate
1975 if (predicate_proj != NULL ) {
1976 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1977 }
1978 }
1980 if (loop->_next) { // sibling
1981 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1982 }
1983 }
1985 //------------------------eliminate_useless_predicates-----------------------------
1986 // Eliminate all inserted predicates if they could not be used by loop predication.
1987 // Note: it will also eliminates loop limits check predicate since it also uses
1988 // Opaque1 node (see Parse::add_predicate()).
1989 void PhaseIdealLoop::eliminate_useless_predicates() {
1990 if (C->predicate_count() == 0)
1991 return; // no predicate left
1993 Unique_Node_List useful_predicates; // to store useful predicates
1994 if (C->has_loops()) {
1995 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1996 }
1998 for (int i = C->predicate_count(); i > 0; i--) {
1999 Node * n = C->predicate_opaque1_node(i-1);
2000 assert(n->Opcode() == Op_Opaque1, "must be");
2001 if (!useful_predicates.member(n)) { // not in the useful list
2002 _igvn.replace_node(n, n->in(1));
2003 }
2004 }
2005 }
2007 //------------------------process_expensive_nodes-----------------------------
2008 // Expensive nodes have their control input set to prevent the GVN
2009 // from commoning them and as a result forcing the resulting node to
2010 // be in a more frequent path. Use CFG information here, to change the
2011 // control inputs so that some expensive nodes can be commoned while
2012 // not executed more frequently.
2013 bool PhaseIdealLoop::process_expensive_nodes() {
2014 assert(OptimizeExpensiveOps, "optimization off?");
2016 // Sort nodes to bring similar nodes together
2017 C->sort_expensive_nodes();
2019 bool progress = false;
2021 for (int i = 0; i < C->expensive_count(); ) {
2022 Node* n = C->expensive_node(i);
2023 int start = i;
2024 // Find nodes similar to n
2025 i++;
2026 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2027 int end = i;
2028 // And compare them two by two
2029 for (int j = start; j < end; j++) {
2030 Node* n1 = C->expensive_node(j);
2031 if (is_node_unreachable(n1)) {
2032 continue;
2033 }
2034 for (int k = j+1; k < end; k++) {
2035 Node* n2 = C->expensive_node(k);
2036 if (is_node_unreachable(n2)) {
2037 continue;
2038 }
2040 assert(n1 != n2, "should be pair of nodes");
2042 Node* c1 = n1->in(0);
2043 Node* c2 = n2->in(0);
2045 Node* parent_c1 = c1;
2046 Node* parent_c2 = c2;
2048 // The call to get_early_ctrl_for_expensive() moves the
2049 // expensive nodes up but stops at loops that are in a if
2050 // branch. See whether we can exit the loop and move above the
2051 // If.
2052 if (c1->is_Loop()) {
2053 parent_c1 = c1->in(1);
2054 }
2055 if (c2->is_Loop()) {
2056 parent_c2 = c2->in(1);
2057 }
2059 if (parent_c1 == parent_c2) {
2060 _igvn._worklist.push(n1);
2061 _igvn._worklist.push(n2);
2062 continue;
2063 }
2065 // Look for identical expensive node up the dominator chain.
2066 if (is_dominator(c1, c2)) {
2067 c2 = c1;
2068 } else if (is_dominator(c2, c1)) {
2069 c1 = c2;
2070 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2071 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2072 // Both branches have the same expensive node so move it up
2073 // before the if.
2074 c1 = c2 = idom(parent_c1->in(0));
2075 }
2076 // Do the actual moves
2077 if (n1->in(0) != c1) {
2078 _igvn.hash_delete(n1);
2079 n1->set_req(0, c1);
2080 _igvn.hash_insert(n1);
2081 _igvn._worklist.push(n1);
2082 progress = true;
2083 }
2084 if (n2->in(0) != c2) {
2085 _igvn.hash_delete(n2);
2086 n2->set_req(0, c2);
2087 _igvn.hash_insert(n2);
2088 _igvn._worklist.push(n2);
2089 progress = true;
2090 }
2091 }
2092 }
2093 }
2095 return progress;
2096 }
2099 //=============================================================================
2100 //----------------------------build_and_optimize-------------------------------
2101 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2102 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2103 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
2104 ResourceMark rm;
2106 int old_progress = C->major_progress();
2107 uint orig_worklist_size = _igvn._worklist.size();
2109 // Reset major-progress flag for the driver's heuristics
2110 C->clear_major_progress();
2112 #ifndef PRODUCT
2113 // Capture for later assert
2114 uint unique = C->unique();
2115 _loop_invokes++;
2116 _loop_work += unique;
2117 #endif
2119 // True if the method has at least 1 irreducible loop
2120 _has_irreducible_loops = false;
2122 _created_loop_node = false;
2124 Arena *a = Thread::current()->resource_area();
2125 VectorSet visited(a);
2126 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2127 _nodes.map(C->unique(), NULL);
2128 memset(_nodes.adr(), 0, wordSize * C->unique());
2130 // Pre-build the top-level outermost loop tree entry
2131 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2132 // Do not need a safepoint at the top level
2133 _ltree_root->_has_sfpt = 1;
2135 // Initialize Dominators.
2136 // Checked in clone_loop_predicate() during beautify_loops().
2137 _idom_size = 0;
2138 _idom = NULL;
2139 _dom_depth = NULL;
2140 _dom_stk = NULL;
2142 // Empty pre-order array
2143 allocate_preorders();
2145 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2146 // IdealLoopTree entries. Data nodes are NOT walked.
2147 build_loop_tree();
2148 // Check for bailout, and return
2149 if (C->failing()) {
2150 return;
2151 }
2153 // No loops after all
2154 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2156 // There should always be an outer loop containing the Root and Return nodes.
2157 // If not, we have a degenerate empty program. Bail out in this case.
2158 if (!has_node(C->root())) {
2159 if (!_verify_only) {
2160 C->clear_major_progress();
2161 C->record_method_not_compilable("empty program detected during loop optimization");
2162 }
2163 return;
2164 }
2166 // Nothing to do, so get out
2167 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
2168 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2169 if (stop_early && !do_expensive_nodes) {
2170 _igvn.optimize(); // Cleanup NeverBranches
2171 return;
2172 }
2174 // Set loop nesting depth
2175 _ltree_root->set_nest( 0 );
2177 // Split shared headers and insert loop landing pads.
2178 // Do not bother doing this on the Root loop of course.
2179 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2180 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2181 if( _ltree_root->_child->beautify_loops( this ) ) {
2182 // Re-build loop tree!
2183 _ltree_root->_child = NULL;
2184 _nodes.clear();
2185 reallocate_preorders();
2186 build_loop_tree();
2187 // Check for bailout, and return
2188 if (C->failing()) {
2189 return;
2190 }
2191 // Reset loop nesting depth
2192 _ltree_root->set_nest( 0 );
2194 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2195 }
2196 }
2198 // Build Dominators for elision of NULL checks & loop finding.
2199 // Since nodes do not have a slot for immediate dominator, make
2200 // a persistent side array for that info indexed on node->_idx.
2201 _idom_size = C->unique();
2202 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2203 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2204 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2205 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2207 Dominators();
2209 if (!_verify_only) {
2210 // As a side effect, Dominators removed any unreachable CFG paths
2211 // into RegionNodes. It doesn't do this test against Root, so
2212 // we do it here.
2213 for( uint i = 1; i < C->root()->req(); i++ ) {
2214 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2215 _igvn.delete_input_of(C->root(), i);
2216 i--; // Rerun same iteration on compressed edges
2217 }
2218 }
2220 // Given dominators, try to find inner loops with calls that must
2221 // always be executed (call dominates loop tail). These loops do
2222 // not need a separate safepoint.
2223 Node_List cisstack(a);
2224 _ltree_root->check_safepts(visited, cisstack);
2225 }
2227 // Walk the DATA nodes and place into loops. Find earliest control
2228 // node. For CFG nodes, the _nodes array starts out and remains
2229 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2230 // _nodes array holds the earliest legal controlling CFG node.
2232 // Allocate stack with enough space to avoid frequent realloc
2233 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2234 Node_Stack nstack( a, stack_size );
2236 visited.Clear();
2237 Node_List worklist(a);
2238 // Don't need C->root() on worklist since
2239 // it will be processed among C->top() inputs
2240 worklist.push( C->top() );
2241 visited.set( C->top()->_idx ); // Set C->top() as visited now
2242 build_loop_early( visited, worklist, nstack );
2244 // Given early legal placement, try finding counted loops. This placement
2245 // is good enough to discover most loop invariants.
2246 if( !_verify_me && !_verify_only )
2247 _ltree_root->counted_loop( this );
2249 // Find latest loop placement. Find ideal loop placement.
2250 visited.Clear();
2251 init_dom_lca_tags();
2252 // Need C->root() on worklist when processing outs
2253 worklist.push( C->root() );
2254 NOT_PRODUCT( C->verify_graph_edges(); )
2255 worklist.push( C->top() );
2256 build_loop_late( visited, worklist, nstack );
2258 if (_verify_only) {
2259 // restore major progress flag
2260 for (int i = 0; i < old_progress; i++)
2261 C->set_major_progress();
2262 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2263 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2264 return;
2265 }
2267 // clear out the dead code after build_loop_late
2268 while (_deadlist.size()) {
2269 _igvn.remove_globally_dead_node(_deadlist.pop());
2270 }
2272 if (stop_early) {
2273 assert(do_expensive_nodes, "why are we here?");
2274 if (process_expensive_nodes()) {
2275 // If we made some progress when processing expensive nodes then
2276 // the IGVN may modify the graph in a way that will allow us to
2277 // make some more progress: we need to try processing expensive
2278 // nodes again.
2279 C->set_major_progress();
2280 }
2281 _igvn.optimize();
2282 return;
2283 }
2285 // Some parser-inserted loop predicates could never be used by loop
2286 // predication or they were moved away from loop during some optimizations.
2287 // For example, peeling. Eliminate them before next loop optimizations.
2288 if (UseLoopPredicate || LoopLimitCheck) {
2289 eliminate_useless_predicates();
2290 }
2292 #ifndef PRODUCT
2293 C->verify_graph_edges();
2294 if (_verify_me) { // Nested verify pass?
2295 // Check to see if the verify mode is broken
2296 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2297 return;
2298 }
2299 if(VerifyLoopOptimizations) verify();
2300 if(TraceLoopOpts && C->has_loops()) {
2301 _ltree_root->dump();
2302 }
2303 #endif
2305 if (skip_loop_opts) {
2306 // Cleanup any modified bits
2307 _igvn.optimize();
2309 if (C->log() != NULL) {
2310 log_loop_tree(_ltree_root, _ltree_root, C->log());
2311 }
2312 return;
2313 }
2315 if (ReassociateInvariants) {
2316 // Reassociate invariants and prep for split_thru_phi
2317 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2318 IdealLoopTree* lpt = iter.current();
2319 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2321 lpt->reassociate_invariants(this);
2323 // Because RCE opportunities can be masked by split_thru_phi,
2324 // look for RCE candidates and inhibit split_thru_phi
2325 // on just their loop-phi's for this pass of loop opts
2326 if (SplitIfBlocks && do_split_ifs) {
2327 if (lpt->policy_range_check(this)) {
2328 lpt->_rce_candidate = 1; // = true
2329 }
2330 }
2331 }
2332 }
2334 // Check for aggressive application of split-if and other transforms
2335 // that require basic-block info (like cloning through Phi's)
2336 if( SplitIfBlocks && do_split_ifs ) {
2337 visited.Clear();
2338 split_if_with_blocks( visited, nstack );
2339 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2340 }
2342 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2343 C->set_major_progress();
2344 }
2346 // Perform loop predication before iteration splitting
2347 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2348 _ltree_root->_child->loop_predication(this);
2349 }
2351 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2352 if (do_intrinsify_fill()) {
2353 C->set_major_progress();
2354 }
2355 }
2357 // Perform iteration-splitting on inner loops. Split iterations to avoid
2358 // range checks or one-shot null checks.
2360 // If split-if's didn't hack the graph too bad (no CFG changes)
2361 // then do loop opts.
2362 if (C->has_loops() && !C->major_progress()) {
2363 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2364 _ltree_root->_child->iteration_split( this, worklist );
2365 // No verify after peeling! GCM has hoisted code out of the loop.
2366 // After peeling, the hoisted code could sink inside the peeled area.
2367 // The peeling code does not try to recompute the best location for
2368 // all the code before the peeled area, so the verify pass will always
2369 // complain about it.
2370 }
2371 // Do verify graph edges in any case
2372 NOT_PRODUCT( C->verify_graph_edges(); );
2374 if (!do_split_ifs) {
2375 // We saw major progress in Split-If to get here. We forced a
2376 // pass with unrolling and not split-if, however more split-if's
2377 // might make progress. If the unrolling didn't make progress
2378 // then the major-progress flag got cleared and we won't try
2379 // another round of Split-If. In particular the ever-common
2380 // instance-of/check-cast pattern requires at least 2 rounds of
2381 // Split-If to clear out.
2382 C->set_major_progress();
2383 }
2385 // Repeat loop optimizations if new loops were seen
2386 if (created_loop_node()) {
2387 C->set_major_progress();
2388 }
2390 // Keep loop predicates and perform optimizations with them
2391 // until no more loop optimizations could be done.
2392 // After that switch predicates off and do more loop optimizations.
2393 if (!C->major_progress() && (C->predicate_count() > 0)) {
2394 C->cleanup_loop_predicates(_igvn);
2395 #ifndef PRODUCT
2396 if (TraceLoopOpts) {
2397 tty->print_cr("PredicatesOff");
2398 }
2399 #endif
2400 C->set_major_progress();
2401 }
2403 // Convert scalar to superword operations at the end of all loop opts.
2404 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2405 // SuperWord transform
2406 SuperWord sw(this);
2407 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2408 IdealLoopTree* lpt = iter.current();
2409 if (lpt->is_counted()) {
2410 sw.transform_loop(lpt);
2411 }
2412 }
2413 }
2415 // Cleanup any modified bits
2416 _igvn.optimize();
2418 // disable assert until issue with split_flow_path is resolved (6742111)
2419 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2420 // "shouldn't introduce irreducible loops");
2422 if (C->log() != NULL) {
2423 log_loop_tree(_ltree_root, _ltree_root, C->log());
2424 }
2425 }
2427 #ifndef PRODUCT
2428 //------------------------------print_statistics-------------------------------
2429 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2430 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2431 void PhaseIdealLoop::print_statistics() {
2432 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2433 }
2435 //------------------------------verify-----------------------------------------
2436 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2437 static int fail; // debug only, so its multi-thread dont care
2438 void PhaseIdealLoop::verify() const {
2439 int old_progress = C->major_progress();
2440 ResourceMark rm;
2441 PhaseIdealLoop loop_verify( _igvn, this );
2442 VectorSet visited(Thread::current()->resource_area());
2444 fail = 0;
2445 verify_compare( C->root(), &loop_verify, visited );
2446 assert( fail == 0, "verify loops failed" );
2447 // Verify loop structure is the same
2448 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2449 // Reset major-progress. It was cleared by creating a verify version of
2450 // PhaseIdealLoop.
2451 for( int i=0; i<old_progress; i++ )
2452 C->set_major_progress();
2453 }
2455 //------------------------------verify_compare---------------------------------
2456 // Make sure me and the given PhaseIdealLoop agree on key data structures
2457 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2458 if( !n ) return;
2459 if( visited.test_set( n->_idx ) ) return;
2460 if( !_nodes[n->_idx] ) { // Unreachable
2461 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2462 return;
2463 }
2465 uint i;
2466 for( i = 0; i < n->req(); i++ )
2467 verify_compare( n->in(i), loop_verify, visited );
2469 // Check the '_nodes' block/loop structure
2470 i = n->_idx;
2471 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2472 if( _nodes[i] != loop_verify->_nodes[i] &&
2473 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2474 tty->print("Mismatched control setting for: ");
2475 n->dump();
2476 if( fail++ > 10 ) return;
2477 Node *c = get_ctrl_no_update(n);
2478 tty->print("We have it as: ");
2479 if( c->in(0) ) c->dump();
2480 else tty->print_cr("N%d",c->_idx);
2481 tty->print("Verify thinks: ");
2482 if( loop_verify->has_ctrl(n) )
2483 loop_verify->get_ctrl_no_update(n)->dump();
2484 else
2485 loop_verify->get_loop_idx(n)->dump();
2486 tty->cr();
2487 }
2488 } else { // We have a loop
2489 IdealLoopTree *us = get_loop_idx(n);
2490 if( loop_verify->has_ctrl(n) ) {
2491 tty->print("Mismatched loop setting for: ");
2492 n->dump();
2493 if( fail++ > 10 ) return;
2494 tty->print("We have it as: ");
2495 us->dump();
2496 tty->print("Verify thinks: ");
2497 loop_verify->get_ctrl_no_update(n)->dump();
2498 tty->cr();
2499 } else if (!C->major_progress()) {
2500 // Loop selection can be messed up if we did a major progress
2501 // operation, like split-if. Do not verify in that case.
2502 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2503 if( us->_head != them->_head || us->_tail != them->_tail ) {
2504 tty->print("Unequals loops for: ");
2505 n->dump();
2506 if( fail++ > 10 ) return;
2507 tty->print("We have it as: ");
2508 us->dump();
2509 tty->print("Verify thinks: ");
2510 them->dump();
2511 tty->cr();
2512 }
2513 }
2514 }
2516 // Check for immediate dominators being equal
2517 if( i >= _idom_size ) {
2518 if( !n->is_CFG() ) return;
2519 tty->print("CFG Node with no idom: ");
2520 n->dump();
2521 return;
2522 }
2523 if( !n->is_CFG() ) return;
2524 if( n == C->root() ) return; // No IDOM here
2526 assert(n->_idx == i, "sanity");
2527 Node *id = idom_no_update(n);
2528 if( id != loop_verify->idom_no_update(n) ) {
2529 tty->print("Unequals idoms for: ");
2530 n->dump();
2531 if( fail++ > 10 ) return;
2532 tty->print("We have it as: ");
2533 id->dump();
2534 tty->print("Verify thinks: ");
2535 loop_verify->idom_no_update(n)->dump();
2536 tty->cr();
2537 }
2539 }
2541 //------------------------------verify_tree------------------------------------
2542 // Verify that tree structures match. Because the CFG can change, siblings
2543 // within the loop tree can be reordered. We attempt to deal with that by
2544 // reordering the verify's loop tree if possible.
2545 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2546 assert( _parent == parent, "Badly formed loop tree" );
2548 // Siblings not in same order? Attempt to re-order.
2549 if( _head != loop->_head ) {
2550 // Find _next pointer to update
2551 IdealLoopTree **pp = &loop->_parent->_child;
2552 while( *pp != loop )
2553 pp = &((*pp)->_next);
2554 // Find proper sibling to be next
2555 IdealLoopTree **nn = &loop->_next;
2556 while( (*nn) && (*nn)->_head != _head )
2557 nn = &((*nn)->_next);
2559 // Check for no match.
2560 if( !(*nn) ) {
2561 // Annoyingly, irreducible loops can pick different headers
2562 // after a major_progress operation, so the rest of the loop
2563 // tree cannot be matched.
2564 if (_irreducible && Compile::current()->major_progress()) return;
2565 assert( 0, "failed to match loop tree" );
2566 }
2568 // Move (*nn) to (*pp)
2569 IdealLoopTree *hit = *nn;
2570 *nn = hit->_next;
2571 hit->_next = loop;
2572 *pp = loop;
2573 loop = hit;
2574 // Now try again to verify
2575 }
2577 assert( _head == loop->_head , "mismatched loop head" );
2578 Node *tail = _tail; // Inline a non-updating version of
2579 while( !tail->in(0) ) // the 'tail()' call.
2580 tail = tail->in(1);
2581 assert( tail == loop->_tail, "mismatched loop tail" );
2583 // Counted loops that are guarded should be able to find their guards
2584 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2585 CountedLoopNode *cl = _head->as_CountedLoop();
2586 Node *init = cl->init_trip();
2587 Node *ctrl = cl->in(LoopNode::EntryControl);
2588 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2589 Node *iff = ctrl->in(0);
2590 assert( iff->Opcode() == Op_If, "" );
2591 Node *bol = iff->in(1);
2592 assert( bol->Opcode() == Op_Bool, "" );
2593 Node *cmp = bol->in(1);
2594 assert( cmp->Opcode() == Op_CmpI, "" );
2595 Node *add = cmp->in(1);
2596 Node *opaq;
2597 if( add->Opcode() == Op_Opaque1 ) {
2598 opaq = add;
2599 } else {
2600 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2601 assert( add == init, "" );
2602 opaq = cmp->in(2);
2603 }
2604 assert( opaq->Opcode() == Op_Opaque1, "" );
2606 }
2608 if (_child != NULL) _child->verify_tree(loop->_child, this);
2609 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2610 // Innermost loops need to verify loop bodies,
2611 // but only if no 'major_progress'
2612 int fail = 0;
2613 if (!Compile::current()->major_progress() && _child == NULL) {
2614 for( uint i = 0; i < _body.size(); i++ ) {
2615 Node *n = _body.at(i);
2616 if (n->outcnt() == 0) continue; // Ignore dead
2617 uint j;
2618 for( j = 0; j < loop->_body.size(); j++ )
2619 if( loop->_body.at(j) == n )
2620 break;
2621 if( j == loop->_body.size() ) { // Not found in loop body
2622 // Last ditch effort to avoid assertion: Its possible that we
2623 // have some users (so outcnt not zero) but are still dead.
2624 // Try to find from root.
2625 if (Compile::current()->root()->find(n->_idx)) {
2626 fail++;
2627 tty->print("We have that verify does not: ");
2628 n->dump();
2629 }
2630 }
2631 }
2632 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2633 Node *n = loop->_body.at(i2);
2634 if (n->outcnt() == 0) continue; // Ignore dead
2635 uint j;
2636 for( j = 0; j < _body.size(); j++ )
2637 if( _body.at(j) == n )
2638 break;
2639 if( j == _body.size() ) { // Not found in loop body
2640 // Last ditch effort to avoid assertion: Its possible that we
2641 // have some users (so outcnt not zero) but are still dead.
2642 // Try to find from root.
2643 if (Compile::current()->root()->find(n->_idx)) {
2644 fail++;
2645 tty->print("Verify has that we do not: ");
2646 n->dump();
2647 }
2648 }
2649 }
2650 assert( !fail, "loop body mismatch" );
2651 }
2652 }
2654 #endif
2656 //------------------------------set_idom---------------------------------------
2657 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2658 uint idx = d->_idx;
2659 if (idx >= _idom_size) {
2660 uint newsize = _idom_size<<1;
2661 while( idx >= newsize ) {
2662 newsize <<= 1;
2663 }
2664 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2665 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2666 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2667 _idom_size = newsize;
2668 }
2669 _idom[idx] = n;
2670 _dom_depth[idx] = dom_depth;
2671 }
2673 //------------------------------recompute_dom_depth---------------------------------------
2674 // The dominator tree is constructed with only parent pointers.
2675 // This recomputes the depth in the tree by first tagging all
2676 // nodes as "no depth yet" marker. The next pass then runs up
2677 // the dom tree from each node marked "no depth yet", and computes
2678 // the depth on the way back down.
2679 void PhaseIdealLoop::recompute_dom_depth() {
2680 uint no_depth_marker = C->unique();
2681 uint i;
2682 // Initialize depth to "no depth yet"
2683 for (i = 0; i < _idom_size; i++) {
2684 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2685 _dom_depth[i] = no_depth_marker;
2686 }
2687 }
2688 if (_dom_stk == NULL) {
2689 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
2690 if (init_size < 10) init_size = 10;
2691 _dom_stk = new GrowableArray<uint>(init_size);
2692 }
2693 // Compute new depth for each node.
2694 for (i = 0; i < _idom_size; i++) {
2695 uint j = i;
2696 // Run up the dom tree to find a node with a depth
2697 while (_dom_depth[j] == no_depth_marker) {
2698 _dom_stk->push(j);
2699 j = _idom[j]->_idx;
2700 }
2701 // Compute the depth on the way back down this tree branch
2702 uint dd = _dom_depth[j] + 1;
2703 while (_dom_stk->length() > 0) {
2704 uint j = _dom_stk->pop();
2705 _dom_depth[j] = dd;
2706 dd++;
2707 }
2708 }
2709 }
2711 //------------------------------sort-------------------------------------------
2712 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2713 // loop tree, not the root.
2714 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2715 if( !innermost ) return loop; // New innermost loop
2717 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2718 assert( loop_preorder, "not yet post-walked loop" );
2719 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2720 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2722 // Insert at start of list
2723 while( l ) { // Insertion sort based on pre-order
2724 if( l == loop ) return innermost; // Already on list!
2725 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2726 assert( l_preorder, "not yet post-walked l" );
2727 // Check header pre-order number to figure proper nesting
2728 if( loop_preorder > l_preorder )
2729 break; // End of insertion
2730 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2731 // Since I split shared headers, you'd think this could not happen.
2732 // BUT: I must first do the preorder numbering before I can discover I
2733 // have shared headers, so the split headers all get the same preorder
2734 // number as the RegionNode they split from.
2735 if( loop_preorder == l_preorder &&
2736 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2737 break; // Also check for shared headers (same pre#)
2738 pp = &l->_parent; // Chain up list
2739 l = *pp;
2740 }
2741 // Link into list
2742 // Point predecessor to me
2743 *pp = loop;
2744 // Point me to successor
2745 IdealLoopTree *p = loop->_parent;
2746 loop->_parent = l; // Point me to successor
2747 if( p ) sort( p, innermost ); // Insert my parents into list as well
2748 return innermost;
2749 }
2751 //------------------------------build_loop_tree--------------------------------
2752 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2753 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2754 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2755 // tightest enclosing IdealLoopTree for post-walked.
2756 //
2757 // During my forward walk I do a short 1-layer lookahead to see if I can find
2758 // a loop backedge with that doesn't have any work on the backedge. This
2759 // helps me construct nested loops with shared headers better.
2760 //
2761 // Once I've done the forward recursion, I do the post-work. For each child
2762 // I check to see if there is a backedge. Backedges define a loop! I
2763 // insert an IdealLoopTree at the target of the backedge.
2764 //
2765 // During the post-work I also check to see if I have several children
2766 // belonging to different loops. If so, then this Node is a decision point
2767 // where control flow can choose to change loop nests. It is at this
2768 // decision point where I can figure out how loops are nested. At this
2769 // time I can properly order the different loop nests from my children.
2770 // Note that there may not be any backedges at the decision point!
2771 //
2772 // Since the decision point can be far removed from the backedges, I can't
2773 // order my loops at the time I discover them. Thus at the decision point
2774 // I need to inspect loop header pre-order numbers to properly nest my
2775 // loops. This means I need to sort my childrens' loops by pre-order.
2776 // The sort is of size number-of-control-children, which generally limits
2777 // it to size 2 (i.e., I just choose between my 2 target loops).
2778 void PhaseIdealLoop::build_loop_tree() {
2779 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
2780 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
2781 Node *n = C->root();
2782 bltstack.push(n);
2783 int pre_order = 1;
2784 int stack_size;
2786 while ( ( stack_size = bltstack.length() ) != 0 ) {
2787 n = bltstack.top(); // Leave node on stack
2788 if ( !is_visited(n) ) {
2789 // ---- Pre-pass Work ----
2790 // Pre-walked but not post-walked nodes need a pre_order number.
2792 set_preorder_visited( n, pre_order ); // set as visited
2794 // ---- Scan over children ----
2795 // Scan first over control projections that lead to loop headers.
2796 // This helps us find inner-to-outer loops with shared headers better.
2798 // Scan children's children for loop headers.
2799 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2800 Node* m = n->raw_out(i); // Child
2801 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2802 // Scan over children's children to find loop
2803 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2804 Node* l = m->fast_out(j);
2805 if( is_visited(l) && // Been visited?
2806 !is_postvisited(l) && // But not post-visited
2807 get_preorder(l) < pre_order ) { // And smaller pre-order
2808 // Found! Scan the DFS down this path before doing other paths
2809 bltstack.push(m);
2810 break;
2811 }
2812 }
2813 }
2814 }
2815 pre_order++;
2816 }
2817 else if ( !is_postvisited(n) ) {
2818 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2819 // such as com.sun.rsasign.am::a.
2820 // For non-recursive version, first, process current children.
2821 // On next iteration, check if additional children were added.
2822 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2823 Node* u = n->raw_out(k);
2824 if ( u->is_CFG() && !is_visited(u) ) {
2825 bltstack.push(u);
2826 }
2827 }
2828 if ( bltstack.length() == stack_size ) {
2829 // There were no additional children, post visit node now
2830 (void)bltstack.pop(); // Remove node from stack
2831 pre_order = build_loop_tree_impl( n, pre_order );
2832 // Check for bailout
2833 if (C->failing()) {
2834 return;
2835 }
2836 // Check to grow _preorders[] array for the case when
2837 // build_loop_tree_impl() adds new nodes.
2838 check_grow_preorders();
2839 }
2840 }
2841 else {
2842 (void)bltstack.pop(); // Remove post-visited node from stack
2843 }
2844 }
2845 }
2847 //------------------------------build_loop_tree_impl---------------------------
2848 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2849 // ---- Post-pass Work ----
2850 // Pre-walked but not post-walked nodes need a pre_order number.
2852 // Tightest enclosing loop for this Node
2853 IdealLoopTree *innermost = NULL;
2855 // For all children, see if any edge is a backedge. If so, make a loop
2856 // for it. Then find the tightest enclosing loop for the self Node.
2857 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2858 Node* m = n->fast_out(i); // Child
2859 if( n == m ) continue; // Ignore control self-cycles
2860 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2862 IdealLoopTree *l; // Child's loop
2863 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2864 // Found a backedge
2865 assert( get_preorder(m) < pre_order, "should be backedge" );
2866 // Check for the RootNode, which is already a LoopNode and is allowed
2867 // to have multiple "backedges".
2868 if( m == C->root()) { // Found the root?
2869 l = _ltree_root; // Root is the outermost LoopNode
2870 } else { // Else found a nested loop
2871 // Insert a LoopNode to mark this loop.
2872 l = new IdealLoopTree(this, m, n);
2873 } // End of Else found a nested loop
2874 if( !has_loop(m) ) // If 'm' does not already have a loop set
2875 set_loop(m, l); // Set loop header to loop now
2877 } else { // Else not a nested loop
2878 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2879 l = get_loop(m); // Get previously determined loop
2880 // If successor is header of a loop (nest), move up-loop till it
2881 // is a member of some outer enclosing loop. Since there are no
2882 // shared headers (I've split them already) I only need to go up
2883 // at most 1 level.
2884 while( l && l->_head == m ) // Successor heads loop?
2885 l = l->_parent; // Move up 1 for me
2886 // If this loop is not properly parented, then this loop
2887 // has no exit path out, i.e. its an infinite loop.
2888 if( !l ) {
2889 // Make loop "reachable" from root so the CFG is reachable. Basically
2890 // insert a bogus loop exit that is never taken. 'm', the loop head,
2891 // points to 'n', one (of possibly many) fall-in paths. There may be
2892 // many backedges as well.
2894 // Here I set the loop to be the root loop. I could have, after
2895 // inserting a bogus loop exit, restarted the recursion and found my
2896 // new loop exit. This would make the infinite loop a first-class
2897 // loop and it would then get properly optimized. What's the use of
2898 // optimizing an infinite loop?
2899 l = _ltree_root; // Oops, found infinite loop
2901 if (!_verify_only) {
2902 // Insert the NeverBranch between 'm' and it's control user.
2903 NeverBranchNode *iff = new (C) NeverBranchNode( m );
2904 _igvn.register_new_node_with_optimizer(iff);
2905 set_loop(iff, l);
2906 Node *if_t = new (C) CProjNode( iff, 0 );
2907 _igvn.register_new_node_with_optimizer(if_t);
2908 set_loop(if_t, l);
2910 Node* cfg = NULL; // Find the One True Control User of m
2911 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2912 Node* x = m->fast_out(j);
2913 if (x->is_CFG() && x != m && x != iff)
2914 { cfg = x; break; }
2915 }
2916 assert(cfg != NULL, "must find the control user of m");
2917 uint k = 0; // Probably cfg->in(0)
2918 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2919 cfg->set_req( k, if_t ); // Now point to NeverBranch
2921 // Now create the never-taken loop exit
2922 Node *if_f = new (C) CProjNode( iff, 1 );
2923 _igvn.register_new_node_with_optimizer(if_f);
2924 set_loop(if_f, l);
2925 // Find frame ptr for Halt. Relies on the optimizer
2926 // V-N'ing. Easier and quicker than searching through
2927 // the program structure.
2928 Node *frame = new (C) ParmNode( C->start(), TypeFunc::FramePtr );
2929 _igvn.register_new_node_with_optimizer(frame);
2930 // Halt & Catch Fire
2931 Node *halt = new (C) HaltNode( if_f, frame );
2932 _igvn.register_new_node_with_optimizer(halt);
2933 set_loop(halt, l);
2934 C->root()->add_req(halt);
2935 }
2936 set_loop(C->root(), _ltree_root);
2937 }
2938 }
2939 // Weeny check for irreducible. This child was already visited (this
2940 // IS the post-work phase). Is this child's loop header post-visited
2941 // as well? If so, then I found another entry into the loop.
2942 if (!_verify_only) {
2943 while( is_postvisited(l->_head) ) {
2944 // found irreducible
2945 l->_irreducible = 1; // = true
2946 l = l->_parent;
2947 _has_irreducible_loops = true;
2948 // Check for bad CFG here to prevent crash, and bailout of compile
2949 if (l == NULL) {
2950 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2951 return pre_order;
2952 }
2953 }
2954 C->set_has_irreducible_loop(_has_irreducible_loops);
2955 }
2957 // This Node might be a decision point for loops. It is only if
2958 // it's children belong to several different loops. The sort call
2959 // does a trivial amount of work if there is only 1 child or all
2960 // children belong to the same loop. If however, the children
2961 // belong to different loops, the sort call will properly set the
2962 // _parent pointers to show how the loops nest.
2963 //
2964 // In any case, it returns the tightest enclosing loop.
2965 innermost = sort( l, innermost );
2966 }
2968 // Def-use info will have some dead stuff; dead stuff will have no
2969 // loop decided on.
2971 // Am I a loop header? If so fix up my parent's child and next ptrs.
2972 if( innermost && innermost->_head == n ) {
2973 assert( get_loop(n) == innermost, "" );
2974 IdealLoopTree *p = innermost->_parent;
2975 IdealLoopTree *l = innermost;
2976 while( p && l->_head == n ) {
2977 l->_next = p->_child; // Put self on parents 'next child'
2978 p->_child = l; // Make self as first child of parent
2979 l = p; // Now walk up the parent chain
2980 p = l->_parent;
2981 }
2982 } else {
2983 // Note that it is possible for a LoopNode to reach here, if the
2984 // backedge has been made unreachable (hence the LoopNode no longer
2985 // denotes a Loop, and will eventually be removed).
2987 // Record tightest enclosing loop for self. Mark as post-visited.
2988 set_loop(n, innermost);
2989 // Also record has_call flag early on
2990 if( innermost ) {
2991 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2992 // Do not count uncommon calls
2993 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2994 Node *iff = n->in(0)->in(0);
2995 // No any calls for vectorized loops.
2996 if( UseSuperWord || !iff->is_If() ||
2997 (n->in(0)->Opcode() == Op_IfFalse &&
2998 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2999 (iff->as_If()->_prob >= 0.01) )
3000 innermost->_has_call = 1;
3001 }
3002 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3003 // Disable loop optimizations if the loop has a scalar replaceable
3004 // allocation. This disabling may cause a potential performance lost
3005 // if the allocation is not eliminated for some reason.
3006 innermost->_allow_optimizations = false;
3007 innermost->_has_call = 1; // = true
3008 } else if (n->Opcode() == Op_SafePoint) {
3009 // Record all safepoints in this loop.
3010 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3011 innermost->_safepts->push(n);
3012 }
3013 }
3014 }
3016 // Flag as post-visited now
3017 set_postvisited(n);
3018 return pre_order;
3019 }
3022 //------------------------------build_loop_early-------------------------------
3023 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3024 // First pass computes the earliest controlling node possible. This is the
3025 // controlling input with the deepest dominating depth.
3026 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3027 while (worklist.size() != 0) {
3028 // Use local variables nstack_top_n & nstack_top_i to cache values
3029 // on nstack's top.
3030 Node *nstack_top_n = worklist.pop();
3031 uint nstack_top_i = 0;
3032 //while_nstack_nonempty:
3033 while (true) {
3034 // Get parent node and next input's index from stack's top.
3035 Node *n = nstack_top_n;
3036 uint i = nstack_top_i;
3037 uint cnt = n->req(); // Count of inputs
3038 if (i == 0) { // Pre-process the node.
3039 if( has_node(n) && // Have either loop or control already?
3040 !has_ctrl(n) ) { // Have loop picked out already?
3041 // During "merge_many_backedges" we fold up several nested loops
3042 // into a single loop. This makes the members of the original
3043 // loop bodies pointing to dead loops; they need to move up
3044 // to the new UNION'd larger loop. I set the _head field of these
3045 // dead loops to NULL and the _parent field points to the owning
3046 // loop. Shades of UNION-FIND algorithm.
3047 IdealLoopTree *ilt;
3048 while( !(ilt = get_loop(n))->_head ) {
3049 // Normally I would use a set_loop here. But in this one special
3050 // case, it is legal (and expected) to change what loop a Node
3051 // belongs to.
3052 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3053 }
3054 // Remove safepoints ONLY if I've already seen I don't need one.
3055 // (the old code here would yank a 2nd safepoint after seeing a
3056 // first one, even though the 1st did not dominate in the loop body
3057 // and thus could be avoided indefinitely)
3058 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3059 is_deleteable_safept(n)) {
3060 Node *in = n->in(TypeFunc::Control);
3061 lazy_replace(n,in); // Pull safepoint now
3062 if (ilt->_safepts != NULL) {
3063 ilt->_safepts->yank(n);
3064 }
3065 // Carry on with the recursion "as if" we are walking
3066 // only the control input
3067 if( !visited.test_set( in->_idx ) ) {
3068 worklist.push(in); // Visit this guy later, using worklist
3069 }
3070 // Get next node from nstack:
3071 // - skip n's inputs processing by setting i > cnt;
3072 // - we also will not call set_early_ctrl(n) since
3073 // has_node(n) == true (see the condition above).
3074 i = cnt + 1;
3075 }
3076 }
3077 } // if (i == 0)
3079 // Visit all inputs
3080 bool done = true; // Assume all n's inputs will be processed
3081 while (i < cnt) {
3082 Node *in = n->in(i);
3083 ++i;
3084 if (in == NULL) continue;
3085 if (in->pinned() && !in->is_CFG())
3086 set_ctrl(in, in->in(0));
3087 int is_visited = visited.test_set( in->_idx );
3088 if (!has_node(in)) { // No controlling input yet?
3089 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3090 assert( !is_visited, "visit only once" );
3091 nstack.push(n, i); // Save parent node and next input's index.
3092 nstack_top_n = in; // Process current input now.
3093 nstack_top_i = 0;
3094 done = false; // Not all n's inputs processed.
3095 break; // continue while_nstack_nonempty;
3096 } else if (!is_visited) {
3097 // This guy has a location picked out for him, but has not yet
3098 // been visited. Happens to all CFG nodes, for instance.
3099 // Visit him using the worklist instead of recursion, to break
3100 // cycles. Since he has a location already we do not need to
3101 // find his location before proceeding with the current Node.
3102 worklist.push(in); // Visit this guy later, using worklist
3103 }
3104 }
3105 if (done) {
3106 // All of n's inputs have been processed, complete post-processing.
3108 // Compute earliest point this Node can go.
3109 // CFG, Phi, pinned nodes already know their controlling input.
3110 if (!has_node(n)) {
3111 // Record earliest legal location
3112 set_early_ctrl( n );
3113 }
3114 if (nstack.is_empty()) {
3115 // Finished all nodes on stack.
3116 // Process next node on the worklist.
3117 break;
3118 }
3119 // Get saved parent node and next input's index.
3120 nstack_top_n = nstack.node();
3121 nstack_top_i = nstack.index();
3122 nstack.pop();
3123 }
3124 } // while (true)
3125 }
3126 }
3128 //------------------------------dom_lca_internal--------------------------------
3129 // Pair-wise LCA
3130 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3131 if( !n1 ) return n2; // Handle NULL original LCA
3132 assert( n1->is_CFG(), "" );
3133 assert( n2->is_CFG(), "" );
3134 // find LCA of all uses
3135 uint d1 = dom_depth(n1);
3136 uint d2 = dom_depth(n2);
3137 while (n1 != n2) {
3138 if (d1 > d2) {
3139 n1 = idom(n1);
3140 d1 = dom_depth(n1);
3141 } else if (d1 < d2) {
3142 n2 = idom(n2);
3143 d2 = dom_depth(n2);
3144 } else {
3145 // Here d1 == d2. Due to edits of the dominator-tree, sections
3146 // of the tree might have the same depth. These sections have
3147 // to be searched more carefully.
3149 // Scan up all the n1's with equal depth, looking for n2.
3150 Node *t1 = idom(n1);
3151 while (dom_depth(t1) == d1) {
3152 if (t1 == n2) return n2;
3153 t1 = idom(t1);
3154 }
3155 // Scan up all the n2's with equal depth, looking for n1.
3156 Node *t2 = idom(n2);
3157 while (dom_depth(t2) == d2) {
3158 if (t2 == n1) return n1;
3159 t2 = idom(t2);
3160 }
3161 // Move up to a new dominator-depth value as well as up the dom-tree.
3162 n1 = t1;
3163 n2 = t2;
3164 d1 = dom_depth(n1);
3165 d2 = dom_depth(n2);
3166 }
3167 }
3168 return n1;
3169 }
3171 //------------------------------compute_idom-----------------------------------
3172 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3173 // IDOMs are correct.
3174 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3175 assert( region->is_Region(), "" );
3176 Node *LCA = NULL;
3177 for( uint i = 1; i < region->req(); i++ ) {
3178 if( region->in(i) != C->top() )
3179 LCA = dom_lca( LCA, region->in(i) );
3180 }
3181 return LCA;
3182 }
3184 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3185 bool had_error = false;
3186 #ifdef ASSERT
3187 if (early != C->root()) {
3188 // Make sure that there's a dominance path from LCA to early
3189 Node* d = LCA;
3190 while (d != early) {
3191 if (d == C->root()) {
3192 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3193 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3194 had_error = true;
3195 break;
3196 }
3197 d = idom(d);
3198 }
3199 }
3200 #endif
3201 return had_error;
3202 }
3205 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3206 // Compute LCA over list of uses
3207 bool had_error = false;
3208 Node *LCA = NULL;
3209 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3210 Node* c = n->fast_out(i);
3211 if (_nodes[c->_idx] == NULL)
3212 continue; // Skip the occasional dead node
3213 if( c->is_Phi() ) { // For Phis, we must land above on the path
3214 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3215 if( c->in(j) == n ) { // Found matching input?
3216 Node *use = c->in(0)->in(j);
3217 if (_verify_only && use->is_top()) continue;
3218 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3219 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3220 }
3221 }
3222 } else {
3223 // For CFG data-users, use is in the block just prior
3224 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3225 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3226 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3227 }
3228 }
3229 assert(!had_error, "bad dominance");
3230 return LCA;
3231 }
3233 //------------------------------get_late_ctrl----------------------------------
3234 // Compute latest legal control.
3235 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3236 assert(early != NULL, "early control should not be NULL");
3238 Node* LCA = compute_lca_of_uses(n, early);
3239 #ifdef ASSERT
3240 if (LCA == C->root() && LCA != early) {
3241 // def doesn't dominate uses so print some useful debugging output
3242 compute_lca_of_uses(n, early, true);
3243 }
3244 #endif
3246 // if this is a load, check for anti-dependent stores
3247 // We use a conservative algorithm to identify potential interfering
3248 // instructions and for rescheduling the load. The users of the memory
3249 // input of this load are examined. Any use which is not a load and is
3250 // dominated by early is considered a potentially interfering store.
3251 // This can produce false positives.
3252 if (n->is_Load() && LCA != early) {
3253 Node_List worklist;
3255 Node *mem = n->in(MemNode::Memory);
3256 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3257 Node* s = mem->fast_out(i);
3258 worklist.push(s);
3259 }
3260 while(worklist.size() != 0 && LCA != early) {
3261 Node* s = worklist.pop();
3262 if (s->is_Load()) {
3263 continue;
3264 } else if (s->is_MergeMem()) {
3265 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3266 Node* s1 = s->fast_out(i);
3267 worklist.push(s1);
3268 }
3269 } else {
3270 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3271 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3272 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3273 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3274 }
3275 }
3276 }
3277 }
3279 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3280 return LCA;
3281 }
3283 // true if CFG node d dominates CFG node n
3284 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3285 if (d == n)
3286 return true;
3287 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3288 uint dd = dom_depth(d);
3289 while (dom_depth(n) >= dd) {
3290 if (n == d)
3291 return true;
3292 n = idom(n);
3293 }
3294 return false;
3295 }
3297 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3298 // Pair-wise LCA with tags.
3299 // Tag each index with the node 'tag' currently being processed
3300 // before advancing up the dominator chain using idom().
3301 // Later calls that find a match to 'tag' know that this path has already
3302 // been considered in the current LCA (which is input 'n1' by convention).
3303 // Since get_late_ctrl() is only called once for each node, the tag array
3304 // does not need to be cleared between calls to get_late_ctrl().
3305 // Algorithm trades a larger constant factor for better asymptotic behavior
3306 //
3307 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3308 uint d1 = dom_depth(n1);
3309 uint d2 = dom_depth(n2);
3311 do {
3312 if (d1 > d2) {
3313 // current lca is deeper than n2
3314 _dom_lca_tags.map(n1->_idx, tag);
3315 n1 = idom(n1);
3316 d1 = dom_depth(n1);
3317 } else if (d1 < d2) {
3318 // n2 is deeper than current lca
3319 Node *memo = _dom_lca_tags[n2->_idx];
3320 if( memo == tag ) {
3321 return n1; // Return the current LCA
3322 }
3323 _dom_lca_tags.map(n2->_idx, tag);
3324 n2 = idom(n2);
3325 d2 = dom_depth(n2);
3326 } else {
3327 // Here d1 == d2. Due to edits of the dominator-tree, sections
3328 // of the tree might have the same depth. These sections have
3329 // to be searched more carefully.
3331 // Scan up all the n1's with equal depth, looking for n2.
3332 _dom_lca_tags.map(n1->_idx, tag);
3333 Node *t1 = idom(n1);
3334 while (dom_depth(t1) == d1) {
3335 if (t1 == n2) return n2;
3336 _dom_lca_tags.map(t1->_idx, tag);
3337 t1 = idom(t1);
3338 }
3339 // Scan up all the n2's with equal depth, looking for n1.
3340 _dom_lca_tags.map(n2->_idx, tag);
3341 Node *t2 = idom(n2);
3342 while (dom_depth(t2) == d2) {
3343 if (t2 == n1) return n1;
3344 _dom_lca_tags.map(t2->_idx, tag);
3345 t2 = idom(t2);
3346 }
3347 // Move up to a new dominator-depth value as well as up the dom-tree.
3348 n1 = t1;
3349 n2 = t2;
3350 d1 = dom_depth(n1);
3351 d2 = dom_depth(n2);
3352 }
3353 } while (n1 != n2);
3354 return n1;
3355 }
3357 //------------------------------init_dom_lca_tags------------------------------
3358 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3359 // Intended use does not involve any growth for the array, so it could
3360 // be of fixed size.
3361 void PhaseIdealLoop::init_dom_lca_tags() {
3362 uint limit = C->unique() + 1;
3363 _dom_lca_tags.map( limit, NULL );
3364 #ifdef ASSERT
3365 for( uint i = 0; i < limit; ++i ) {
3366 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3367 }
3368 #endif // ASSERT
3369 }
3371 //------------------------------clear_dom_lca_tags------------------------------
3372 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3373 // Intended use does not involve any growth for the array, so it could
3374 // be of fixed size.
3375 void PhaseIdealLoop::clear_dom_lca_tags() {
3376 uint limit = C->unique() + 1;
3377 _dom_lca_tags.map( limit, NULL );
3378 _dom_lca_tags.clear();
3379 #ifdef ASSERT
3380 for( uint i = 0; i < limit; ++i ) {
3381 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3382 }
3383 #endif // ASSERT
3384 }
3386 //------------------------------build_loop_late--------------------------------
3387 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3388 // Second pass finds latest legal placement, and ideal loop placement.
3389 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3390 while (worklist.size() != 0) {
3391 Node *n = worklist.pop();
3392 // Only visit once
3393 if (visited.test_set(n->_idx)) continue;
3394 uint cnt = n->outcnt();
3395 uint i = 0;
3396 while (true) {
3397 assert( _nodes[n->_idx], "no dead nodes" );
3398 // Visit all children
3399 if (i < cnt) {
3400 Node* use = n->raw_out(i);
3401 ++i;
3402 // Check for dead uses. Aggressively prune such junk. It might be
3403 // dead in the global sense, but still have local uses so I cannot
3404 // easily call 'remove_dead_node'.
3405 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3406 // Due to cycles, we might not hit the same fixed point in the verify
3407 // pass as we do in the regular pass. Instead, visit such phis as
3408 // simple uses of the loop head.
3409 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3410 if( !visited.test(use->_idx) )
3411 worklist.push(use);
3412 } else if( !visited.test_set(use->_idx) ) {
3413 nstack.push(n, i); // Save parent and next use's index.
3414 n = use; // Process all children of current use.
3415 cnt = use->outcnt();
3416 i = 0;
3417 }
3418 } else {
3419 // Do not visit around the backedge of loops via data edges.
3420 // push dead code onto a worklist
3421 _deadlist.push(use);
3422 }
3423 } else {
3424 // All of n's children have been processed, complete post-processing.
3425 build_loop_late_post(n);
3426 if (nstack.is_empty()) {
3427 // Finished all nodes on stack.
3428 // Process next node on the worklist.
3429 break;
3430 }
3431 // Get saved parent node and next use's index. Visit the rest of uses.
3432 n = nstack.node();
3433 cnt = n->outcnt();
3434 i = nstack.index();
3435 nstack.pop();
3436 }
3437 }
3438 }
3439 }
3441 //------------------------------build_loop_late_post---------------------------
3442 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3443 // Second pass finds latest legal placement, and ideal loop placement.
3444 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3446 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3447 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3448 }
3450 #ifdef ASSERT
3451 if (_verify_only && !n->is_CFG()) {
3452 // Check def-use domination.
3453 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
3454 }
3455 #endif
3457 // CFG and pinned nodes already handled
3458 if( n->in(0) ) {
3459 if( n->in(0)->is_top() ) return; // Dead?
3461 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3462 // _must_ be pinned (they have to observe their control edge of course).
3463 // Unlike Stores (which modify an unallocable resource, the memory
3464 // state), Mods/Loads can float around. So free them up.
3465 bool pinned = true;
3466 switch( n->Opcode() ) {
3467 case Op_DivI:
3468 case Op_DivF:
3469 case Op_DivD:
3470 case Op_ModI:
3471 case Op_ModF:
3472 case Op_ModD:
3473 case Op_LoadB: // Same with Loads; they can sink
3474 case Op_LoadUB: // during loop optimizations.
3475 case Op_LoadUS:
3476 case Op_LoadD:
3477 case Op_LoadF:
3478 case Op_LoadI:
3479 case Op_LoadKlass:
3480 case Op_LoadNKlass:
3481 case Op_LoadL:
3482 case Op_LoadS:
3483 case Op_LoadP:
3484 case Op_LoadN:
3485 case Op_LoadRange:
3486 case Op_LoadD_unaligned:
3487 case Op_LoadL_unaligned:
3488 case Op_StrComp: // Does a bunch of load-like effects
3489 case Op_StrEquals:
3490 case Op_StrIndexOf:
3491 case Op_AryEq:
3492 pinned = false;
3493 }
3494 if( pinned ) {
3495 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3496 if( !chosen_loop->_child ) // Inner loop?
3497 chosen_loop->_body.push(n); // Collect inner loops
3498 return;
3499 }
3500 } else { // No slot zero
3501 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3502 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3503 return;
3504 }
3505 assert(!n->is_CFG() || n->outcnt() == 0, "");
3506 }
3508 // Do I have a "safe range" I can select over?
3509 Node *early = get_ctrl(n);// Early location already computed
3511 // Compute latest point this Node can go
3512 Node *LCA = get_late_ctrl( n, early );
3513 // LCA is NULL due to uses being dead
3514 if( LCA == NULL ) {
3515 #ifdef ASSERT
3516 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3517 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3518 }
3519 #endif
3520 _nodes.map(n->_idx, 0); // This node is useless
3521 _deadlist.push(n);
3522 return;
3523 }
3524 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3526 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3527 Node *least = legal; // Best legal position so far
3528 while( early != legal ) { // While not at earliest legal
3529 #ifdef ASSERT
3530 if (legal->is_Start() && !early->is_Root()) {
3531 // Bad graph. Print idom path and fail.
3532 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
3533 assert(false, "Bad graph detected in build_loop_late");
3534 }
3535 #endif
3536 // Find least loop nesting depth
3537 legal = idom(legal); // Bump up the IDOM tree
3538 // Check for lower nesting depth
3539 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3540 least = legal;
3541 }
3542 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3544 // Try not to place code on a loop entry projection
3545 // which can inhibit range check elimination.
3546 if (least != early) {
3547 Node* ctrl_out = least->unique_ctrl_out();
3548 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3549 least == ctrl_out->in(LoopNode::EntryControl)) {
3550 Node* least_dom = idom(least);
3551 if (get_loop(least_dom)->is_member(get_loop(least))) {
3552 least = least_dom;
3553 }
3554 }
3555 }
3557 #ifdef ASSERT
3558 // If verifying, verify that 'verify_me' has a legal location
3559 // and choose it as our location.
3560 if( _verify_me ) {
3561 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3562 Node *legal = LCA;
3563 while( early != legal ) { // While not at earliest legal
3564 if( legal == v_ctrl ) break; // Check for prior good location
3565 legal = idom(legal) ;// Bump up the IDOM tree
3566 }
3567 // Check for prior good location
3568 if( legal == v_ctrl ) least = legal; // Keep prior if found
3569 }
3570 #endif
3572 // Assign discovered "here or above" point
3573 least = find_non_split_ctrl(least);
3574 set_ctrl(n, least);
3576 // Collect inner loop bodies
3577 IdealLoopTree *chosen_loop = get_loop(least);
3578 if( !chosen_loop->_child ) // Inner loop?
3579 chosen_loop->_body.push(n);// Collect inner loops
3580 }
3582 #ifdef ASSERT
3583 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
3584 tty->print_cr("%s", msg);
3585 tty->print("n: "); n->dump();
3586 tty->print("early(n): "); early->dump();
3587 if (n->in(0) != NULL && !n->in(0)->is_top() &&
3588 n->in(0) != early && !n->in(0)->is_Root()) {
3589 tty->print("n->in(0): "); n->in(0)->dump();
3590 }
3591 for (uint i = 1; i < n->req(); i++) {
3592 Node* in1 = n->in(i);
3593 if (in1 != NULL && in1 != n && !in1->is_top()) {
3594 tty->print("n->in(%d): ", i); in1->dump();
3595 Node* in1_early = get_ctrl(in1);
3596 tty->print("early(n->in(%d)): ", i); in1_early->dump();
3597 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
3598 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3599 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3600 }
3601 for (uint j = 1; j < in1->req(); j++) {
3602 Node* in2 = in1->in(j);
3603 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3604 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3605 Node* in2_early = get_ctrl(in2);
3606 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3607 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
3608 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3609 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3610 }
3611 }
3612 }
3613 }
3614 }
3615 tty->cr();
3616 tty->print("LCA(n): "); LCA->dump();
3617 for (uint i = 0; i < n->outcnt(); i++) {
3618 Node* u1 = n->raw_out(i);
3619 if (u1 == n)
3620 continue;
3621 tty->print("n->out(%d): ", i); u1->dump();
3622 if (u1->is_CFG()) {
3623 for (uint j = 0; j < u1->outcnt(); j++) {
3624 Node* u2 = u1->raw_out(j);
3625 if (u2 != u1 && u2 != n && u2->is_CFG()) {
3626 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3627 }
3628 }
3629 } else {
3630 Node* u1_later = get_ctrl(u1);
3631 tty->print("later(n->out(%d)): ", i); u1_later->dump();
3632 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
3633 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3634 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3635 }
3636 for (uint j = 0; j < u1->outcnt(); j++) {
3637 Node* u2 = u1->raw_out(j);
3638 if (u2 == n || u2 == u1)
3639 continue;
3640 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3641 if (!u2->is_CFG()) {
3642 Node* u2_later = get_ctrl(u2);
3643 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3644 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
3645 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3646 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3647 }
3648 }
3649 }
3650 }
3651 }
3652 tty->cr();
3653 int ct = 0;
3654 Node *dbg_legal = LCA;
3655 while(!dbg_legal->is_Start() && ct < 100) {
3656 tty->print("idom[%d] ",ct); dbg_legal->dump();
3657 ct++;
3658 dbg_legal = idom(dbg_legal);
3659 }
3660 tty->cr();
3661 }
3662 #endif
3664 #ifndef PRODUCT
3665 //------------------------------dump-------------------------------------------
3666 void PhaseIdealLoop::dump( ) const {
3667 ResourceMark rm;
3668 Arena* arena = Thread::current()->resource_area();
3669 Node_Stack stack(arena, C->live_nodes() >> 2);
3670 Node_List rpo_list;
3671 VectorSet visited(arena);
3672 visited.set(C->top()->_idx);
3673 rpo( C->root(), stack, visited, rpo_list );
3674 // Dump root loop indexed by last element in PO order
3675 dump( _ltree_root, rpo_list.size(), rpo_list );
3676 }
3678 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3679 loop->dump_head();
3681 // Now scan for CFG nodes in the same loop
3682 for( uint j=idx; j > 0; j-- ) {
3683 Node *n = rpo_list[j-1];
3684 if( !_nodes[n->_idx] ) // Skip dead nodes
3685 continue;
3686 if( get_loop(n) != loop ) { // Wrong loop nest
3687 if( get_loop(n)->_head == n && // Found nested loop?
3688 get_loop(n)->_parent == loop )
3689 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3690 continue;
3691 }
3693 // Dump controlling node
3694 for( uint x = 0; x < loop->_nest; x++ )
3695 tty->print(" ");
3696 tty->print("C");
3697 if( n == C->root() ) {
3698 n->dump();
3699 } else {
3700 Node* cached_idom = idom_no_update(n);
3701 Node *computed_idom = n->in(0);
3702 if( n->is_Region() ) {
3703 computed_idom = compute_idom(n);
3704 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3705 // any MultiBranch ctrl node), so apply a similar transform to
3706 // the cached idom returned from idom_no_update.
3707 cached_idom = find_non_split_ctrl(cached_idom);
3708 }
3709 tty->print(" ID:%d",computed_idom->_idx);
3710 n->dump();
3711 if( cached_idom != computed_idom ) {
3712 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3713 computed_idom->_idx, cached_idom->_idx);
3714 }
3715 }
3716 // Dump nodes it controls
3717 for( uint k = 0; k < _nodes.Size(); k++ ) {
3718 // (k < C->unique() && get_ctrl(find(k)) == n)
3719 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3720 Node *m = C->root()->find(k);
3721 if( m && m->outcnt() > 0 ) {
3722 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3723 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3724 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3725 }
3726 for( uint j = 0; j < loop->_nest; j++ )
3727 tty->print(" ");
3728 tty->print(" ");
3729 m->dump();
3730 }
3731 }
3732 }
3733 }
3734 }
3736 // Collect a R-P-O for the whole CFG.
3737 // Result list is in post-order (scan backwards for RPO)
3738 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3739 stk.push(start, 0);
3740 visited.set(start->_idx);
3742 while (stk.is_nonempty()) {
3743 Node* m = stk.node();
3744 uint idx = stk.index();
3745 if (idx < m->outcnt()) {
3746 stk.set_index(idx + 1);
3747 Node* n = m->raw_out(idx);
3748 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3749 stk.push(n, 0);
3750 }
3751 } else {
3752 rpo_list.push(m);
3753 stk.pop();
3754 }
3755 }
3756 }
3757 #endif
3760 //=============================================================================
3761 //------------------------------LoopTreeIterator-----------------------------------
3763 // Advance to next loop tree using a preorder, left-to-right traversal.
3764 void LoopTreeIterator::next() {
3765 assert(!done(), "must not be done.");
3766 if (_curnt->_child != NULL) {
3767 _curnt = _curnt->_child;
3768 } else if (_curnt->_next != NULL) {
3769 _curnt = _curnt->_next;
3770 } else {
3771 while (_curnt != _root && _curnt->_next == NULL) {
3772 _curnt = _curnt->_parent;
3773 }
3774 if (_curnt == _root) {
3775 _curnt = NULL;
3776 assert(done(), "must be done.");
3777 } else {
3778 assert(_curnt->_next != NULL, "must be more to do");
3779 _curnt = _curnt->_next;
3780 }
3781 }
3782 }