1 /*
2 * Copyright (c) 1998, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "memory/resourceArea.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/idealGraphPrinter.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/superword.hpp"
41
42 //=============================================================================
43 //------------------------------is_loop_iv-------------------------------------
44 // Determine if a node is Counted loop induction variable.
45 // The method is declared in node.hpp.
46 const Node* Node::is_loop_iv() const {
47 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
48 this->as_Phi()->region()->is_CountedLoop() &&
49 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
50 return this;
51 } else {
52 return NULL;
53 }
54 }
55
56 //=============================================================================
57 //------------------------------dump_spec--------------------------------------
58 // Dump special per-node info
59 #ifndef PRODUCT
60 void LoopNode::dump_spec(outputStream *st) const {
61 if (is_inner_loop()) st->print( "inner " );
62 if (is_partial_peel_loop()) st->print( "partial_peel " );
63 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
64 }
65 #endif
66
67 //------------------------------is_valid_counted_loop-------------------------
68 bool LoopNode::is_valid_counted_loop() const {
69 if (is_CountedLoop()) {
70 CountedLoopNode* l = as_CountedLoop();
71 CountedLoopEndNode* le = l->loopexit();
72 if (le != NULL &&
73 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
74 Node* phi = l->phi();
75 Node* exit = le->proj_out(0 /* false */);
76 if (exit != NULL && exit->Opcode() == Opcodes::Op_IfFalse &&
77 phi != NULL && phi->is_Phi() &&
78 phi->in(LoopNode::LoopBackControl) == l->incr() &&
79 le->loopnode() == l && le->stride_is_con()) {
80 return true;
81 }
82 }
83 }
84 return false;
85 }
86
87 //------------------------------get_early_ctrl---------------------------------
88 // Compute earliest legal control
89 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
90 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
91 uint i;
92 Node *early;
93 if (n->in(0) && !n->is_expensive()) {
94 early = n->in(0);
95 if (!early->is_CFG()) // Might be a non-CFG multi-def
96 early = get_ctrl(early); // So treat input as a straight data input
97 i = 1;
98 } else {
99 early = get_ctrl(n->in(1));
100 i = 2;
101 }
102 uint e_d = dom_depth(early);
103 assert( early, "" );
104 for (; i < n->req(); i++) {
105 Node *cin = get_ctrl(n->in(i));
106 assert( cin, "" );
107 // Keep deepest dominator depth
108 uint c_d = dom_depth(cin);
109 if (c_d > e_d) { // Deeper guy?
110 early = cin; // Keep deepest found so far
111 e_d = c_d;
112 } else if (c_d == e_d && // Same depth?
113 early != cin) { // If not equal, must use slower algorithm
114 // If same depth but not equal, one _must_ dominate the other
115 // and we want the deeper (i.e., dominated) guy.
116 Node *n1 = early;
117 Node *n2 = cin;
118 while (1) {
119 n1 = idom(n1); // Walk up until break cycle
120 n2 = idom(n2);
121 if (n1 == cin || // Walked early up to cin
122 dom_depth(n2) < c_d)
123 break; // early is deeper; keep him
124 if (n2 == early || // Walked cin up to early
125 dom_depth(n1) < c_d) {
126 early = cin; // cin is deeper; keep him
127 break;
128 }
129 }
130 e_d = dom_depth(early); // Reset depth register cache
131 }
132 }
133
134 // Return earliest legal location
135 assert(early == find_non_split_ctrl(early), "unexpected early control");
136
137 if (n->is_expensive() && !_verify_only && !_verify_me) {
138 assert(n->in(0), "should have control input");
139 early = get_early_ctrl_for_expensive(n, early);
140 }
141
142 return early;
143 }
144
145 //------------------------------get_early_ctrl_for_expensive---------------------------------
146 // Move node up the dominator tree as high as legal while still beneficial
147 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
148 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
149 assert(OptimizeExpensiveOps, "optimization off?");
150
151 Node* ctl = n->in(0);
152 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
153 uint min_dom_depth = dom_depth(earliest);
154 #ifdef ASSERT
155 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
156 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
157 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
158 }
159 #endif
160 if (dom_depth(ctl) < min_dom_depth) {
161 return earliest;
162 }
163
164 while (1) {
165 Node *next = ctl;
166 // Moving the node out of a loop on the projection of a If
167 // confuses loop predication. So once we hit a Loop in a If branch
168 // that doesn't branch to an UNC, we stop. The code that process
169 // expensive nodes will notice the loop and skip over it to try to
170 // move the node further up.
171 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
172 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
173 break;
174 }
175 next = idom(ctl->in(1)->in(0));
176 } else if (ctl->is_Proj()) {
177 // We only move it up along a projection if the projection is
178 // the single control projection for its parent: same code path,
179 // if it's a If with UNC or fallthrough of a call.
180 Node* parent_ctl = ctl->in(0);
181 if (parent_ctl == NULL) {
182 break;
183 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
184 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
185 } else if (parent_ctl->is_If()) {
186 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
187 break;
188 }
189 assert(idom(ctl) == parent_ctl, "strange");
190 next = idom(parent_ctl);
191 } else if (ctl->is_CatchProj()) {
192 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
193 break;
194 }
195 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
196 next = parent_ctl->in(0)->in(0)->in(0);
197 } else {
198 // Check if parent control has a single projection (this
199 // control is the only possible successor of the parent
200 // control). If so, we can try to move the node above the
201 // parent control.
202 int nb_ctl_proj = 0;
203 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
204 Node *p = parent_ctl->fast_out(i);
205 if (p->is_Proj() && p->is_CFG()) {
206 nb_ctl_proj++;
207 if (nb_ctl_proj > 1) {
208 break;
209 }
210 }
211 }
212
213 if (nb_ctl_proj > 1) {
214 break;
215 }
216 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node");
217 assert(idom(ctl) == parent_ctl, "strange");
218 next = idom(parent_ctl);
219 }
220 } else {
221 next = idom(ctl);
222 }
223 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
224 break;
225 }
226 ctl = next;
227 }
228
229 if (ctl != n->in(0)) {
230 _igvn.replace_input_of(n, 0, ctl);
231 _igvn.hash_insert(n);
232 }
233
234 return ctl;
235 }
236
237
238 //------------------------------set_early_ctrl---------------------------------
239 // Set earliest legal control
240 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
241 Node *early = get_early_ctrl(n);
242
243 // Record earliest legal location
244 set_ctrl(n, early);
245 }
246
247 //------------------------------set_subtree_ctrl-------------------------------
248 // set missing _ctrl entries on new nodes
249 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
250 // Already set? Get out.
251 if( _nodes[n->_idx] ) return;
252 // Recursively set _nodes array to indicate where the Node goes
253 uint i;
254 for( i = 0; i < n->req(); ++i ) {
255 Node *m = n->in(i);
256 if( m && m != C->root() )
257 set_subtree_ctrl( m );
258 }
259
260 // Fixup self
261 set_early_ctrl( n );
262 }
263
264 //------------------------------is_counted_loop--------------------------------
265 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
266 PhaseGVN *gvn = &_igvn;
267
268 // Counted loop head must be a good RegionNode with only 3 not NULL
269 // control input edges: Self, Entry, LoopBack.
270 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
271 return false;
272 }
273 Node *init_control = x->in(LoopNode::EntryControl);
274 Node *back_control = x->in(LoopNode::LoopBackControl);
275 if (init_control == NULL || back_control == NULL) // Partially dead
276 return false;
277 // Must also check for TOP when looking for a dead loop
278 if (init_control->is_top() || back_control->is_top())
279 return false;
280
281 // Allow funny placement of Safepoint
282 if (back_control->Opcode() == Opcodes::Op_SafePoint)
283 back_control = back_control->in(TypeFunc::Control);
284
285 // Controlling test for loop
286 Node *iftrue = back_control;
287 Opcodes iftrue_op = iftrue->Opcode();
288 if (iftrue_op != Opcodes::Op_IfTrue &&
289 iftrue_op != Opcodes::Op_IfFalse)
290 // I have a weird back-control. Probably the loop-exit test is in
291 // the middle of the loop and I am looking at some trailing control-flow
292 // merge point. To fix this I would have to partially peel the loop.
293 return false; // Obscure back-control
294
295 // Get boolean guarding loop-back test
296 Node *iff = iftrue->in(0);
297 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
298 return false;
299 BoolNode *test = iff->in(1)->as_Bool();
300 BoolTest::mask bt = test->_test._test;
301 float cl_prob = iff->as_If()->_prob;
302 if (iftrue_op == Opcodes::Op_IfFalse) {
303 bt = BoolTest(bt).negate();
304 cl_prob = 1.0 - cl_prob;
305 }
306 // Get backedge compare
307 Node *cmp = test->in(1);
308 Opcodes cmp_op = cmp->Opcode();
309 if (cmp_op != Opcodes::Op_CmpI)
310 return false; // Avoid pointer & float compares
311
312 // Find the trip-counter increment & limit. Limit must be loop invariant.
313 Node *incr = cmp->in(1);
314 Node *limit = cmp->in(2);
315
316 // ---------
317 // need 'loop()' test to tell if limit is loop invariant
318 // ---------
319
320 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
321 Node *tmp = incr; // Then reverse order into the CmpI
322 incr = limit;
323 limit = tmp;
324 bt = BoolTest(bt).commute(); // And commute the exit test
325 }
326 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
327 return false;
328 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
329 return false;
330
331 Node* phi_incr = NULL;
332 // Trip-counter increment must be commutative & associative.
333 if (incr->Opcode() == Opcodes::Op_CastII) {
334 incr = incr->in(1);
335 }
336 if (incr->is_Phi()) {
337 if (incr->as_Phi()->region() != x || incr->req() != 3)
338 return false; // Not simple trip counter expression
339 phi_incr = incr;
340 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
341 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
342 return false;
343 }
344
345 Node* trunc1 = NULL;
346 Node* trunc2 = NULL;
347 const TypeInt* iv_trunc_t = NULL;
348 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
349 return false; // Funny increment opcode
350 }
351 assert(incr->Opcode() == Opcodes::Op_AddI, "wrong increment code");
352
353 // Get merge point
354 Node *xphi = incr->in(1);
355 Node *stride = incr->in(2);
356 if (!stride->is_Con()) { // Oops, swap these
357 if (!xphi->is_Con()) // Is the other guy a constant?
358 return false; // Nope, unknown stride, bail out
359 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
360 xphi = stride;
361 stride = tmp;
362 }
363 if (xphi->Opcode() == Opcodes::Op_CastII) {
364 xphi = xphi->in(1);
365 }
366 // Stride must be constant
367 int stride_con = stride->get_int();
368 if (stride_con == 0)
369 return false; // missed some peephole opt
370
371 if (!xphi->is_Phi())
372 return false; // Too much math on the trip counter
373 if (phi_incr != NULL && phi_incr != xphi)
374 return false;
375 PhiNode *phi = xphi->as_Phi();
376
377 // Phi must be of loop header; backedge must wrap to increment
378 if (phi->region() != x)
379 return false;
380 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
381 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
382 return false;
383 }
384 Node *init_trip = phi->in(LoopNode::EntryControl);
385
386 // If iv trunc type is smaller than int, check for possible wrap.
387 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
388 assert(trunc1 != NULL, "must have found some truncation");
389
390 // Get a better type for the phi (filtered thru if's)
391 const TypeInt* phi_ft = filtered_type(phi);
392
393 // Can iv take on a value that will wrap?
394 //
395 // Ensure iv's limit is not within "stride" of the wrap value.
396 //
397 // Example for "short" type
398 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
399 // If the stride is +10, then the last value of the induction
400 // variable before the increment (phi_ft->_hi) must be
401 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
402 // ensure no truncation occurs after the increment.
403
404 if (stride_con > 0) {
405 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
406 iv_trunc_t->_lo > phi_ft->_lo) {
407 return false; // truncation may occur
408 }
409 } else if (stride_con < 0) {
410 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
411 iv_trunc_t->_hi < phi_ft->_hi) {
412 return false; // truncation may occur
413 }
414 }
415 // No possibility of wrap so truncation can be discarded
416 // Promote iv type to Int
417 } else {
418 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
419 }
420
421 // If the condition is inverted and we will be rolling
422 // through MININT to MAXINT, then bail out.
423 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
424 // Odd stride
425 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
426 // Count down loop rolls through MAXINT
427 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
428 // Count up loop rolls through MININT
429 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
430 return false; // Bail out
431 }
432
433 const TypeInt* init_t = gvn->type(init_trip)->is_int();
434 const TypeInt* limit_t = gvn->type(limit)->is_int();
435
436 if (stride_con > 0) {
437 jlong init_p = (jlong)init_t->_lo + stride_con;
438 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
439 return false; // cyclic loop or this loop trips only once
440 } else {
441 jlong init_p = (jlong)init_t->_hi + stride_con;
442 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
443 return false; // cyclic loop or this loop trips only once
444 }
445
446 if (phi_incr != NULL) {
447 // check if there is a possiblity of IV overflowing after the first increment
448 if (stride_con > 0) {
449 if (init_t->_hi > max_jint - stride_con) {
450 return false;
451 }
452 } else {
453 if (init_t->_lo < min_jint - stride_con) {
454 return false;
455 }
456 }
457 }
458
459 // =================================================
460 // ---- SUCCESS! Found A Trip-Counted Loop! -----
461 //
462 assert(x->Opcode() == Opcodes::Op_Loop, "regular loops only");
463 C->print_method(PHASE_BEFORE_CLOOPS, 3);
464
465 Node *hook = new Node(6);
466
467 // ===================================================
468 // Generate loop limit check to avoid integer overflow
469 // in cases like next (cyclic loops):
470 //
471 // for (i=0; i <= max_jint; i++) {}
472 // for (i=0; i < max_jint; i+=2) {}
473 //
474 //
475 // Limit check predicate depends on the loop test:
476 //
477 // for(;i != limit; i++) --> limit <= (max_jint)
478 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
479 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
480 //
481
482 // Check if limit is excluded to do more precise int overflow check.
483 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
484 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
485
486 // If compare points directly to the phi we need to adjust
487 // the compare so that it points to the incr. Limit have
488 // to be adjusted to keep trip count the same and the
489 // adjusted limit should be checked for int overflow.
490 if (phi_incr != NULL) {
491 stride_m += stride_con;
492 }
493
494 if (limit->is_Con()) {
495 int limit_con = limit->get_int();
496 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
497 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
498 // Bailout: it could be integer overflow.
499 return false;
500 }
501 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
502 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
503 // Limit's type may satisfy the condition, for example,
504 // when it is an array length.
505 } else {
506 // Generate loop's limit check.
507 // Loop limit check predicate should be near the loop.
508 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
509 if (!limit_check_proj) {
510 // The limit check predicate is not generated if this method trapped here before.
511 #ifdef ASSERT
512 if (TraceLoopLimitCheck) {
513 tty->print("missing loop limit check:");
514 loop->dump_head();
515 x->dump(1);
516 }
517 #endif
518 return false;
519 }
520
521 IfNode* check_iff = limit_check_proj->in(0)->as_If();
522 Node* cmp_limit;
523 Node* bol;
524
525 if (stride_con > 0) {
526 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
527 bol = new BoolNode(cmp_limit, BoolTest::le);
528 } else {
529 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
530 bol = new BoolNode(cmp_limit, BoolTest::ge);
531 }
532 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
533 bol = _igvn.register_new_node_with_optimizer(bol);
534 set_subtree_ctrl(bol);
535
536 // Replace condition in original predicate but preserve Opaque node
537 // so that previous predicates could be found.
538 assert(check_iff->in(1)->Opcode() == Opcodes::Op_Conv2B &&
539 check_iff->in(1)->in(1)->Opcode() == Opcodes::Op_Opaque1, "");
540 Node* opq = check_iff->in(1)->in(1);
541 _igvn.replace_input_of(opq, 1, bol);
542 // Update ctrl.
543 set_ctrl(opq, check_iff->in(0));
544 set_ctrl(check_iff->in(1), check_iff->in(0));
545
546 #ifndef PRODUCT
547 // report that the loop predication has been actually performed
548 // for this loop
549 if (TraceLoopLimitCheck) {
550 tty->print_cr("Counted Loop Limit Check generated:");
551 debug_only( bol->dump(2); )
552 }
553 #endif
554 }
555
556 if (phi_incr != NULL) {
557 // If compare points directly to the phi we need to adjust
558 // the compare so that it points to the incr. Limit have
559 // to be adjusted to keep trip count the same and we
560 // should avoid int overflow.
561 //
562 // i = init; do {} while(i++ < limit);
563 // is converted to
564 // i = init; do {} while(++i < limit+1);
565 //
566 limit = gvn->transform(new AddINode(limit, stride));
567 }
568
569 // Now we need to canonicalize loop condition.
570 if (bt == BoolTest::ne) {
571 assert(stride_con == 1 || stride_con == -1, "simple increment only");
572 // 'ne' can be replaced with 'lt' only when init < limit.
573 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
574 bt = BoolTest::lt;
575 // 'ne' can be replaced with 'gt' only when init > limit.
576 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
577 bt = BoolTest::gt;
578 }
579
580 if (incl_limit) {
581 // The limit check guaranties that 'limit <= (max_jint - stride)' so
582 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
583 //
584 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
585 limit = gvn->transform(new AddINode(limit, one));
586 if (bt == BoolTest::le)
587 bt = BoolTest::lt;
588 else if (bt == BoolTest::ge)
589 bt = BoolTest::gt;
590 else
591 ShouldNotReachHere();
592 }
593 set_subtree_ctrl( limit );
594
595 if (!UseCountedLoopSafepoints) {
596 // Check for SafePoint on backedge and remove
597 Node *sfpt = x->in(LoopNode::LoopBackControl);
598 if (sfpt->Opcode() == Opcodes::Op_SafePoint && is_deleteable_safept(sfpt)) {
599 lazy_replace( sfpt, iftrue );
600 if (loop->_safepts != NULL) {
601 loop->_safepts->yank(sfpt);
602 }
603 loop->_tail = iftrue;
604 }
605 }
606
607 // Build a canonical trip test.
608 // Clone code, as old values may be in use.
609 incr = incr->clone();
610 incr->set_req(1,phi);
611 incr->set_req(2,stride);
612 incr = _igvn.register_new_node_with_optimizer(incr);
613 set_early_ctrl( incr );
614 _igvn.rehash_node_delayed(phi);
615 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
616
617 // If phi type is more restrictive than Int, raise to
618 // Int to prevent (almost) infinite recursion in igvn
619 // which can only handle integer types for constants or minint..maxint.
620 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
621 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
622 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
623 nphi = _igvn.register_new_node_with_optimizer(nphi);
624 set_ctrl(nphi, get_ctrl(phi));
625 _igvn.replace_node(phi, nphi);
626 phi = nphi->as_Phi();
627 }
628 cmp = cmp->clone();
629 cmp->set_req(1,incr);
630 cmp->set_req(2,limit);
631 cmp = _igvn.register_new_node_with_optimizer(cmp);
632 set_ctrl(cmp, iff->in(0));
633
634 test = test->clone()->as_Bool();
635 (*(BoolTest*)&test->_test)._test = bt;
636 test->set_req(1,cmp);
637 _igvn.register_new_node_with_optimizer(test);
638 set_ctrl(test, iff->in(0));
639
640 // Replace the old IfNode with a new LoopEndNode
641 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
642 IfNode *le = lex->as_If();
643 uint dd = dom_depth(iff);
644 set_idom(le, le->in(0), dd); // Update dominance for loop exit
645 set_loop(le, loop);
646
647 // Get the loop-exit control
648 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Opcodes::Op_IfTrue));
649
650 // Need to swap loop-exit and loop-back control?
651 if (iftrue_op == Opcodes::Op_IfFalse) {
652 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
653 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
654
655 loop->_tail = back_control = ift2;
656 set_loop(ift2, loop);
657 set_loop(iff2, get_loop(iffalse));
658
659 // Lazy update of 'get_ctrl' mechanism.
660 lazy_replace(iffalse, iff2);
661 lazy_replace(iftrue, ift2);
662
663 // Swap names
664 iffalse = iff2;
665 iftrue = ift2;
666 } else {
667 _igvn.rehash_node_delayed(iffalse);
668 _igvn.rehash_node_delayed(iftrue);
669 iffalse->set_req_X( 0, le, &_igvn );
670 iftrue ->set_req_X( 0, le, &_igvn );
671 }
672
673 set_idom(iftrue, le, dd+1);
674 set_idom(iffalse, le, dd+1);
675 assert(iff->outcnt() == 0, "should be dead now");
676 lazy_replace( iff, le ); // fix 'get_ctrl'
677
678 // Now setup a new CountedLoopNode to replace the existing LoopNode
679 CountedLoopNode *l = new CountedLoopNode(init_control, back_control);
680 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
681 // The following assert is approximately true, and defines the intention
682 // of can_be_counted_loop. It fails, however, because phase->type
683 // is not yet initialized for this loop and its parts.
684 //assert(l->can_be_counted_loop(this), "sanity");
685 _igvn.register_new_node_with_optimizer(l);
686 set_loop(l, loop);
687 loop->_head = l;
688 // Fix all data nodes placed at the old loop head.
689 // Uses the lazy-update mechanism of 'get_ctrl'.
690 lazy_replace( x, l );
691 set_idom(l, init_control, dom_depth(x));
692
693 if (!UseCountedLoopSafepoints) {
694 // Check for immediately preceding SafePoint and remove
695 Node *sfpt2 = le->in(0);
696 if (sfpt2->Opcode() == Opcodes::Op_SafePoint && is_deleteable_safept(sfpt2)) {
697 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
698 if (loop->_safepts != NULL) {
699 loop->_safepts->yank(sfpt2);
700 }
701 }
702 }
703
704 // Free up intermediate goo
705 _igvn.remove_dead_node(hook);
706
707 #ifdef ASSERT
708 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
709 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
710 #endif
711 #ifndef PRODUCT
712 if (TraceLoopOpts) {
713 tty->print("Counted ");
714 loop->dump_head();
715 }
716 #endif
717
718 C->print_method(PHASE_AFTER_CLOOPS, 3);
719
720 // Capture bounds of the loop in the induction variable Phi before
721 // subsequent transformation (iteration splitting) obscures the
722 // bounds
723 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
724
725 return true;
726 }
727
728 //----------------------exact_limit-------------------------------------------
729 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
730 assert(loop->_head->is_CountedLoop(), "");
731 CountedLoopNode *cl = loop->_head->as_CountedLoop();
732 assert(cl->is_valid_counted_loop(), "");
733
734 if (ABS(cl->stride_con()) == 1 ||
735 cl->limit()->Opcode() == Opcodes::Op_LoopLimit) {
736 // Old code has exact limit (it could be incorrect in case of int overflow).
737 // Loop limit is exact with stride == 1. And loop may already have exact limit.
738 return cl->limit();
739 }
740 Node *limit = NULL;
741 #ifdef ASSERT
742 BoolTest::mask bt = cl->loopexit()->test_trip();
743 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
744 #endif
745 if (cl->has_exact_trip_count()) {
746 // Simple case: loop has constant boundaries.
747 // Use jlongs to avoid integer overflow.
748 int stride_con = cl->stride_con();
749 jlong init_con = cl->init_trip()->get_int();
750 jlong limit_con = cl->limit()->get_int();
751 julong trip_cnt = cl->trip_count();
752 jlong final_con = init_con + trip_cnt*stride_con;
753 int final_int = (int)final_con;
754 // The final value should be in integer range since the loop
755 // is counted and the limit was checked for overflow.
756 assert(final_con == (jlong)final_int, "final value should be integer");
757 limit = _igvn.intcon(final_int);
758 } else {
759 // Create new LoopLimit node to get exact limit (final iv value).
760 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
761 register_new_node(limit, cl->in(LoopNode::EntryControl));
762 }
763 assert(limit != NULL, "sanity");
764 return limit;
765 }
766
767 //------------------------------Ideal------------------------------------------
768 // Return a node which is more "ideal" than the current node.
769 // Attempt to convert into a counted-loop.
770 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
771 if (!can_be_counted_loop(phase)) {
772 phase->C->set_major_progress();
773 }
774 return RegionNode::Ideal(phase, can_reshape);
775 }
776
777
778 //=============================================================================
779 //------------------------------Ideal------------------------------------------
780 // Return a node which is more "ideal" than the current node.
781 // Attempt to convert into a counted-loop.
782 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
783 return RegionNode::Ideal(phase, can_reshape);
784 }
785
786 //------------------------------dump_spec--------------------------------------
787 // Dump special per-node info
788 #ifndef PRODUCT
789 void CountedLoopNode::dump_spec(outputStream *st) const {
790 LoopNode::dump_spec(st);
791 if (stride_is_con()) {
792 st->print("stride: %d ",stride_con());
793 }
794 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
795 if (is_main_loop()) st->print("main of N%d", _idx);
796 if (is_post_loop()) st->print("post of N%d", _main_idx);
797 }
798 #endif
799
800 //=============================================================================
801 int CountedLoopEndNode::stride_con() const {
802 return stride()->bottom_type()->is_int()->get_con();
803 }
804
805 //=============================================================================
806 //------------------------------Value-----------------------------------------
807 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
808 const Type* init_t = phase->type(in(Init));
809 const Type* limit_t = phase->type(in(Limit));
810 const Type* stride_t = phase->type(in(Stride));
811 // Either input is TOP ==> the result is TOP
812 if (init_t == Type::TOP) return Type::TOP;
813 if (limit_t == Type::TOP) return Type::TOP;
814 if (stride_t == Type::TOP) return Type::TOP;
815
816 int stride_con = stride_t->is_int()->get_con();
817 if (stride_con == 1)
818 return NULL; // Identity
819
820 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
821 // Use jlongs to avoid integer overflow.
822 jlong init_con = init_t->is_int()->get_con();
823 jlong limit_con = limit_t->is_int()->get_con();
824 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
825 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
826 jlong final_con = init_con + stride_con*trip_count;
827 int final_int = (int)final_con;
828 // The final value should be in integer range since the loop
829 // is counted and the limit was checked for overflow.
830 assert(final_con == (jlong)final_int, "final value should be integer");
831 return TypeInt::make(final_int);
832 }
833
834 return bottom_type(); // TypeInt::INT
835 }
836
837 //------------------------------Ideal------------------------------------------
838 // Return a node which is more "ideal" than the current node.
839 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
840 if (phase->type(in(Init)) == Type::TOP ||
841 phase->type(in(Limit)) == Type::TOP ||
842 phase->type(in(Stride)) == Type::TOP)
843 return NULL; // Dead
844
845 int stride_con = phase->type(in(Stride))->is_int()->get_con();
846 if (stride_con == 1)
847 return NULL; // Identity
848
849 if (in(Init)->is_Con() && in(Limit)->is_Con())
850 return NULL; // Value
851
852 // Delay following optimizations until all loop optimizations
853 // done to keep Ideal graph simple.
854 if (!can_reshape || phase->C->major_progress())
855 return NULL;
856
857 const TypeInt* init_t = phase->type(in(Init) )->is_int();
858 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
859 int stride_p;
860 jlong lim, ini;
861 julong max;
862 if (stride_con > 0) {
863 stride_p = stride_con;
864 lim = limit_t->_hi;
865 ini = init_t->_lo;
866 max = (julong)max_jint;
867 } else {
868 stride_p = -stride_con;
869 lim = init_t->_hi;
870 ini = limit_t->_lo;
871 max = (julong)min_jint;
872 }
873 julong range = lim - ini + stride_p;
874 if (range <= max) {
875 // Convert to integer expression if it is not overflow.
876 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
877 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
878 Node *bias = phase->transform(new AddINode(range, stride_m));
879 Node *trip = phase->transform(new DivINode(0, bias, in(Stride)));
880 Node *span = phase->transform(new MulINode(trip, in(Stride)));
881 return new AddINode(span, in(Init)); // exact limit
882 }
883
884 if (is_power_of_2(stride_p) || // divisor is 2^n
885 !Matcher::has_match_rule(Opcodes::Op_LoopLimit)) { // or no specialized Mach node?
886 // Convert to long expression to avoid integer overflow
887 // and let igvn optimizer convert this division.
888 //
889 Node* init = phase->transform( new ConvI2LNode(in(Init)));
890 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
891 Node* stride = phase->longcon(stride_con);
892 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
893
894 Node *range = phase->transform(new SubLNode(limit, init));
895 Node *bias = phase->transform(new AddLNode(range, stride_m));
896 Node *span;
897 if (stride_con > 0 && is_power_of_2(stride_p)) {
898 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
899 // and avoid generating rounding for division. Zero trip guard should
900 // guarantee that init < limit but sometimes the guard is missing and
901 // we can get situation when init > limit. Note, for the empty loop
902 // optimization zero trip guard is generated explicitly which leaves
903 // only RCE predicate where exact limit is used and the predicate
904 // will simply fail forcing recompilation.
905 Node* neg_stride = phase->longcon(-stride_con);
906 span = phase->transform(new AndLNode(bias, neg_stride));
907 } else {
908 Node *trip = phase->transform(new DivLNode(0, bias, stride));
909 span = phase->transform(new MulLNode(trip, stride));
910 }
911 // Convert back to int
912 Node *span_int = phase->transform(new ConvL2INode(span));
913 return new AddINode(span_int, in(Init)); // exact limit
914 }
915
916 return NULL; // No progress
917 }
918
919 //------------------------------Identity---------------------------------------
920 // If stride == 1 return limit node.
921 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
922 int stride_con = phase->type(in(Stride))->is_int()->get_con();
923 if (stride_con == 1 || stride_con == -1)
924 return in(Limit);
925 return this;
926 }
927
928 //=============================================================================
929 //----------------------match_incr_with_optional_truncation--------------------
930 // Match increment with optional truncation:
931 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
932 // Return NULL for failure. Success returns the increment node.
933 Node* CountedLoopNode::match_incr_with_optional_truncation(
934 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
935 // Quick cutouts:
936 if (expr == NULL || expr->req() != 3) return NULL;
937
938 Node *t1 = NULL;
939 Node *t2 = NULL;
940 const TypeInt* trunc_t = TypeInt::INT;
941 Node* n1 = expr;
942 Opcodes n1op = n1->Opcode();
943
944 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
945 if (n1op == Opcodes::Op_AndI &&
946 n1->in(2)->is_Con() &&
947 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
948 // %%% This check should match any mask of 2**K-1.
949 t1 = n1;
950 n1 = t1->in(1);
951 n1op = n1->Opcode();
952 trunc_t = TypeInt::CHAR;
953 } else if (n1op == Opcodes::Op_RShiftI &&
954 n1->in(1) != NULL &&
955 n1->in(1)->Opcode() == Opcodes::Op_LShiftI &&
956 n1->in(2) == n1->in(1)->in(2) &&
957 n1->in(2)->is_Con()) {
958 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
959 // %%% This check should match any shift in [1..31].
960 if (shift == 16 || shift == 8) {
961 t1 = n1;
962 t2 = t1->in(1);
963 n1 = t2->in(1);
964 n1op = n1->Opcode();
965 if (shift == 16) {
966 trunc_t = TypeInt::SHORT;
967 } else if (shift == 8) {
968 trunc_t = TypeInt::BYTE;
969 }
970 }
971 }
972
973 // If (maybe after stripping) it is an AddI, we won:
974 if (n1op == Opcodes::Op_AddI) {
975 *trunc1 = t1;
976 *trunc2 = t2;
977 *trunc_type = trunc_t;
978 return n1;
979 }
980
981 // failed
982 return NULL;
983 }
984
985
986 //------------------------------filtered_type--------------------------------
987 // Return a type based on condition control flow
988 // A successful return will be a type that is restricted due
989 // to a series of dominating if-tests, such as:
990 // if (i < 10) {
991 // if (i > 0) {
992 // here: "i" type is [1..10)
993 // }
994 // }
995 // or a control flow merge
996 // if (i < 10) {
997 // do {
998 // phi( , ) -- at top of loop type is [min_int..10)
999 // i = ?
1000 // } while ( i < 10)
1001 //
1002 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1003 assert(n && n->bottom_type()->is_int(), "must be int");
1004 const TypeInt* filtered_t = NULL;
1005 if (!n->is_Phi()) {
1006 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1007 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1008
1009 } else {
1010 Node* phi = n->as_Phi();
1011 Node* region = phi->in(0);
1012 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1013 if (region && region != C->top()) {
1014 for (uint i = 1; i < phi->req(); i++) {
1015 Node* val = phi->in(i);
1016 Node* use_c = region->in(i);
1017 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1018 if (val_t != NULL) {
1019 if (filtered_t == NULL) {
1020 filtered_t = val_t;
1021 } else {
1022 filtered_t = filtered_t->meet(val_t)->is_int();
1023 }
1024 }
1025 }
1026 }
1027 }
1028 const TypeInt* n_t = _igvn.type(n)->is_int();
1029 if (filtered_t != NULL) {
1030 n_t = n_t->join(filtered_t)->is_int();
1031 }
1032 return n_t;
1033 }
1034
1035
1036 //------------------------------filtered_type_from_dominators--------------------------------
1037 // Return a possibly more restrictive type for val based on condition control flow of dominators
1038 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1039 if (val->is_Con()) {
1040 return val->bottom_type()->is_int();
1041 }
1042 uint if_limit = 10; // Max number of dominating if's visited
1043 const TypeInt* rtn_t = NULL;
1044
1045 if (use_ctrl && use_ctrl != C->top()) {
1046 Node* val_ctrl = get_ctrl(val);
1047 uint val_dom_depth = dom_depth(val_ctrl);
1048 Node* pred = use_ctrl;
1049 uint if_cnt = 0;
1050 while (if_cnt < if_limit) {
1051 if ((pred->Opcode() == Opcodes::Op_IfTrue || pred->Opcode() == Opcodes::Op_IfFalse)) {
1052 if_cnt++;
1053 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1054 if (if_t != NULL) {
1055 if (rtn_t == NULL) {
1056 rtn_t = if_t;
1057 } else {
1058 rtn_t = rtn_t->join(if_t)->is_int();
1059 }
1060 }
1061 }
1062 pred = idom(pred);
1063 if (pred == NULL || pred == C->top()) {
1064 break;
1065 }
1066 // Stop if going beyond definition block of val
1067 if (dom_depth(pred) < val_dom_depth) {
1068 break;
1069 }
1070 }
1071 }
1072 return rtn_t;
1073 }
1074
1075
1076 //------------------------------dump_spec--------------------------------------
1077 // Dump special per-node info
1078 #ifndef PRODUCT
1079 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1080 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1081 BoolTest bt( test_trip()); // Added this for g++.
1082
1083 st->print("[");
1084 bt.dump_on(st);
1085 st->print("]");
1086 }
1087 st->print(" ");
1088 IfNode::dump_spec(st);
1089 }
1090 #endif
1091
1092 //=============================================================================
1093 //------------------------------is_member--------------------------------------
1094 // Is 'l' a member of 'this'?
1095 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1096 while( l->_nest > _nest ) l = l->_parent;
1097 return l == this;
1098 }
1099
1100 //------------------------------set_nest---------------------------------------
1101 // Set loop tree nesting depth. Accumulate _has_call bits.
1102 int IdealLoopTree::set_nest( uint depth ) {
1103 _nest = depth;
1104 int bits = _has_call;
1105 if( _child ) bits |= _child->set_nest(depth+1);
1106 if( bits ) _has_call = 1;
1107 if( _next ) bits |= _next ->set_nest(depth );
1108 return bits;
1109 }
1110
1111 //------------------------------split_fall_in----------------------------------
1112 // Split out multiple fall-in edges from the loop header. Move them to a
1113 // private RegionNode before the loop. This becomes the loop landing pad.
1114 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1115 PhaseIterGVN &igvn = phase->_igvn;
1116 uint i;
1117
1118 // Make a new RegionNode to be the landing pad.
1119 Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1120 phase->set_loop(landing_pad,_parent);
1121 // Gather all the fall-in control paths into the landing pad
1122 uint icnt = fall_in_cnt;
1123 uint oreq = _head->req();
1124 for( i = oreq-1; i>0; i-- )
1125 if( !phase->is_member( this, _head->in(i) ) )
1126 landing_pad->set_req(icnt--,_head->in(i));
1127
1128 // Peel off PhiNode edges as well
1129 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1130 Node *oj = _head->fast_out(j);
1131 if( oj->is_Phi() ) {
1132 PhiNode* old_phi = oj->as_Phi();
1133 assert( old_phi->region() == _head, "" );
1134 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1135 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1136 uint icnt = fall_in_cnt;
1137 for( i = oreq-1; i>0; i-- ) {
1138 if( !phase->is_member( this, _head->in(i) ) ) {
1139 p->init_req(icnt--, old_phi->in(i));
1140 // Go ahead and clean out old edges from old phi
1141 old_phi->del_req(i);
1142 }
1143 }
1144 // Search for CSE's here, because ZKM.jar does a lot of
1145 // loop hackery and we need to be a little incremental
1146 // with the CSE to avoid O(N^2) node blow-up.
1147 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1148 if( p2 ) { // Found CSE
1149 p->destruct(); // Recover useless new node
1150 p = p2; // Use old node
1151 } else {
1152 igvn.register_new_node_with_optimizer(p, old_phi);
1153 }
1154 // Make old Phi refer to new Phi.
1155 old_phi->add_req(p);
1156 // Check for the special case of making the old phi useless and
1157 // disappear it. In JavaGrande I have a case where this useless
1158 // Phi is the loop limit and prevents recognizing a CountedLoop
1159 // which in turn prevents removing an empty loop.
1160 Node *id_old_phi = old_phi->Identity( &igvn );
1161 if( id_old_phi != old_phi ) { // Found a simple identity?
1162 // Note that I cannot call 'replace_node' here, because
1163 // that will yank the edge from old_phi to the Region and
1164 // I'm mid-iteration over the Region's uses.
1165 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1166 Node* use = old_phi->last_out(i);
1167 igvn.rehash_node_delayed(use);
1168 uint uses_found = 0;
1169 for (uint j = 0; j < use->len(); j++) {
1170 if (use->in(j) == old_phi) {
1171 if (j < use->req()) use->set_req (j, id_old_phi);
1172 else use->set_prec(j, id_old_phi);
1173 uses_found++;
1174 }
1175 }
1176 i -= uses_found; // we deleted 1 or more copies of this edge
1177 }
1178 }
1179 igvn._worklist.push(old_phi);
1180 }
1181 }
1182 // Finally clean out the fall-in edges from the RegionNode
1183 for( i = oreq-1; i>0; i-- ) {
1184 if( !phase->is_member( this, _head->in(i) ) ) {
1185 _head->del_req(i);
1186 }
1187 }
1188 igvn.rehash_node_delayed(_head);
1189 // Transform landing pad
1190 igvn.register_new_node_with_optimizer(landing_pad, _head);
1191 // Insert landing pad into the header
1192 _head->add_req(landing_pad);
1193 }
1194
1195 //------------------------------split_outer_loop-------------------------------
1196 // Split out the outermost loop from this shared header.
1197 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1198 PhaseIterGVN &igvn = phase->_igvn;
1199
1200 // Find index of outermost loop; it should also be my tail.
1201 uint outer_idx = 1;
1202 while( _head->in(outer_idx) != _tail ) outer_idx++;
1203
1204 // Make a LoopNode for the outermost loop.
1205 Node *ctl = _head->in(LoopNode::EntryControl);
1206 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1207 outer = igvn.register_new_node_with_optimizer(outer, _head);
1208 phase->set_created_loop_node();
1209
1210 // Outermost loop falls into '_head' loop
1211 _head->set_req(LoopNode::EntryControl, outer);
1212 _head->del_req(outer_idx);
1213 // Split all the Phis up between '_head' loop and 'outer' loop.
1214 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1215 Node *out = _head->fast_out(j);
1216 if( out->is_Phi() ) {
1217 PhiNode *old_phi = out->as_Phi();
1218 assert( old_phi->region() == _head, "" );
1219 Node *phi = PhiNode::make_blank(outer, old_phi);
1220 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1221 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1222 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1223 // Make old Phi point to new Phi on the fall-in path
1224 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1225 old_phi->del_req(outer_idx);
1226 }
1227 }
1228
1229 // Use the new loop head instead of the old shared one
1230 _head = outer;
1231 phase->set_loop(_head, this);
1232 }
1233
1234 //------------------------------fix_parent-------------------------------------
1235 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1236 loop->_parent = parent;
1237 if( loop->_child ) fix_parent( loop->_child, loop );
1238 if( loop->_next ) fix_parent( loop->_next , parent );
1239 }
1240
1241 //------------------------------estimate_path_freq-----------------------------
1242 static float estimate_path_freq( Node *n ) {
1243 // Try to extract some path frequency info
1244 IfNode *iff;
1245 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1246 Opcodes nop = n->Opcode();
1247 if( nop == Opcodes::Op_SafePoint ) { // Skip any safepoint
1248 n = n->in(0);
1249 continue;
1250 }
1251 if( nop == Opcodes::Op_CatchProj ) { // Get count from a prior call
1252 // Assume call does not always throw exceptions: means the call-site
1253 // count is also the frequency of the fall-through path.
1254 assert( n->is_CatchProj(), "" );
1255 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1256 return 0.0f; // Assume call exception path is rare
1257 Node *call = n->in(0)->in(0)->in(0);
1258 assert( call->is_Call(), "expect a call here" );
1259 const JVMState *jvms = ((CallNode*)call)->jvms();
1260 ciMethodData* methodData = jvms->method()->method_data();
1261 if (!methodData->is_mature()) return 0.0f; // No call-site data
1262 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1263 if ((data == NULL) || !data->is_CounterData()) {
1264 // no call profile available, try call's control input
1265 n = n->in(0);
1266 continue;
1267 }
1268 return data->as_CounterData()->count()/FreqCountInvocations;
1269 }
1270 // See if there's a gating IF test
1271 Node *n_c = n->in(0);
1272 if( !n_c->is_If() ) break; // No estimate available
1273 iff = n_c->as_If();
1274 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1275 // Compute how much count comes on this path
1276 return ((nop == Opcodes::Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1277 // Have no count info. Skip dull uncommon-trap like branches.
1278 if( (nop == Opcodes::Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1279 (nop == Opcodes::Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1280 break;
1281 // Skip through never-taken branch; look for a real loop exit.
1282 n = iff->in(0);
1283 }
1284 return 0.0f; // No estimate available
1285 }
1286
1287 //------------------------------merge_many_backedges---------------------------
1288 // Merge all the backedges from the shared header into a private Region.
1289 // Feed that region as the one backedge to this loop.
1290 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1291 uint i;
1292
1293 // Scan for the top 2 hottest backedges
1294 float hotcnt = 0.0f;
1295 float warmcnt = 0.0f;
1296 uint hot_idx = 0;
1297 // Loop starts at 2 because slot 1 is the fall-in path
1298 for( i = 2; i < _head->req(); i++ ) {
1299 float cnt = estimate_path_freq(_head->in(i));
1300 if( cnt > hotcnt ) { // Grab hottest path
1301 warmcnt = hotcnt;
1302 hotcnt = cnt;
1303 hot_idx = i;
1304 } else if( cnt > warmcnt ) { // And 2nd hottest path
1305 warmcnt = cnt;
1306 }
1307 }
1308
1309 // See if the hottest backedge is worthy of being an inner loop
1310 // by being much hotter than the next hottest backedge.
1311 if( hotcnt <= 0.0001 ||
1312 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1313
1314 // Peel out the backedges into a private merge point; peel
1315 // them all except optionally hot_idx.
1316 PhaseIterGVN &igvn = phase->_igvn;
1317
1318 Node *hot_tail = NULL;
1319 // Make a Region for the merge point
1320 Node *r = new RegionNode(1);
1321 for( i = 2; i < _head->req(); i++ ) {
1322 if( i != hot_idx )
1323 r->add_req( _head->in(i) );
1324 else hot_tail = _head->in(i);
1325 }
1326 igvn.register_new_node_with_optimizer(r, _head);
1327 // Plug region into end of loop _head, followed by hot_tail
1328 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1329 igvn.replace_input_of(_head, 2, r);
1330 if( hot_idx ) _head->add_req(hot_tail);
1331
1332 // Split all the Phis up between '_head' loop and the Region 'r'
1333 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1334 Node *out = _head->fast_out(j);
1335 if( out->is_Phi() ) {
1336 PhiNode* n = out->as_Phi();
1337 igvn.hash_delete(n); // Delete from hash before hacking edges
1338 Node *hot_phi = NULL;
1339 Node *phi = new PhiNode(r, n->type(), n->adr_type());
1340 // Check all inputs for the ones to peel out
1341 uint j = 1;
1342 for( uint i = 2; i < n->req(); i++ ) {
1343 if( i != hot_idx )
1344 phi->set_req( j++, n->in(i) );
1345 else hot_phi = n->in(i);
1346 }
1347 // Register the phi but do not transform until whole place transforms
1348 igvn.register_new_node_with_optimizer(phi, n);
1349 // Add the merge phi to the old Phi
1350 while( n->req() > 3 ) n->del_req( n->req()-1 );
1351 igvn.replace_input_of(n, 2, phi);
1352 if( hot_idx ) n->add_req(hot_phi);
1353 }
1354 }
1355
1356
1357 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1358 // of self loop tree. Turn self into a loop headed by _head and with
1359 // tail being the new merge point.
1360 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1361 phase->set_loop(_tail,ilt); // Adjust tail
1362 _tail = r; // Self's tail is new merge point
1363 phase->set_loop(r,this);
1364 ilt->_child = _child; // New guy has my children
1365 _child = ilt; // Self has new guy as only child
1366 ilt->_parent = this; // new guy has self for parent
1367 ilt->_nest = _nest; // Same nesting depth (for now)
1368
1369 // Starting with 'ilt', look for child loop trees using the same shared
1370 // header. Flatten these out; they will no longer be loops in the end.
1371 IdealLoopTree **pilt = &_child;
1372 while( ilt ) {
1373 if( ilt->_head == _head ) {
1374 uint i;
1375 for( i = 2; i < _head->req(); i++ )
1376 if( _head->in(i) == ilt->_tail )
1377 break; // Still a loop
1378 if( i == _head->req() ) { // No longer a loop
1379 // Flatten ilt. Hang ilt's "_next" list from the end of
1380 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1381 IdealLoopTree **cp = &ilt->_child;
1382 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1383 *cp = ilt->_next; // Hang next list at end of child list
1384 *pilt = ilt->_child; // Move child up to replace ilt
1385 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1386 ilt = ilt->_child; // Repeat using new ilt
1387 continue; // do not advance over ilt->_child
1388 }
1389 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1390 phase->set_loop(_head,ilt);
1391 }
1392 pilt = &ilt->_child; // Advance to next
1393 ilt = *pilt;
1394 }
1395
1396 if( _child ) fix_parent( _child, this );
1397 }
1398
1399 //------------------------------beautify_loops---------------------------------
1400 // Split shared headers and insert loop landing pads.
1401 // Insert a LoopNode to replace the RegionNode.
1402 // Return TRUE if loop tree is structurally changed.
1403 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1404 bool result = false;
1405 // Cache parts in locals for easy
1406 PhaseIterGVN &igvn = phase->_igvn;
1407
1408 igvn.hash_delete(_head); // Yank from hash before hacking edges
1409
1410 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1411 int fall_in_cnt = 0;
1412 for( uint i = 1; i < _head->req(); i++ )
1413 if( !phase->is_member( this, _head->in(i) ) )
1414 fall_in_cnt++;
1415 assert( fall_in_cnt, "at least 1 fall-in path" );
1416 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1417 split_fall_in( phase, fall_in_cnt );
1418
1419 // Swap inputs to the _head and all Phis to move the fall-in edge to
1420 // the left.
1421 fall_in_cnt = 1;
1422 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1423 fall_in_cnt++;
1424 if( fall_in_cnt > 1 ) {
1425 // Since I am just swapping inputs I do not need to update def-use info
1426 Node *tmp = _head->in(1);
1427 igvn.rehash_node_delayed(_head);
1428 _head->set_req( 1, _head->in(fall_in_cnt) );
1429 _head->set_req( fall_in_cnt, tmp );
1430 // Swap also all Phis
1431 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1432 Node* phi = _head->fast_out(i);
1433 if( phi->is_Phi() ) {
1434 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
1435 tmp = phi->in(1);
1436 phi->set_req( 1, phi->in(fall_in_cnt) );
1437 phi->set_req( fall_in_cnt, tmp );
1438 }
1439 }
1440 }
1441 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1442 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1443
1444 // If I am a shared header (multiple backedges), peel off the many
1445 // backedges into a private merge point and use the merge point as
1446 // the one true backedge.
1447 if( _head->req() > 3 ) {
1448 // Merge the many backedges into a single backedge but leave
1449 // the hottest backedge as separate edge for the following peel.
1450 merge_many_backedges( phase );
1451 result = true;
1452 }
1453
1454 // If I have one hot backedge, peel off myself loop.
1455 // I better be the outermost loop.
1456 if (_head->req() > 3 && !_irreducible) {
1457 split_outer_loop( phase );
1458 result = true;
1459
1460 } else if (!_head->is_Loop() && !_irreducible) {
1461 // Make a new LoopNode to replace the old loop head
1462 Node *l = new LoopNode( _head->in(1), _head->in(2) );
1463 l = igvn.register_new_node_with_optimizer(l, _head);
1464 phase->set_created_loop_node();
1465 // Go ahead and replace _head
1466 phase->_igvn.replace_node( _head, l );
1467 _head = l;
1468 phase->set_loop(_head, this);
1469 }
1470
1471 // Now recursively beautify nested loops
1472 if( _child ) result |= _child->beautify_loops( phase );
1473 if( _next ) result |= _next ->beautify_loops( phase );
1474 return result;
1475 }
1476
1477 //------------------------------allpaths_check_safepts----------------------------
1478 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1479 // encountered. Helper for check_safepts.
1480 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1481 assert(stack.size() == 0, "empty stack");
1482 stack.push(_tail);
1483 visited.Clear();
1484 visited.set(_tail->_idx);
1485 while (stack.size() > 0) {
1486 Node* n = stack.pop();
1487 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1488 // Terminate this path
1489 } else if (n->Opcode() == Opcodes::Op_SafePoint) {
1490 if (_phase->get_loop(n) != this) {
1491 if (_required_safept == NULL) _required_safept = new Node_List();
1492 _required_safept->push(n); // save the one closest to the tail
1493 }
1494 // Terminate this path
1495 } else {
1496 uint start = n->is_Region() ? 1 : 0;
1497 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1498 for (uint i = start; i < end; i++) {
1499 Node* in = n->in(i);
1500 assert(in->is_CFG(), "must be");
1501 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1502 stack.push(in);
1503 }
1504 }
1505 }
1506 }
1507 }
1508
1509 //------------------------------check_safepts----------------------------
1510 // Given dominators, try to find loops with calls that must always be
1511 // executed (call dominates loop tail). These loops do not need non-call
1512 // safepoints (ncsfpt).
1513 //
1514 // A complication is that a safepoint in a inner loop may be needed
1515 // by an outer loop. In the following, the inner loop sees it has a
1516 // call (block 3) on every path from the head (block 2) to the
1517 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1518 // in block 2, _but_ this leaves the outer loop without a safepoint.
1519 //
1520 // entry 0
1521 // |
1522 // v
1523 // outer 1,2 +->1
1524 // | |
1525 // | v
1526 // | 2<---+ ncsfpt in 2
1527 // |_/|\ |
1528 // | v |
1529 // inner 2,3 / 3 | call in 3
1530 // / | |
1531 // v +--+
1532 // exit 4
1533 //
1534 //
1535 // This method creates a list (_required_safept) of ncsfpt nodes that must
1536 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1537 // is first looked for in the lists for the outer loops of the current loop.
1538 //
1539 // The insights into the problem:
1540 // A) counted loops are okay
1541 // B) innermost loops are okay (only an inner loop can delete
1542 // a ncsfpt needed by an outer loop)
1543 // C) a loop is immune from an inner loop deleting a safepoint
1544 // if the loop has a call on the idom-path
1545 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1546 // idom-path that is not in a nested loop
1547 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1548 // loop needs to be prevented from deletion by an inner loop
1549 //
1550 // There are two analyses:
1551 // 1) The first, and cheaper one, scans the loop body from
1552 // tail to head following the idom (immediate dominator)
1553 // chain, looking for the cases (C,D,E) above.
1554 // Since inner loops are scanned before outer loops, there is summary
1555 // information about inner loops. Inner loops can be skipped over
1556 // when the tail of an inner loop is encountered.
1557 //
1558 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1559 // the idom path (which is rare), scans all predecessor control paths
1560 // from the tail to the head, terminating a path when a call or sfpt
1561 // is encountered, to find the ncsfpt's that are closest to the tail.
1562 //
1563 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1564 // Bottom up traversal
1565 IdealLoopTree* ch = _child;
1566 if (_child) _child->check_safepts(visited, stack);
1567 if (_next) _next ->check_safepts(visited, stack);
1568
1569 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1570 bool has_call = false; // call on dom-path
1571 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1572 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1573 // Scan the dom-path nodes from tail to head
1574 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1575 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1576 has_call = true;
1577 _has_sfpt = 1; // Then no need for a safept!
1578 break;
1579 } else if (n->Opcode() == Opcodes::Op_SafePoint) {
1580 if (_phase->get_loop(n) == this) {
1581 has_local_ncsfpt = true;
1582 break;
1583 }
1584 if (nonlocal_ncsfpt == NULL) {
1585 nonlocal_ncsfpt = n; // save the one closest to the tail
1586 }
1587 } else {
1588 IdealLoopTree* nlpt = _phase->get_loop(n);
1589 if (this != nlpt) {
1590 // If at an inner loop tail, see if the inner loop has already
1591 // recorded seeing a call on the dom-path (and stop.) If not,
1592 // jump to the head of the inner loop.
1593 assert(is_member(nlpt), "nested loop");
1594 Node* tail = nlpt->_tail;
1595 if (tail->in(0)->is_If()) tail = tail->in(0);
1596 if (n == tail) {
1597 // If inner loop has call on dom-path, so does outer loop
1598 if (nlpt->_has_sfpt) {
1599 has_call = true;
1600 _has_sfpt = 1;
1601 break;
1602 }
1603 // Skip to head of inner loop
1604 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1605 n = nlpt->_head;
1606 }
1607 }
1608 }
1609 }
1610 // Record safept's that this loop needs preserved when an
1611 // inner loop attempts to delete it's safepoints.
1612 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1613 if (nonlocal_ncsfpt != NULL) {
1614 if (_required_safept == NULL) _required_safept = new Node_List();
1615 _required_safept->push(nonlocal_ncsfpt);
1616 } else {
1617 // Failed to find a suitable safept on the dom-path. Now use
1618 // an all paths walk from tail to head, looking for safepoints to preserve.
1619 allpaths_check_safepts(visited, stack);
1620 }
1621 }
1622 }
1623 }
1624
1625 //---------------------------is_deleteable_safept----------------------------
1626 // Is safept not required by an outer loop?
1627 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1628 assert(sfpt->Opcode() == Opcodes::Op_SafePoint, "");
1629 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1630 while (lp != NULL) {
1631 Node_List* sfpts = lp->_required_safept;
1632 if (sfpts != NULL) {
1633 for (uint i = 0; i < sfpts->size(); i++) {
1634 if (sfpt == sfpts->at(i))
1635 return false;
1636 }
1637 }
1638 lp = lp->_parent;
1639 }
1640 return true;
1641 }
1642
1643 //---------------------------replace_parallel_iv-------------------------------
1644 // Replace parallel induction variable (parallel to trip counter)
1645 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1646 assert(loop->_head->is_CountedLoop(), "");
1647 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1648 if (!cl->is_valid_counted_loop())
1649 return; // skip malformed counted loop
1650 Node *incr = cl->incr();
1651 if (incr == NULL)
1652 return; // Dead loop?
1653 Node *init = cl->init_trip();
1654 Node *phi = cl->phi();
1655 int stride_con = cl->stride_con();
1656
1657 // Visit all children, looking for Phis
1658 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1659 Node *out = cl->out(i);
1660 // Look for other phis (secondary IVs). Skip dead ones
1661 if (!out->is_Phi() || out == phi || !has_node(out))
1662 continue;
1663 PhiNode* phi2 = out->as_Phi();
1664 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1665 // Look for induction variables of the form: X += constant
1666 if (phi2->region() != loop->_head ||
1667 incr2->req() != 3 ||
1668 incr2->in(1) != phi2 ||
1669 incr2 == incr ||
1670 incr2->Opcode() != Opcodes::Op_AddI ||
1671 !incr2->in(2)->is_Con())
1672 continue;
1673
1674 // Check for parallel induction variable (parallel to trip counter)
1675 // via an affine function. In particular, count-down loops with
1676 // count-up array indices are common. We only RCE references off
1677 // the trip-counter, so we need to convert all these to trip-counter
1678 // expressions.
1679 Node *init2 = phi2->in( LoopNode::EntryControl );
1680 int stride_con2 = incr2->in(2)->get_int();
1681
1682 // The general case here gets a little tricky. We want to find the
1683 // GCD of all possible parallel IV's and make a new IV using this
1684 // GCD for the loop. Then all possible IVs are simple multiples of
1685 // the GCD. In practice, this will cover very few extra loops.
1686 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1687 // where +/-1 is the common case, but other integer multiples are
1688 // also easy to handle.
1689 int ratio_con = stride_con2/stride_con;
1690
1691 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1692 #ifndef PRODUCT
1693 if (TraceLoopOpts) {
1694 tty->print("Parallel IV: %d ", phi2->_idx);
1695 loop->dump_head();
1696 }
1697 #endif
1698 // Convert to using the trip counter. The parallel induction
1699 // variable differs from the trip counter by a loop-invariant
1700 // amount, the difference between their respective initial values.
1701 // It is scaled by the 'ratio_con'.
1702 Node* ratio = _igvn.intcon(ratio_con);
1703 set_ctrl(ratio, C->root());
1704 Node* ratio_init = new MulINode(init, ratio);
1705 _igvn.register_new_node_with_optimizer(ratio_init, init);
1706 set_early_ctrl(ratio_init);
1707 Node* diff = new SubINode(init2, ratio_init);
1708 _igvn.register_new_node_with_optimizer(diff, init2);
1709 set_early_ctrl(diff);
1710 Node* ratio_idx = new MulINode(phi, ratio);
1711 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1712 set_ctrl(ratio_idx, cl);
1713 Node* add = new AddINode(ratio_idx, diff);
1714 _igvn.register_new_node_with_optimizer(add);
1715 set_ctrl(add, cl);
1716 _igvn.replace_node( phi2, add );
1717 // Sometimes an induction variable is unused
1718 if (add->outcnt() == 0) {
1719 _igvn.remove_dead_node(add);
1720 }
1721 --i; // deleted this phi; rescan starting with next position
1722 continue;
1723 }
1724 }
1725 }
1726
1727 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
1728 Node* keep = NULL;
1729 if (keep_one) {
1730 // Look for a safepoint on the idom-path.
1731 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
1732 if (i->Opcode() == Opcodes::Op_SafePoint && phase->get_loop(i) == this) {
1733 keep = i;
1734 break; // Found one
1735 }
1736 }
1737 }
1738
1739 // Don't remove any safepoints if it is requested to keep a single safepoint and
1740 // no safepoint was found on idom-path. It is not safe to remove any safepoint
1741 // in this case since there's no safepoint dominating all paths in the loop body.
1742 bool prune = !keep_one || keep != NULL;
1743
1744 // Delete other safepoints in this loop.
1745 Node_List* sfpts = _safepts;
1746 if (prune && sfpts != NULL) {
1747 assert(keep == NULL || keep->Opcode() == Opcodes::Op_SafePoint, "not safepoint");
1748 for (uint i = 0; i < sfpts->size(); i++) {
1749 Node* n = sfpts->at(i);
1750 assert(phase->get_loop(n) == this, "");
1751 if (n != keep && phase->is_deleteable_safept(n)) {
1752 phase->lazy_replace(n, n->in(TypeFunc::Control));
1753 }
1754 }
1755 }
1756 }
1757
1758 //------------------------------counted_loop-----------------------------------
1759 // Convert to counted loops where possible
1760 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1761
1762 // For grins, set the inner-loop flag here
1763 if (!_child) {
1764 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1765 }
1766
1767 if (_head->is_CountedLoop() ||
1768 phase->is_counted_loop(_head, this)) {
1769
1770 if (!UseCountedLoopSafepoints) {
1771 // Indicate we do not need a safepoint here
1772 _has_sfpt = 1;
1773 }
1774
1775 // Remove safepoints
1776 bool keep_one_sfpt = !(_has_call || _has_sfpt);
1777 remove_safepoints(phase, keep_one_sfpt);
1778
1779 // Look for induction variables
1780 phase->replace_parallel_iv(this);
1781
1782 } else if (_parent != NULL && !_irreducible) {
1783 // Not a counted loop. Keep one safepoint.
1784 bool keep_one_sfpt = true;
1785 remove_safepoints(phase, keep_one_sfpt);
1786 }
1787
1788 // Recursively
1789 if (_child) _child->counted_loop( phase );
1790 if (_next) _next ->counted_loop( phase );
1791 }
1792
1793 #ifndef PRODUCT
1794 //------------------------------dump_head--------------------------------------
1795 // Dump 1 liner for loop header info
1796 void IdealLoopTree::dump_head( ) const {
1797 for (uint i=0; i<_nest; i++)
1798 tty->print(" ");
1799 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1800 if (_irreducible) tty->print(" IRREDUCIBLE");
1801 Node* entry = _head->in(LoopNode::EntryControl);
1802 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1803 if (predicate != NULL ) {
1804 tty->print(" limit_check");
1805 entry = entry->in(0)->in(0);
1806 }
1807 if (UseLoopPredicate) {
1808 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1809 if (entry != NULL) {
1810 tty->print(" predicated");
1811 }
1812 }
1813 if (_head->is_CountedLoop()) {
1814 CountedLoopNode *cl = _head->as_CountedLoop();
1815 tty->print(" counted");
1816
1817 Node* init_n = cl->init_trip();
1818 if (init_n != NULL && init_n->is_Con())
1819 tty->print(" [%d,", cl->init_trip()->get_int());
1820 else
1821 tty->print(" [int,");
1822 Node* limit_n = cl->limit();
1823 if (limit_n != NULL && limit_n->is_Con())
1824 tty->print("%d),", cl->limit()->get_int());
1825 else
1826 tty->print("int),");
1827 int stride_con = cl->stride_con();
1828 if (stride_con > 0) tty->print("+");
1829 tty->print("%d", stride_con);
1830
1831 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
1832
1833 if (cl->is_pre_loop ()) tty->print(" pre" );
1834 if (cl->is_main_loop()) tty->print(" main");
1835 if (cl->is_post_loop()) tty->print(" post");
1836 if (cl->is_vectorized_loop()) tty->print(" vector");
1837 if (cl->range_checks_present()) tty->print(" rc ");
1838 if (cl->is_multiversioned()) tty->print(" multi ");
1839 }
1840 if (_has_call) tty->print(" has_call");
1841 if (_has_sfpt) tty->print(" has_sfpt");
1842 if (_rce_candidate) tty->print(" rce");
1843 if (_safepts != NULL && _safepts->size() > 0) {
1844 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
1845 }
1846 if (_required_safept != NULL && _required_safept->size() > 0) {
1847 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
1848 }
1849 tty->cr();
1850 }
1851
1852 //------------------------------dump-------------------------------------------
1853 // Dump loops by loop tree
1854 void IdealLoopTree::dump( ) const {
1855 dump_head();
1856 if (_child) _child->dump();
1857 if (_next) _next ->dump();
1858 }
1859
1860 #endif
1861
1862 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1863 if (loop == root) {
1864 if (loop->_child != NULL) {
1865 log->begin_head("loop_tree");
1866 log->end_head();
1867 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1868 log->tail("loop_tree");
1869 assert(loop->_next == NULL, "what?");
1870 }
1871 } else {
1872 Node* head = loop->_head;
1873 log->begin_head("loop");
1874 log->print(" idx='%d' ", head->_idx);
1875 if (loop->_irreducible) log->print("irreducible='1' ");
1876 if (head->is_Loop()) {
1877 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1878 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1879 }
1880 if (head->is_CountedLoop()) {
1881 CountedLoopNode* cl = head->as_CountedLoop();
1882 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1883 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1884 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1885 }
1886 log->end_head();
1887 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1888 log->tail("loop");
1889 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1890 }
1891 }
1892
1893 //---------------------collect_potentially_useful_predicates-----------------------
1894 // Helper function to collect potentially useful predicates to prevent them from
1895 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1896 void PhaseIdealLoop::collect_potentially_useful_predicates(
1897 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1898 if (loop->_child) { // child
1899 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1900 }
1901
1902 // self (only loops that we can apply loop predication may use their predicates)
1903 if (loop->_head->is_Loop() &&
1904 !loop->_irreducible &&
1905 !loop->tail()->is_top()) {
1906 LoopNode* lpn = loop->_head->as_Loop();
1907 Node* entry = lpn->in(LoopNode::EntryControl);
1908 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1909 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1910 assert(entry->in(0)->in(1)->in(1)->Opcode() == Opcodes::Op_Opaque1, "must be");
1911 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1912 entry = entry->in(0)->in(0);
1913 }
1914 predicate_proj = find_predicate(entry); // Predicate
1915 if (predicate_proj != NULL ) {
1916 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1917 }
1918 }
1919
1920 if (loop->_next) { // sibling
1921 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1922 }
1923 }
1924
1925 //------------------------eliminate_useless_predicates-----------------------------
1926 // Eliminate all inserted predicates if they could not be used by loop predication.
1927 // Note: it will also eliminates loop limits check predicate since it also uses
1928 // Opaque1 node (see Parse::add_predicate()).
1929 void PhaseIdealLoop::eliminate_useless_predicates() {
1930 if (C->predicate_count() == 0)
1931 return; // no predicate left
1932
1933 Unique_Node_List useful_predicates; // to store useful predicates
1934 if (C->has_loops()) {
1935 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1936 }
1937
1938 for (int i = C->predicate_count(); i > 0; i--) {
1939 Node * n = C->predicate_opaque1_node(i-1);
1940 assert(n->Opcode() == Opcodes::Op_Opaque1, "must be");
1941 if (!useful_predicates.member(n)) { // not in the useful list
1942 _igvn.replace_node(n, n->in(1));
1943 }
1944 }
1945 }
1946
1947 //------------------------process_expensive_nodes-----------------------------
1948 // Expensive nodes have their control input set to prevent the GVN
1949 // from commoning them and as a result forcing the resulting node to
1950 // be in a more frequent path. Use CFG information here, to change the
1951 // control inputs so that some expensive nodes can be commoned while
1952 // not executed more frequently.
1953 bool PhaseIdealLoop::process_expensive_nodes() {
1954 assert(OptimizeExpensiveOps, "optimization off?");
1955
1956 // Sort nodes to bring similar nodes together
1957 C->sort_expensive_nodes();
1958
1959 bool progress = false;
1960
1961 for (int i = 0; i < C->expensive_count(); ) {
1962 Node* n = C->expensive_node(i);
1963 int start = i;
1964 // Find nodes similar to n
1965 i++;
1966 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
1967 int end = i;
1968 // And compare them two by two
1969 for (int j = start; j < end; j++) {
1970 Node* n1 = C->expensive_node(j);
1971 if (is_node_unreachable(n1)) {
1972 continue;
1973 }
1974 for (int k = j+1; k < end; k++) {
1975 Node* n2 = C->expensive_node(k);
1976 if (is_node_unreachable(n2)) {
1977 continue;
1978 }
1979
1980 assert(n1 != n2, "should be pair of nodes");
1981
1982 Node* c1 = n1->in(0);
1983 Node* c2 = n2->in(0);
1984
1985 Node* parent_c1 = c1;
1986 Node* parent_c2 = c2;
1987
1988 // The call to get_early_ctrl_for_expensive() moves the
1989 // expensive nodes up but stops at loops that are in a if
1990 // branch. See whether we can exit the loop and move above the
1991 // If.
1992 if (c1->is_Loop()) {
1993 parent_c1 = c1->in(1);
1994 }
1995 if (c2->is_Loop()) {
1996 parent_c2 = c2->in(1);
1997 }
1998
1999 if (parent_c1 == parent_c2) {
2000 _igvn._worklist.push(n1);
2001 _igvn._worklist.push(n2);
2002 continue;
2003 }
2004
2005 // Look for identical expensive node up the dominator chain.
2006 if (is_dominator(c1, c2)) {
2007 c2 = c1;
2008 } else if (is_dominator(c2, c1)) {
2009 c1 = c2;
2010 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2011 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2012 // Both branches have the same expensive node so move it up
2013 // before the if.
2014 c1 = c2 = idom(parent_c1->in(0));
2015 }
2016 // Do the actual moves
2017 if (n1->in(0) != c1) {
2018 _igvn.hash_delete(n1);
2019 n1->set_req(0, c1);
2020 _igvn.hash_insert(n1);
2021 _igvn._worklist.push(n1);
2022 progress = true;
2023 }
2024 if (n2->in(0) != c2) {
2025 _igvn.hash_delete(n2);
2026 n2->set_req(0, c2);
2027 _igvn.hash_insert(n2);
2028 _igvn._worklist.push(n2);
2029 progress = true;
2030 }
2031 }
2032 }
2033 }
2034
2035 return progress;
2036 }
2037
2038
2039 //=============================================================================
2040 //----------------------------build_and_optimize-------------------------------
2041 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2042 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2043 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
2044 ResourceMark rm;
2045
2046 int old_progress = C->major_progress();
2047 uint orig_worklist_size = _igvn._worklist.size();
2048
2049 // Reset major-progress flag for the driver's heuristics
2050 C->clear_major_progress();
2051
2052 #ifndef PRODUCT
2053 // Capture for later assert
2054 uint unique = C->unique();
2055 _loop_invokes++;
2056 _loop_work += unique;
2057 #endif
2058
2059 // True if the method has at least 1 irreducible loop
2060 _has_irreducible_loops = false;
2061
2062 _created_loop_node = false;
2063
2064 Arena *a = Thread::current()->resource_area();
2065 VectorSet visited(a);
2066 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2067 _nodes.map(C->unique(), NULL);
2068 memset(_nodes.adr(), 0, wordSize * C->unique());
2069
2070 // Pre-build the top-level outermost loop tree entry
2071 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2072 // Do not need a safepoint at the top level
2073 _ltree_root->_has_sfpt = 1;
2074
2075 // Initialize Dominators.
2076 // Checked in clone_loop_predicate() during beautify_loops().
2077 _idom_size = 0;
2078 _idom = NULL;
2079 _dom_depth = NULL;
2080 _dom_stk = NULL;
2081
2082 // Empty pre-order array
2083 allocate_preorders();
2084
2085 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2086 // IdealLoopTree entries. Data nodes are NOT walked.
2087 build_loop_tree();
2088 // Check for bailout, and return
2089 if (C->failing()) {
2090 return;
2091 }
2092
2093 // No loops after all
2094 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2095
2096 // There should always be an outer loop containing the Root and Return nodes.
2097 // If not, we have a degenerate empty program. Bail out in this case.
2098 if (!has_node(C->root())) {
2099 if (!_verify_only) {
2100 C->clear_major_progress();
2101 C->record_method_not_compilable("empty program detected during loop optimization");
2102 }
2103 return;
2104 }
2105
2106 // Nothing to do, so get out
2107 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
2108 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2109 if (stop_early && !do_expensive_nodes) {
2110 _igvn.optimize(); // Cleanup NeverBranches
2111 return;
2112 }
2113
2114 // Set loop nesting depth
2115 _ltree_root->set_nest( 0 );
2116
2117 // Split shared headers and insert loop landing pads.
2118 // Do not bother doing this on the Root loop of course.
2119 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2120 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2121 if( _ltree_root->_child->beautify_loops( this ) ) {
2122 // Re-build loop tree!
2123 _ltree_root->_child = NULL;
2124 _nodes.clear();
2125 reallocate_preorders();
2126 build_loop_tree();
2127 // Check for bailout, and return
2128 if (C->failing()) {
2129 return;
2130 }
2131 // Reset loop nesting depth
2132 _ltree_root->set_nest( 0 );
2133
2134 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2135 }
2136 }
2137
2138 // Build Dominators for elision of NULL checks & loop finding.
2139 // Since nodes do not have a slot for immediate dominator, make
2140 // a persistent side array for that info indexed on node->_idx.
2141 _idom_size = C->unique();
2142 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2143 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2144 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2145 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2146
2147 Dominators();
2148
2149 if (!_verify_only) {
2150 // As a side effect, Dominators removed any unreachable CFG paths
2151 // into RegionNodes. It doesn't do this test against Root, so
2152 // we do it here.
2153 for( uint i = 1; i < C->root()->req(); i++ ) {
2154 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2155 _igvn.delete_input_of(C->root(), i);
2156 i--; // Rerun same iteration on compressed edges
2157 }
2158 }
2159
2160 // Given dominators, try to find inner loops with calls that must
2161 // always be executed (call dominates loop tail). These loops do
2162 // not need a separate safepoint.
2163 Node_List cisstack(a);
2164 _ltree_root->check_safepts(visited, cisstack);
2165 }
2166
2167 // Walk the DATA nodes and place into loops. Find earliest control
2168 // node. For CFG nodes, the _nodes array starts out and remains
2169 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2170 // _nodes array holds the earliest legal controlling CFG node.
2171
2172 // Allocate stack with enough space to avoid frequent realloc
2173 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2174 Node_Stack nstack( a, stack_size );
2175
2176 visited.Clear();
2177 Node_List worklist(a);
2178 // Don't need C->root() on worklist since
2179 // it will be processed among C->top() inputs
2180 worklist.push( C->top() );
2181 visited.set( C->top()->_idx ); // Set C->top() as visited now
2182 build_loop_early( visited, worklist, nstack );
2183
2184 // Given early legal placement, try finding counted loops. This placement
2185 // is good enough to discover most loop invariants.
2186 if( !_verify_me && !_verify_only )
2187 _ltree_root->counted_loop( this );
2188
2189 // Find latest loop placement. Find ideal loop placement.
2190 visited.Clear();
2191 init_dom_lca_tags();
2192 // Need C->root() on worklist when processing outs
2193 worklist.push( C->root() );
2194 NOT_PRODUCT( C->verify_graph_edges(); )
2195 worklist.push( C->top() );
2196 build_loop_late( visited, worklist, nstack );
2197
2198 if (_verify_only) {
2199 // restore major progress flag
2200 for (int i = 0; i < old_progress; i++)
2201 C->set_major_progress();
2202 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2203 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2204 return;
2205 }
2206
2207 // clear out the dead code after build_loop_late
2208 while (_deadlist.size()) {
2209 _igvn.remove_globally_dead_node(_deadlist.pop());
2210 }
2211
2212 if (stop_early) {
2213 assert(do_expensive_nodes, "why are we here?");
2214 if (process_expensive_nodes()) {
2215 // If we made some progress when processing expensive nodes then
2216 // the IGVN may modify the graph in a way that will allow us to
2217 // make some more progress: we need to try processing expensive
2218 // nodes again.
2219 C->set_major_progress();
2220 }
2221 _igvn.optimize();
2222 return;
2223 }
2224
2225 // Some parser-inserted loop predicates could never be used by loop
2226 // predication or they were moved away from loop during some optimizations.
2227 // For example, peeling. Eliminate them before next loop optimizations.
2228 eliminate_useless_predicates();
2229
2230 #ifndef PRODUCT
2231 C->verify_graph_edges();
2232 if (_verify_me) { // Nested verify pass?
2233 // Check to see if the verify mode is broken
2234 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2235 return;
2236 }
2237 if(VerifyLoopOptimizations) verify();
2238 if(TraceLoopOpts && C->has_loops()) {
2239 _ltree_root->dump();
2240 }
2241 #endif
2242
2243 if (skip_loop_opts) {
2244 // restore major progress flag
2245 for (int i = 0; i < old_progress; i++) {
2246 C->set_major_progress();
2247 }
2248
2249 // Cleanup any modified bits
2250 _igvn.optimize();
2251
2252 if (C->log() != NULL) {
2253 log_loop_tree(_ltree_root, _ltree_root, C->log());
2254 }
2255 return;
2256 }
2257
2258 if (ReassociateInvariants) {
2259 // Reassociate invariants and prep for split_thru_phi
2260 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2261 IdealLoopTree* lpt = iter.current();
2262 bool is_counted = lpt->is_counted();
2263 if (!is_counted || !lpt->is_inner()) continue;
2264
2265 // check for vectorized loops, any reassociation of invariants was already done
2266 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2267
2268 lpt->reassociate_invariants(this);
2269
2270 // Because RCE opportunities can be masked by split_thru_phi,
2271 // look for RCE candidates and inhibit split_thru_phi
2272 // on just their loop-phi's for this pass of loop opts
2273 if (SplitIfBlocks && do_split_ifs) {
2274 if (lpt->policy_range_check(this)) {
2275 lpt->_rce_candidate = 1; // = true
2276 }
2277 }
2278 }
2279 }
2280
2281 // Check for aggressive application of split-if and other transforms
2282 // that require basic-block info (like cloning through Phi's)
2283 if( SplitIfBlocks && do_split_ifs ) {
2284 visited.Clear();
2285 split_if_with_blocks( visited, nstack );
2286 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2287 }
2288
2289 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2290 C->set_major_progress();
2291 }
2292
2293 // Perform loop predication before iteration splitting
2294 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2295 _ltree_root->_child->loop_predication(this);
2296 }
2297
2298 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2299 if (do_intrinsify_fill()) {
2300 C->set_major_progress();
2301 }
2302 }
2303
2304 // Perform iteration-splitting on inner loops. Split iterations to avoid
2305 // range checks or one-shot null checks.
2306
2307 // If split-if's didn't hack the graph too bad (no CFG changes)
2308 // then do loop opts.
2309 if (C->has_loops() && !C->major_progress()) {
2310 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2311 _ltree_root->_child->iteration_split( this, worklist );
2312 // No verify after peeling! GCM has hoisted code out of the loop.
2313 // After peeling, the hoisted code could sink inside the peeled area.
2314 // The peeling code does not try to recompute the best location for
2315 // all the code before the peeled area, so the verify pass will always
2316 // complain about it.
2317 }
2318 // Do verify graph edges in any case
2319 NOT_PRODUCT( C->verify_graph_edges(); );
2320
2321 if (!do_split_ifs) {
2322 // We saw major progress in Split-If to get here. We forced a
2323 // pass with unrolling and not split-if, however more split-if's
2324 // might make progress. If the unrolling didn't make progress
2325 // then the major-progress flag got cleared and we won't try
2326 // another round of Split-If. In particular the ever-common
2327 // instance-of/check-cast pattern requires at least 2 rounds of
2328 // Split-If to clear out.
2329 C->set_major_progress();
2330 }
2331
2332 // Repeat loop optimizations if new loops were seen
2333 if (created_loop_node()) {
2334 C->set_major_progress();
2335 }
2336
2337 // Keep loop predicates and perform optimizations with them
2338 // until no more loop optimizations could be done.
2339 // After that switch predicates off and do more loop optimizations.
2340 if (!C->major_progress() && (C->predicate_count() > 0)) {
2341 C->cleanup_loop_predicates(_igvn);
2342 if (TraceLoopOpts) {
2343 tty->print_cr("PredicatesOff");
2344 }
2345 C->set_major_progress();
2346 }
2347
2348 // Convert scalar to superword operations at the end of all loop opts.
2349 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2350 // SuperWord transform
2351 SuperWord sw(this);
2352 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2353 IdealLoopTree* lpt = iter.current();
2354 if (lpt->is_counted()) {
2355 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
2356
2357 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
2358 // Check that the rce'd post loop is encountered first, multiversion after all
2359 // major main loop optimization are concluded
2360 if (!C->major_progress()) {
2361 IdealLoopTree *lpt_next = lpt->_next;
2362 if (lpt_next && lpt_next->is_counted()) {
2363 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
2364 has_range_checks(lpt_next);
2365 if (cl->is_post_loop() && cl->range_checks_present()) {
2366 if (!cl->is_multiversioned()) {
2367 if (multi_version_post_loops(lpt, lpt_next) == false) {
2368 // Cause the rce loop to be optimized away if we fail
2369 cl->mark_is_multiversioned();
2370 cl->set_slp_max_unroll(0);
2371 poison_rce_post_loop(lpt);
2372 }
2373 }
2374 }
2375 }
2376 sw.transform_loop(lpt, true);
2377 }
2378 } else if (cl->is_main_loop()) {
2379 sw.transform_loop(lpt, true);
2380 }
2381 }
2382 }
2383 }
2384
2385 // Cleanup any modified bits
2386 _igvn.optimize();
2387
2388 // disable assert until issue with split_flow_path is resolved (6742111)
2389 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2390 // "shouldn't introduce irreducible loops");
2391
2392 if (C->log() != NULL) {
2393 log_loop_tree(_ltree_root, _ltree_root, C->log());
2394 }
2395 }
2396
2397 #ifndef PRODUCT
2398 //------------------------------print_statistics-------------------------------
2399 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2400 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2401 void PhaseIdealLoop::print_statistics() {
2402 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2403 }
2404
2405 //------------------------------verify-----------------------------------------
2406 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2407 static int fail; // debug only, so its multi-thread dont care
2408 void PhaseIdealLoop::verify() const {
2409 int old_progress = C->major_progress();
2410 ResourceMark rm;
2411 PhaseIdealLoop loop_verify( _igvn, this );
2412 VectorSet visited(Thread::current()->resource_area());
2413
2414 fail = 0;
2415 verify_compare( C->root(), &loop_verify, visited );
2416 assert( fail == 0, "verify loops failed" );
2417 // Verify loop structure is the same
2418 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2419 // Reset major-progress. It was cleared by creating a verify version of
2420 // PhaseIdealLoop.
2421 for( int i=0; i<old_progress; i++ )
2422 C->set_major_progress();
2423 }
2424
2425 //------------------------------verify_compare---------------------------------
2426 // Make sure me and the given PhaseIdealLoop agree on key data structures
2427 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2428 if( !n ) return;
2429 if( visited.test_set( n->_idx ) ) return;
2430 if( !_nodes[n->_idx] ) { // Unreachable
2431 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2432 return;
2433 }
2434
2435 uint i;
2436 for( i = 0; i < n->req(); i++ )
2437 verify_compare( n->in(i), loop_verify, visited );
2438
2439 // Check the '_nodes' block/loop structure
2440 i = n->_idx;
2441 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2442 if( _nodes[i] != loop_verify->_nodes[i] &&
2443 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2444 tty->print("Mismatched control setting for: ");
2445 n->dump();
2446 if( fail++ > 10 ) return;
2447 Node *c = get_ctrl_no_update(n);
2448 tty->print("We have it as: ");
2449 if( c->in(0) ) c->dump();
2450 else tty->print_cr("N%d",c->_idx);
2451 tty->print("Verify thinks: ");
2452 if( loop_verify->has_ctrl(n) )
2453 loop_verify->get_ctrl_no_update(n)->dump();
2454 else
2455 loop_verify->get_loop_idx(n)->dump();
2456 tty->cr();
2457 }
2458 } else { // We have a loop
2459 IdealLoopTree *us = get_loop_idx(n);
2460 if( loop_verify->has_ctrl(n) ) {
2461 tty->print("Mismatched loop setting for: ");
2462 n->dump();
2463 if( fail++ > 10 ) return;
2464 tty->print("We have it as: ");
2465 us->dump();
2466 tty->print("Verify thinks: ");
2467 loop_verify->get_ctrl_no_update(n)->dump();
2468 tty->cr();
2469 } else if (!C->major_progress()) {
2470 // Loop selection can be messed up if we did a major progress
2471 // operation, like split-if. Do not verify in that case.
2472 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2473 if( us->_head != them->_head || us->_tail != them->_tail ) {
2474 tty->print("Unequals loops for: ");
2475 n->dump();
2476 if( fail++ > 10 ) return;
2477 tty->print("We have it as: ");
2478 us->dump();
2479 tty->print("Verify thinks: ");
2480 them->dump();
2481 tty->cr();
2482 }
2483 }
2484 }
2485
2486 // Check for immediate dominators being equal
2487 if( i >= _idom_size ) {
2488 if( !n->is_CFG() ) return;
2489 tty->print("CFG Node with no idom: ");
2490 n->dump();
2491 return;
2492 }
2493 if( !n->is_CFG() ) return;
2494 if( n == C->root() ) return; // No IDOM here
2495
2496 assert(n->_idx == i, "sanity");
2497 Node *id = idom_no_update(n);
2498 if( id != loop_verify->idom_no_update(n) ) {
2499 tty->print("Unequals idoms for: ");
2500 n->dump();
2501 if( fail++ > 10 ) return;
2502 tty->print("We have it as: ");
2503 id->dump();
2504 tty->print("Verify thinks: ");
2505 loop_verify->idom_no_update(n)->dump();
2506 tty->cr();
2507 }
2508
2509 }
2510
2511 //------------------------------verify_tree------------------------------------
2512 // Verify that tree structures match. Because the CFG can change, siblings
2513 // within the loop tree can be reordered. We attempt to deal with that by
2514 // reordering the verify's loop tree if possible.
2515 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2516 assert( _parent == parent, "Badly formed loop tree" );
2517
2518 // Siblings not in same order? Attempt to re-order.
2519 if( _head != loop->_head ) {
2520 // Find _next pointer to update
2521 IdealLoopTree **pp = &loop->_parent->_child;
2522 while( *pp != loop )
2523 pp = &((*pp)->_next);
2524 // Find proper sibling to be next
2525 IdealLoopTree **nn = &loop->_next;
2526 while( (*nn) && (*nn)->_head != _head )
2527 nn = &((*nn)->_next);
2528
2529 // Check for no match.
2530 if( !(*nn) ) {
2531 // Annoyingly, irreducible loops can pick different headers
2532 // after a major_progress operation, so the rest of the loop
2533 // tree cannot be matched.
2534 if (_irreducible && Compile::current()->major_progress()) return;
2535 assert( 0, "failed to match loop tree" );
2536 }
2537
2538 // Move (*nn) to (*pp)
2539 IdealLoopTree *hit = *nn;
2540 *nn = hit->_next;
2541 hit->_next = loop;
2542 *pp = loop;
2543 loop = hit;
2544 // Now try again to verify
2545 }
2546
2547 assert( _head == loop->_head , "mismatched loop head" );
2548 Node *tail = _tail; // Inline a non-updating version of
2549 while( !tail->in(0) ) // the 'tail()' call.
2550 tail = tail->in(1);
2551 assert( tail == loop->_tail, "mismatched loop tail" );
2552
2553 // Counted loops that are guarded should be able to find their guards
2554 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2555 CountedLoopNode *cl = _head->as_CountedLoop();
2556 Node *init = cl->init_trip();
2557 Node *ctrl = cl->in(LoopNode::EntryControl);
2558 assert( ctrl->Opcode() == Opcodes::Op_IfTrue || ctrl->Opcode() == Opcodes::Op_IfFalse, "" );
2559 Node *iff = ctrl->in(0);
2560 assert( iff->Opcode() == Opcodes::Op_If, "" );
2561 Node *bol = iff->in(1);
2562 assert( bol->Opcode() == Opcodes::Op_Bool, "" );
2563 Node *cmp = bol->in(1);
2564 assert( cmp->Opcode() == Opcodes::Op_CmpI, "" );
2565 Node *add = cmp->in(1);
2566 Node *opaq;
2567 if( add->Opcode() == Opcodes::Op_Opaque1 ) {
2568 opaq = add;
2569 } else {
2570 assert( add->Opcode() == Opcodes::Op_AddI || add->Opcode() == Opcodes::Op_ConI , "" );
2571 assert( add == init, "" );
2572 opaq = cmp->in(2);
2573 }
2574 assert( opaq->Opcode() == Opcodes::Op_Opaque1, "" );
2575
2576 }
2577
2578 if (_child != NULL) _child->verify_tree(loop->_child, this);
2579 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2580 // Innermost loops need to verify loop bodies,
2581 // but only if no 'major_progress'
2582 int fail = 0;
2583 if (!Compile::current()->major_progress() && _child == NULL) {
2584 for( uint i = 0; i < _body.size(); i++ ) {
2585 Node *n = _body.at(i);
2586 if (n->outcnt() == 0) continue; // Ignore dead
2587 uint j;
2588 for( j = 0; j < loop->_body.size(); j++ )
2589 if( loop->_body.at(j) == n )
2590 break;
2591 if( j == loop->_body.size() ) { // Not found in loop body
2592 // Last ditch effort to avoid assertion: Its possible that we
2593 // have some users (so outcnt not zero) but are still dead.
2594 // Try to find from root.
2595 if (Compile::current()->root()->find(n->_idx)) {
2596 fail++;
2597 tty->print("We have that verify does not: ");
2598 n->dump();
2599 }
2600 }
2601 }
2602 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2603 Node *n = loop->_body.at(i2);
2604 if (n->outcnt() == 0) continue; // Ignore dead
2605 uint j;
2606 for( j = 0; j < _body.size(); j++ )
2607 if( _body.at(j) == n )
2608 break;
2609 if( j == _body.size() ) { // Not found in loop body
2610 // Last ditch effort to avoid assertion: Its possible that we
2611 // have some users (so outcnt not zero) but are still dead.
2612 // Try to find from root.
2613 if (Compile::current()->root()->find(n->_idx)) {
2614 fail++;
2615 tty->print("Verify has that we do not: ");
2616 n->dump();
2617 }
2618 }
2619 }
2620 assert( !fail, "loop body mismatch" );
2621 }
2622 }
2623
2624 #endif
2625
2626 //------------------------------set_idom---------------------------------------
2627 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2628 uint idx = d->_idx;
2629 if (idx >= _idom_size) {
2630 uint newsize = _idom_size<<1;
2631 while( idx >= newsize ) {
2632 newsize <<= 1;
2633 }
2634 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2635 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2636 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2637 _idom_size = newsize;
2638 }
2639 _idom[idx] = n;
2640 _dom_depth[idx] = dom_depth;
2641 }
2642
2643 //------------------------------recompute_dom_depth---------------------------------------
2644 // The dominator tree is constructed with only parent pointers.
2645 // This recomputes the depth in the tree by first tagging all
2646 // nodes as "no depth yet" marker. The next pass then runs up
2647 // the dom tree from each node marked "no depth yet", and computes
2648 // the depth on the way back down.
2649 void PhaseIdealLoop::recompute_dom_depth() {
2650 uint no_depth_marker = C->unique();
2651 uint i;
2652 // Initialize depth to "no depth yet"
2653 for (i = 0; i < _idom_size; i++) {
2654 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2655 _dom_depth[i] = no_depth_marker;
2656 }
2657 }
2658 if (_dom_stk == NULL) {
2659 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
2660 if (init_size < 10) init_size = 10;
2661 _dom_stk = new GrowableArray<uint>(init_size);
2662 }
2663 // Compute new depth for each node.
2664 for (i = 0; i < _idom_size; i++) {
2665 uint j = i;
2666 // Run up the dom tree to find a node with a depth
2667 while (_dom_depth[j] == no_depth_marker) {
2668 _dom_stk->push(j);
2669 j = _idom[j]->_idx;
2670 }
2671 // Compute the depth on the way back down this tree branch
2672 uint dd = _dom_depth[j] + 1;
2673 while (_dom_stk->length() > 0) {
2674 uint j = _dom_stk->pop();
2675 _dom_depth[j] = dd;
2676 dd++;
2677 }
2678 }
2679 }
2680
2681 //------------------------------sort-------------------------------------------
2682 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2683 // loop tree, not the root.
2684 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2685 if( !innermost ) return loop; // New innermost loop
2686
2687 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2688 assert( loop_preorder, "not yet post-walked loop" );
2689 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2690 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2691
2692 // Insert at start of list
2693 while( l ) { // Insertion sort based on pre-order
2694 if( l == loop ) return innermost; // Already on list!
2695 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2696 assert( l_preorder, "not yet post-walked l" );
2697 // Check header pre-order number to figure proper nesting
2698 if( loop_preorder > l_preorder )
2699 break; // End of insertion
2700 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2701 // Since I split shared headers, you'd think this could not happen.
2702 // BUT: I must first do the preorder numbering before I can discover I
2703 // have shared headers, so the split headers all get the same preorder
2704 // number as the RegionNode they split from.
2705 if( loop_preorder == l_preorder &&
2706 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2707 break; // Also check for shared headers (same pre#)
2708 pp = &l->_parent; // Chain up list
2709 l = *pp;
2710 }
2711 // Link into list
2712 // Point predecessor to me
2713 *pp = loop;
2714 // Point me to successor
2715 IdealLoopTree *p = loop->_parent;
2716 loop->_parent = l; // Point me to successor
2717 if( p ) sort( p, innermost ); // Insert my parents into list as well
2718 return innermost;
2719 }
2720
2721 //------------------------------build_loop_tree--------------------------------
2722 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2723 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2724 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2725 // tightest enclosing IdealLoopTree for post-walked.
2726 //
2727 // During my forward walk I do a short 1-layer lookahead to see if I can find
2728 // a loop backedge with that doesn't have any work on the backedge. This
2729 // helps me construct nested loops with shared headers better.
2730 //
2731 // Once I've done the forward recursion, I do the post-work. For each child
2732 // I check to see if there is a backedge. Backedges define a loop! I
2733 // insert an IdealLoopTree at the target of the backedge.
2734 //
2735 // During the post-work I also check to see if I have several children
2736 // belonging to different loops. If so, then this Node is a decision point
2737 // where control flow can choose to change loop nests. It is at this
2738 // decision point where I can figure out how loops are nested. At this
2739 // time I can properly order the different loop nests from my children.
2740 // Note that there may not be any backedges at the decision point!
2741 //
2742 // Since the decision point can be far removed from the backedges, I can't
2743 // order my loops at the time I discover them. Thus at the decision point
2744 // I need to inspect loop header pre-order numbers to properly nest my
2745 // loops. This means I need to sort my childrens' loops by pre-order.
2746 // The sort is of size number-of-control-children, which generally limits
2747 // it to size 2 (i.e., I just choose between my 2 target loops).
2748 void PhaseIdealLoop::build_loop_tree() {
2749 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
2750 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
2751 Node *n = C->root();
2752 bltstack.push(n);
2753 int pre_order = 1;
2754 int stack_size;
2755
2756 while ( ( stack_size = bltstack.length() ) != 0 ) {
2757 n = bltstack.top(); // Leave node on stack
2758 if ( !is_visited(n) ) {
2759 // ---- Pre-pass Work ----
2760 // Pre-walked but not post-walked nodes need a pre_order number.
2761
2762 set_preorder_visited( n, pre_order ); // set as visited
2763
2764 // ---- Scan over children ----
2765 // Scan first over control projections that lead to loop headers.
2766 // This helps us find inner-to-outer loops with shared headers better.
2767
2768 // Scan children's children for loop headers.
2769 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2770 Node* m = n->raw_out(i); // Child
2771 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2772 // Scan over children's children to find loop
2773 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2774 Node* l = m->fast_out(j);
2775 if( is_visited(l) && // Been visited?
2776 !is_postvisited(l) && // But not post-visited
2777 get_preorder(l) < pre_order ) { // And smaller pre-order
2778 // Found! Scan the DFS down this path before doing other paths
2779 bltstack.push(m);
2780 break;
2781 }
2782 }
2783 }
2784 }
2785 pre_order++;
2786 }
2787 else if ( !is_postvisited(n) ) {
2788 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2789 // such as com.sun.rsasign.am::a.
2790 // For non-recursive version, first, process current children.
2791 // On next iteration, check if additional children were added.
2792 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2793 Node* u = n->raw_out(k);
2794 if ( u->is_CFG() && !is_visited(u) ) {
2795 bltstack.push(u);
2796 }
2797 }
2798 if ( bltstack.length() == stack_size ) {
2799 // There were no additional children, post visit node now
2800 (void)bltstack.pop(); // Remove node from stack
2801 pre_order = build_loop_tree_impl( n, pre_order );
2802 // Check for bailout
2803 if (C->failing()) {
2804 return;
2805 }
2806 // Check to grow _preorders[] array for the case when
2807 // build_loop_tree_impl() adds new nodes.
2808 check_grow_preorders();
2809 }
2810 }
2811 else {
2812 (void)bltstack.pop(); // Remove post-visited node from stack
2813 }
2814 }
2815 }
2816
2817 //------------------------------build_loop_tree_impl---------------------------
2818 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2819 // ---- Post-pass Work ----
2820 // Pre-walked but not post-walked nodes need a pre_order number.
2821
2822 // Tightest enclosing loop for this Node
2823 IdealLoopTree *innermost = NULL;
2824
2825 // For all children, see if any edge is a backedge. If so, make a loop
2826 // for it. Then find the tightest enclosing loop for the self Node.
2827 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2828 Node* m = n->fast_out(i); // Child
2829 if( n == m ) continue; // Ignore control self-cycles
2830 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2831
2832 IdealLoopTree *l; // Child's loop
2833 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2834 // Found a backedge
2835 assert( get_preorder(m) < pre_order, "should be backedge" );
2836 // Check for the RootNode, which is already a LoopNode and is allowed
2837 // to have multiple "backedges".
2838 if( m == C->root()) { // Found the root?
2839 l = _ltree_root; // Root is the outermost LoopNode
2840 } else { // Else found a nested loop
2841 // Insert a LoopNode to mark this loop.
2842 l = new IdealLoopTree(this, m, n);
2843 } // End of Else found a nested loop
2844 if( !has_loop(m) ) // If 'm' does not already have a loop set
2845 set_loop(m, l); // Set loop header to loop now
2846
2847 } else { // Else not a nested loop
2848 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2849 l = get_loop(m); // Get previously determined loop
2850 // If successor is header of a loop (nest), move up-loop till it
2851 // is a member of some outer enclosing loop. Since there are no
2852 // shared headers (I've split them already) I only need to go up
2853 // at most 1 level.
2854 while( l && l->_head == m ) // Successor heads loop?
2855 l = l->_parent; // Move up 1 for me
2856 // If this loop is not properly parented, then this loop
2857 // has no exit path out, i.e. its an infinite loop.
2858 if( !l ) {
2859 // Make loop "reachable" from root so the CFG is reachable. Basically
2860 // insert a bogus loop exit that is never taken. 'm', the loop head,
2861 // points to 'n', one (of possibly many) fall-in paths. There may be
2862 // many backedges as well.
2863
2864 // Here I set the loop to be the root loop. I could have, after
2865 // inserting a bogus loop exit, restarted the recursion and found my
2866 // new loop exit. This would make the infinite loop a first-class
2867 // loop and it would then get properly optimized. What's the use of
2868 // optimizing an infinite loop?
2869 l = _ltree_root; // Oops, found infinite loop
2870
2871 if (!_verify_only) {
2872 // Insert the NeverBranch between 'm' and it's control user.
2873 NeverBranchNode *iff = new NeverBranchNode( m );
2874 _igvn.register_new_node_with_optimizer(iff);
2875 set_loop(iff, l);
2876 Node *if_t = new CProjNode( iff, 0 );
2877 _igvn.register_new_node_with_optimizer(if_t);
2878 set_loop(if_t, l);
2879
2880 Node* cfg = NULL; // Find the One True Control User of m
2881 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2882 Node* x = m->fast_out(j);
2883 if (x->is_CFG() && x != m && x != iff)
2884 { cfg = x; break; }
2885 }
2886 assert(cfg != NULL, "must find the control user of m");
2887 uint k = 0; // Probably cfg->in(0)
2888 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2889 cfg->set_req( k, if_t ); // Now point to NeverBranch
2890 _igvn._worklist.push(cfg);
2891
2892 // Now create the never-taken loop exit
2893 Node *if_f = new CProjNode( iff, 1 );
2894 _igvn.register_new_node_with_optimizer(if_f);
2895 set_loop(if_f, l);
2896 // Find frame ptr for Halt. Relies on the optimizer
2897 // V-N'ing. Easier and quicker than searching through
2898 // the program structure.
2899 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
2900 _igvn.register_new_node_with_optimizer(frame);
2901 // Halt & Catch Fire
2902 Node *halt = new HaltNode( if_f, frame );
2903 _igvn.register_new_node_with_optimizer(halt);
2904 set_loop(halt, l);
2905 C->root()->add_req(halt);
2906 }
2907 set_loop(C->root(), _ltree_root);
2908 }
2909 }
2910 // Weeny check for irreducible. This child was already visited (this
2911 // IS the post-work phase). Is this child's loop header post-visited
2912 // as well? If so, then I found another entry into the loop.
2913 if (!_verify_only) {
2914 while( is_postvisited(l->_head) ) {
2915 // found irreducible
2916 l->_irreducible = 1; // = true
2917 l = l->_parent;
2918 _has_irreducible_loops = true;
2919 // Check for bad CFG here to prevent crash, and bailout of compile
2920 if (l == NULL) {
2921 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2922 return pre_order;
2923 }
2924 }
2925 C->set_has_irreducible_loop(_has_irreducible_loops);
2926 }
2927
2928 // This Node might be a decision point for loops. It is only if
2929 // it's children belong to several different loops. The sort call
2930 // does a trivial amount of work if there is only 1 child or all
2931 // children belong to the same loop. If however, the children
2932 // belong to different loops, the sort call will properly set the
2933 // _parent pointers to show how the loops nest.
2934 //
2935 // In any case, it returns the tightest enclosing loop.
2936 innermost = sort( l, innermost );
2937 }
2938
2939 // Def-use info will have some dead stuff; dead stuff will have no
2940 // loop decided on.
2941
2942 // Am I a loop header? If so fix up my parent's child and next ptrs.
2943 if( innermost && innermost->_head == n ) {
2944 assert( get_loop(n) == innermost, "" );
2945 IdealLoopTree *p = innermost->_parent;
2946 IdealLoopTree *l = innermost;
2947 while( p && l->_head == n ) {
2948 l->_next = p->_child; // Put self on parents 'next child'
2949 p->_child = l; // Make self as first child of parent
2950 l = p; // Now walk up the parent chain
2951 p = l->_parent;
2952 }
2953 } else {
2954 // Note that it is possible for a LoopNode to reach here, if the
2955 // backedge has been made unreachable (hence the LoopNode no longer
2956 // denotes a Loop, and will eventually be removed).
2957
2958 // Record tightest enclosing loop for self. Mark as post-visited.
2959 set_loop(n, innermost);
2960 // Also record has_call flag early on
2961 if( innermost ) {
2962 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2963 // Do not count uncommon calls
2964 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2965 Node *iff = n->in(0)->in(0);
2966 // No any calls for vectorized loops.
2967 if( UseSuperWord || !iff->is_If() ||
2968 (n->in(0)->Opcode() == Opcodes::Op_IfFalse &&
2969 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2970 (iff->as_If()->_prob >= 0.01) )
2971 innermost->_has_call = 1;
2972 }
2973 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2974 // Disable loop optimizations if the loop has a scalar replaceable
2975 // allocation. This disabling may cause a potential performance lost
2976 // if the allocation is not eliminated for some reason.
2977 innermost->_allow_optimizations = false;
2978 innermost->_has_call = 1; // = true
2979 } else if (n->Opcode() == Opcodes::Op_SafePoint) {
2980 // Record all safepoints in this loop.
2981 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
2982 innermost->_safepts->push(n);
2983 }
2984 }
2985 }
2986
2987 // Flag as post-visited now
2988 set_postvisited(n);
2989 return pre_order;
2990 }
2991
2992
2993 //------------------------------build_loop_early-------------------------------
2994 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2995 // First pass computes the earliest controlling node possible. This is the
2996 // controlling input with the deepest dominating depth.
2997 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2998 while (worklist.size() != 0) {
2999 // Use local variables nstack_top_n & nstack_top_i to cache values
3000 // on nstack's top.
3001 Node *nstack_top_n = worklist.pop();
3002 uint nstack_top_i = 0;
3003 //while_nstack_nonempty:
3004 while (true) {
3005 // Get parent node and next input's index from stack's top.
3006 Node *n = nstack_top_n;
3007 uint i = nstack_top_i;
3008 uint cnt = n->req(); // Count of inputs
3009 if (i == 0) { // Pre-process the node.
3010 if( has_node(n) && // Have either loop or control already?
3011 !has_ctrl(n) ) { // Have loop picked out already?
3012 // During "merge_many_backedges" we fold up several nested loops
3013 // into a single loop. This makes the members of the original
3014 // loop bodies pointing to dead loops; they need to move up
3015 // to the new UNION'd larger loop. I set the _head field of these
3016 // dead loops to NULL and the _parent field points to the owning
3017 // loop. Shades of UNION-FIND algorithm.
3018 IdealLoopTree *ilt;
3019 while( !(ilt = get_loop(n))->_head ) {
3020 // Normally I would use a set_loop here. But in this one special
3021 // case, it is legal (and expected) to change what loop a Node
3022 // belongs to.
3023 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3024 }
3025 // Remove safepoints ONLY if I've already seen I don't need one.
3026 // (the old code here would yank a 2nd safepoint after seeing a
3027 // first one, even though the 1st did not dominate in the loop body
3028 // and thus could be avoided indefinitely)
3029 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Opcodes::Op_SafePoint &&
3030 is_deleteable_safept(n)) {
3031 Node *in = n->in(TypeFunc::Control);
3032 lazy_replace(n,in); // Pull safepoint now
3033 if (ilt->_safepts != NULL) {
3034 ilt->_safepts->yank(n);
3035 }
3036 // Carry on with the recursion "as if" we are walking
3037 // only the control input
3038 if( !visited.test_set( in->_idx ) ) {
3039 worklist.push(in); // Visit this guy later, using worklist
3040 }
3041 // Get next node from nstack:
3042 // - skip n's inputs processing by setting i > cnt;
3043 // - we also will not call set_early_ctrl(n) since
3044 // has_node(n) == true (see the condition above).
3045 i = cnt + 1;
3046 }
3047 }
3048 } // if (i == 0)
3049
3050 // Visit all inputs
3051 bool done = true; // Assume all n's inputs will be processed
3052 while (i < cnt) {
3053 Node *in = n->in(i);
3054 ++i;
3055 if (in == NULL) continue;
3056 if (in->pinned() && !in->is_CFG())
3057 set_ctrl(in, in->in(0));
3058 int is_visited = visited.test_set( in->_idx );
3059 if (!has_node(in)) { // No controlling input yet?
3060 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3061 assert( !is_visited, "visit only once" );
3062 nstack.push(n, i); // Save parent node and next input's index.
3063 nstack_top_n = in; // Process current input now.
3064 nstack_top_i = 0;
3065 done = false; // Not all n's inputs processed.
3066 break; // continue while_nstack_nonempty;
3067 } else if (!is_visited) {
3068 // This guy has a location picked out for him, but has not yet
3069 // been visited. Happens to all CFG nodes, for instance.
3070 // Visit him using the worklist instead of recursion, to break
3071 // cycles. Since he has a location already we do not need to
3072 // find his location before proceeding with the current Node.
3073 worklist.push(in); // Visit this guy later, using worklist
3074 }
3075 }
3076 if (done) {
3077 // All of n's inputs have been processed, complete post-processing.
3078
3079 // Compute earliest point this Node can go.
3080 // CFG, Phi, pinned nodes already know their controlling input.
3081 if (!has_node(n)) {
3082 // Record earliest legal location
3083 set_early_ctrl( n );
3084 }
3085 if (nstack.is_empty()) {
3086 // Finished all nodes on stack.
3087 // Process next node on the worklist.
3088 break;
3089 }
3090 // Get saved parent node and next input's index.
3091 nstack_top_n = nstack.node();
3092 nstack_top_i = nstack.index();
3093 nstack.pop();
3094 }
3095 } // while (true)
3096 }
3097 }
3098
3099 //------------------------------dom_lca_internal--------------------------------
3100 // Pair-wise LCA
3101 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3102 if( !n1 ) return n2; // Handle NULL original LCA
3103 assert( n1->is_CFG(), "" );
3104 assert( n2->is_CFG(), "" );
3105 // find LCA of all uses
3106 uint d1 = dom_depth(n1);
3107 uint d2 = dom_depth(n2);
3108 while (n1 != n2) {
3109 if (d1 > d2) {
3110 n1 = idom(n1);
3111 d1 = dom_depth(n1);
3112 } else if (d1 < d2) {
3113 n2 = idom(n2);
3114 d2 = dom_depth(n2);
3115 } else {
3116 // Here d1 == d2. Due to edits of the dominator-tree, sections
3117 // of the tree might have the same depth. These sections have
3118 // to be searched more carefully.
3119
3120 // Scan up all the n1's with equal depth, looking for n2.
3121 Node *t1 = idom(n1);
3122 while (dom_depth(t1) == d1) {
3123 if (t1 == n2) return n2;
3124 t1 = idom(t1);
3125 }
3126 // Scan up all the n2's with equal depth, looking for n1.
3127 Node *t2 = idom(n2);
3128 while (dom_depth(t2) == d2) {
3129 if (t2 == n1) return n1;
3130 t2 = idom(t2);
3131 }
3132 // Move up to a new dominator-depth value as well as up the dom-tree.
3133 n1 = t1;
3134 n2 = t2;
3135 d1 = dom_depth(n1);
3136 d2 = dom_depth(n2);
3137 }
3138 }
3139 return n1;
3140 }
3141
3142 //------------------------------compute_idom-----------------------------------
3143 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3144 // IDOMs are correct.
3145 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3146 assert( region->is_Region(), "" );
3147 Node *LCA = NULL;
3148 for( uint i = 1; i < region->req(); i++ ) {
3149 if( region->in(i) != C->top() )
3150 LCA = dom_lca( LCA, region->in(i) );
3151 }
3152 return LCA;
3153 }
3154
3155 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3156 bool had_error = false;
3157 #ifdef ASSERT
3158 if (early != C->root()) {
3159 // Make sure that there's a dominance path from LCA to early
3160 Node* d = LCA;
3161 while (d != early) {
3162 if (d == C->root()) {
3163 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3164 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3165 had_error = true;
3166 break;
3167 }
3168 d = idom(d);
3169 }
3170 }
3171 #endif
3172 return had_error;
3173 }
3174
3175
3176 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3177 // Compute LCA over list of uses
3178 bool had_error = false;
3179 Node *LCA = NULL;
3180 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3181 Node* c = n->fast_out(i);
3182 if (_nodes[c->_idx] == NULL)
3183 continue; // Skip the occasional dead node
3184 if( c->is_Phi() ) { // For Phis, we must land above on the path
3185 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3186 if( c->in(j) == n ) { // Found matching input?
3187 Node *use = c->in(0)->in(j);
3188 if (_verify_only && use->is_top()) continue;
3189 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3190 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3191 }
3192 }
3193 } else {
3194 // For CFG data-users, use is in the block just prior
3195 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3196 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3197 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3198 }
3199 }
3200 assert(!had_error, "bad dominance");
3201 return LCA;
3202 }
3203
3204 // Check the shape of the graph at the loop entry. In some cases,
3205 // the shape of the graph does not match the shape outlined below.
3206 // That is caused by the Opaque1 node "protecting" the shape of
3207 // the graph being removed by, for example, the IGVN performed
3208 // in PhaseIdealLoop::build_and_optimize().
3209 //
3210 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3211 // loop unswitching, and IGVN, or a combination of them) can freely change
3212 // the graph's shape. As a result, the graph shape outlined below cannot
3213 // be guaranteed anymore.
3214 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3215 if (!cl->is_main_loop() && !cl->is_post_loop()) {
3216 return false;
3217 }
3218 Node* ctrl = cl->in(LoopNode::EntryControl);
3219 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3220 return false;
3221 }
3222 Node* iffm = ctrl->in(0);
3223 if (iffm == NULL || !iffm->is_If()) {
3224 return false;
3225 }
3226 Node* bolzm = iffm->in(1);
3227 if (bolzm == NULL || !bolzm->is_Bool()) {
3228 return false;
3229 }
3230 Node* cmpzm = bolzm->in(1);
3231 if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3232 return false;
3233 }
3234 // compares can get conditionally flipped
3235 bool found_opaque = false;
3236 for (uint i = 1; i < cmpzm->req(); i++) {
3237 Node* opnd = cmpzm->in(i);
3238 if (opnd && opnd->Opcode() == Opcodes::Op_Opaque1) {
3239 found_opaque = true;
3240 break;
3241 }
3242 }
3243 if (!found_opaque) {
3244 return false;
3245 }
3246 return true;
3247 }
3248
3249 //------------------------------get_late_ctrl----------------------------------
3250 // Compute latest legal control.
3251 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3252 assert(early != NULL, "early control should not be NULL");
3253
3254 Node* LCA = compute_lca_of_uses(n, early);
3255 #ifdef ASSERT
3256 if (LCA == C->root() && LCA != early) {
3257 // def doesn't dominate uses so print some useful debugging output
3258 compute_lca_of_uses(n, early, true);
3259 }
3260 #endif
3261
3262 // if this is a load, check for anti-dependent stores
3263 // We use a conservative algorithm to identify potential interfering
3264 // instructions and for rescheduling the load. The users of the memory
3265 // input of this load are examined. Any use which is not a load and is
3266 // dominated by early is considered a potentially interfering store.
3267 // This can produce false positives.
3268 if (n->is_Load() && LCA != early) {
3269 Node_List worklist;
3270
3271 Node *mem = n->in(MemNode::Memory);
3272 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3273 Node* s = mem->fast_out(i);
3274 worklist.push(s);
3275 }
3276 while(worklist.size() != 0 && LCA != early) {
3277 Node* s = worklist.pop();
3278 if (s->is_Load()) {
3279 continue;
3280 } else if (s->is_MergeMem()) {
3281 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3282 Node* s1 = s->fast_out(i);
3283 worklist.push(s1);
3284 }
3285 } else {
3286 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3287 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3288 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3289 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3290 }
3291 }
3292 }
3293 }
3294
3295 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3296 return LCA;
3297 }
3298
3299 // true if CFG node d dominates CFG node n
3300 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3301 if (d == n)
3302 return true;
3303 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3304 uint dd = dom_depth(d);
3305 while (dom_depth(n) >= dd) {
3306 if (n == d)
3307 return true;
3308 n = idom(n);
3309 }
3310 return false;
3311 }
3312
3313 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3314 // Pair-wise LCA with tags.
3315 // Tag each index with the node 'tag' currently being processed
3316 // before advancing up the dominator chain using idom().
3317 // Later calls that find a match to 'tag' know that this path has already
3318 // been considered in the current LCA (which is input 'n1' by convention).
3319 // Since get_late_ctrl() is only called once for each node, the tag array
3320 // does not need to be cleared between calls to get_late_ctrl().
3321 // Algorithm trades a larger constant factor for better asymptotic behavior
3322 //
3323 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3324 uint d1 = dom_depth(n1);
3325 uint d2 = dom_depth(n2);
3326
3327 do {
3328 if (d1 > d2) {
3329 // current lca is deeper than n2
3330 _dom_lca_tags.map(n1->_idx, tag);
3331 n1 = idom(n1);
3332 d1 = dom_depth(n1);
3333 } else if (d1 < d2) {
3334 // n2 is deeper than current lca
3335 Node *memo = _dom_lca_tags[n2->_idx];
3336 if( memo == tag ) {
3337 return n1; // Return the current LCA
3338 }
3339 _dom_lca_tags.map(n2->_idx, tag);
3340 n2 = idom(n2);
3341 d2 = dom_depth(n2);
3342 } else {
3343 // Here d1 == d2. Due to edits of the dominator-tree, sections
3344 // of the tree might have the same depth. These sections have
3345 // to be searched more carefully.
3346
3347 // Scan up all the n1's with equal depth, looking for n2.
3348 _dom_lca_tags.map(n1->_idx, tag);
3349 Node *t1 = idom(n1);
3350 while (dom_depth(t1) == d1) {
3351 if (t1 == n2) return n2;
3352 _dom_lca_tags.map(t1->_idx, tag);
3353 t1 = idom(t1);
3354 }
3355 // Scan up all the n2's with equal depth, looking for n1.
3356 _dom_lca_tags.map(n2->_idx, tag);
3357 Node *t2 = idom(n2);
3358 while (dom_depth(t2) == d2) {
3359 if (t2 == n1) return n1;
3360 _dom_lca_tags.map(t2->_idx, tag);
3361 t2 = idom(t2);
3362 }
3363 // Move up to a new dominator-depth value as well as up the dom-tree.
3364 n1 = t1;
3365 n2 = t2;
3366 d1 = dom_depth(n1);
3367 d2 = dom_depth(n2);
3368 }
3369 } while (n1 != n2);
3370 return n1;
3371 }
3372
3373 //------------------------------init_dom_lca_tags------------------------------
3374 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3375 // Intended use does not involve any growth for the array, so it could
3376 // be of fixed size.
3377 void PhaseIdealLoop::init_dom_lca_tags() {
3378 uint limit = C->unique() + 1;
3379 _dom_lca_tags.map( limit, NULL );
3380 #ifdef ASSERT
3381 for( uint i = 0; i < limit; ++i ) {
3382 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3383 }
3384 #endif // ASSERT
3385 }
3386
3387 //------------------------------clear_dom_lca_tags------------------------------
3388 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3389 // Intended use does not involve any growth for the array, so it could
3390 // be of fixed size.
3391 void PhaseIdealLoop::clear_dom_lca_tags() {
3392 uint limit = C->unique() + 1;
3393 _dom_lca_tags.map( limit, NULL );
3394 _dom_lca_tags.clear();
3395 #ifdef ASSERT
3396 for( uint i = 0; i < limit; ++i ) {
3397 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3398 }
3399 #endif // ASSERT
3400 }
3401
3402 //------------------------------build_loop_late--------------------------------
3403 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3404 // Second pass finds latest legal placement, and ideal loop placement.
3405 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3406 while (worklist.size() != 0) {
3407 Node *n = worklist.pop();
3408 // Only visit once
3409 if (visited.test_set(n->_idx)) continue;
3410 uint cnt = n->outcnt();
3411 uint i = 0;
3412 while (true) {
3413 assert( _nodes[n->_idx], "no dead nodes" );
3414 // Visit all children
3415 if (i < cnt) {
3416 Node* use = n->raw_out(i);
3417 ++i;
3418 // Check for dead uses. Aggressively prune such junk. It might be
3419 // dead in the global sense, but still have local uses so I cannot
3420 // easily call 'remove_dead_node'.
3421 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3422 // Due to cycles, we might not hit the same fixed point in the verify
3423 // pass as we do in the regular pass. Instead, visit such phis as
3424 // simple uses of the loop head.
3425 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3426 if( !visited.test(use->_idx) )
3427 worklist.push(use);
3428 } else if( !visited.test_set(use->_idx) ) {
3429 nstack.push(n, i); // Save parent and next use's index.
3430 n = use; // Process all children of current use.
3431 cnt = use->outcnt();
3432 i = 0;
3433 }
3434 } else {
3435 // Do not visit around the backedge of loops via data edges.
3436 // push dead code onto a worklist
3437 _deadlist.push(use);
3438 }
3439 } else {
3440 // All of n's children have been processed, complete post-processing.
3441 build_loop_late_post(n);
3442 if (nstack.is_empty()) {
3443 // Finished all nodes on stack.
3444 // Process next node on the worklist.
3445 break;
3446 }
3447 // Get saved parent node and next use's index. Visit the rest of uses.
3448 n = nstack.node();
3449 cnt = n->outcnt();
3450 i = nstack.index();
3451 nstack.pop();
3452 }
3453 }
3454 }
3455 }
3456
3457 //------------------------------build_loop_late_post---------------------------
3458 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3459 // Second pass finds latest legal placement, and ideal loop placement.
3460 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3461
3462 if (n->req() == 2 && (n->Opcode() == Opcodes::Op_ConvI2L || n->Opcode() == Opcodes::Op_CastII) && !C->major_progress() && !_verify_only) {
3463 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3464 }
3465
3466 #ifdef ASSERT
3467 if (_verify_only && !n->is_CFG()) {
3468 // Check def-use domination.
3469 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
3470 }
3471 #endif
3472
3473 // CFG and pinned nodes already handled
3474 if( n->in(0) ) {
3475 if( n->in(0)->is_top() ) return; // Dead?
3476
3477 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3478 // _must_ be pinned (they have to observe their control edge of course).
3479 // Unlike Stores (which modify an unallocable resource, the memory
3480 // state), Mods/Loads can float around. So free them up.
3481 bool pinned = true;
3482 switch( n->Opcode() ) {
3483 case Opcodes::Op_DivI:
3484 case Opcodes::Op_DivF:
3485 case Opcodes::Op_DivD:
3486 case Opcodes::Op_ModI:
3487 case Opcodes::Op_ModF:
3488 case Opcodes::Op_ModD:
3489 case Opcodes::Op_LoadB: // Same with Loads; they can sink
3490 case Opcodes::Op_LoadUB: // during loop optimizations.
3491 case Opcodes::Op_LoadUS:
3492 case Opcodes::Op_LoadD:
3493 case Opcodes::Op_LoadF:
3494 case Opcodes::Op_LoadI:
3495 case Opcodes::Op_LoadKlass:
3496 case Opcodes::Op_LoadNKlass:
3497 case Opcodes::Op_LoadL:
3498 case Opcodes::Op_LoadS:
3499 case Opcodes::Op_LoadP:
3500 case Opcodes::Op_LoadN:
3501 case Opcodes::Op_LoadRange:
3502 case Opcodes::Op_LoadD_unaligned:
3503 case Opcodes::Op_LoadL_unaligned:
3504 case Opcodes::Op_StrComp: // Does a bunch of load-like effects
3505 case Opcodes::Op_StrEquals:
3506 case Opcodes::Op_StrIndexOf:
3507 case Opcodes::Op_StrIndexOfChar:
3508 case Opcodes::Op_AryEq:
3509 case Opcodes::Op_HasNegatives:
3510 pinned = false;
3511 }
3512 if( pinned ) {
3513 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3514 if( !chosen_loop->_child ) // Inner loop?
3515 chosen_loop->_body.push(n); // Collect inner loops
3516 return;
3517 }
3518 } else { // No slot zero
3519 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3520 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3521 return;
3522 }
3523 assert(!n->is_CFG() || n->outcnt() == 0, "");
3524 }
3525
3526 // Do I have a "safe range" I can select over?
3527 Node *early = get_ctrl(n);// Early location already computed
3528
3529 // Compute latest point this Node can go
3530 Node *LCA = get_late_ctrl( n, early );
3531 // LCA is NULL due to uses being dead
3532 if( LCA == NULL ) {
3533 #ifdef ASSERT
3534 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3535 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3536 }
3537 #endif
3538 _nodes.map(n->_idx, 0); // This node is useless
3539 _deadlist.push(n);
3540 return;
3541 }
3542 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3543
3544 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3545 Node *least = legal; // Best legal position so far
3546 while( early != legal ) { // While not at earliest legal
3547 #ifdef ASSERT
3548 if (legal->is_Start() && !early->is_Root()) {
3549 // Bad graph. Print idom path and fail.
3550 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
3551 assert(false, "Bad graph detected in build_loop_late");
3552 }
3553 #endif
3554 // Find least loop nesting depth
3555 legal = idom(legal); // Bump up the IDOM tree
3556 // Check for lower nesting depth
3557 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3558 least = legal;
3559 }
3560 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3561
3562 // Try not to place code on a loop entry projection
3563 // which can inhibit range check elimination.
3564 if (least != early) {
3565 Node* ctrl_out = least->unique_ctrl_out();
3566 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3567 least == ctrl_out->in(LoopNode::EntryControl)) {
3568 Node* least_dom = idom(least);
3569 if (get_loop(least_dom)->is_member(get_loop(least))) {
3570 least = least_dom;
3571 }
3572 }
3573 }
3574
3575 #ifdef ASSERT
3576 // If verifying, verify that 'verify_me' has a legal location
3577 // and choose it as our location.
3578 if( _verify_me ) {
3579 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3580 Node *legal = LCA;
3581 while( early != legal ) { // While not at earliest legal
3582 if( legal == v_ctrl ) break; // Check for prior good location
3583 legal = idom(legal) ;// Bump up the IDOM tree
3584 }
3585 // Check for prior good location
3586 if( legal == v_ctrl ) least = legal; // Keep prior if found
3587 }
3588 #endif
3589
3590 // Assign discovered "here or above" point
3591 least = find_non_split_ctrl(least);
3592 set_ctrl(n, least);
3593
3594 // Collect inner loop bodies
3595 IdealLoopTree *chosen_loop = get_loop(least);
3596 if( !chosen_loop->_child ) // Inner loop?
3597 chosen_loop->_body.push(n);// Collect inner loops
3598 }
3599
3600 #ifdef ASSERT
3601 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
3602 tty->print_cr("%s", msg);
3603 tty->print("n: "); n->dump();
3604 tty->print("early(n): "); early->dump();
3605 if (n->in(0) != NULL && !n->in(0)->is_top() &&
3606 n->in(0) != early && !n->in(0)->is_Root()) {
3607 tty->print("n->in(0): "); n->in(0)->dump();
3608 }
3609 for (uint i = 1; i < n->req(); i++) {
3610 Node* in1 = n->in(i);
3611 if (in1 != NULL && in1 != n && !in1->is_top()) {
3612 tty->print("n->in(%d): ", i); in1->dump();
3613 Node* in1_early = get_ctrl(in1);
3614 tty->print("early(n->in(%d)): ", i); in1_early->dump();
3615 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
3616 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3617 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3618 }
3619 for (uint j = 1; j < in1->req(); j++) {
3620 Node* in2 = in1->in(j);
3621 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3622 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3623 Node* in2_early = get_ctrl(in2);
3624 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3625 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
3626 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3627 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3628 }
3629 }
3630 }
3631 }
3632 }
3633 tty->cr();
3634 tty->print("LCA(n): "); LCA->dump();
3635 for (uint i = 0; i < n->outcnt(); i++) {
3636 Node* u1 = n->raw_out(i);
3637 if (u1 == n)
3638 continue;
3639 tty->print("n->out(%d): ", i); u1->dump();
3640 if (u1->is_CFG()) {
3641 for (uint j = 0; j < u1->outcnt(); j++) {
3642 Node* u2 = u1->raw_out(j);
3643 if (u2 != u1 && u2 != n && u2->is_CFG()) {
3644 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3645 }
3646 }
3647 } else {
3648 Node* u1_later = get_ctrl(u1);
3649 tty->print("later(n->out(%d)): ", i); u1_later->dump();
3650 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
3651 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3652 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3653 }
3654 for (uint j = 0; j < u1->outcnt(); j++) {
3655 Node* u2 = u1->raw_out(j);
3656 if (u2 == n || u2 == u1)
3657 continue;
3658 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3659 if (!u2->is_CFG()) {
3660 Node* u2_later = get_ctrl(u2);
3661 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3662 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
3663 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3664 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3665 }
3666 }
3667 }
3668 }
3669 }
3670 tty->cr();
3671 int ct = 0;
3672 Node *dbg_legal = LCA;
3673 while(!dbg_legal->is_Start() && ct < 100) {
3674 tty->print("idom[%d] ",ct); dbg_legal->dump();
3675 ct++;
3676 dbg_legal = idom(dbg_legal);
3677 }
3678 tty->cr();
3679 }
3680 #endif
3681
3682 #ifndef PRODUCT
3683 //------------------------------dump-------------------------------------------
3684 void PhaseIdealLoop::dump( ) const {
3685 ResourceMark rm;
3686 Arena* arena = Thread::current()->resource_area();
3687 Node_Stack stack(arena, C->live_nodes() >> 2);
3688 Node_List rpo_list;
3689 VectorSet visited(arena);
3690 visited.set(C->top()->_idx);
3691 rpo( C->root(), stack, visited, rpo_list );
3692 // Dump root loop indexed by last element in PO order
3693 dump( _ltree_root, rpo_list.size(), rpo_list );
3694 }
3695
3696 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3697 loop->dump_head();
3698
3699 // Now scan for CFG nodes in the same loop
3700 for( uint j=idx; j > 0; j-- ) {
3701 Node *n = rpo_list[j-1];
3702 if( !_nodes[n->_idx] ) // Skip dead nodes
3703 continue;
3704 if( get_loop(n) != loop ) { // Wrong loop nest
3705 if( get_loop(n)->_head == n && // Found nested loop?
3706 get_loop(n)->_parent == loop )
3707 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3708 continue;
3709 }
3710
3711 // Dump controlling node
3712 for( uint x = 0; x < loop->_nest; x++ )
3713 tty->print(" ");
3714 tty->print("C");
3715 if( n == C->root() ) {
3716 n->dump();
3717 } else {
3718 Node* cached_idom = idom_no_update(n);
3719 Node *computed_idom = n->in(0);
3720 if( n->is_Region() ) {
3721 computed_idom = compute_idom(n);
3722 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3723 // any MultiBranch ctrl node), so apply a similar transform to
3724 // the cached idom returned from idom_no_update.
3725 cached_idom = find_non_split_ctrl(cached_idom);
3726 }
3727 tty->print(" ID:%d",computed_idom->_idx);
3728 n->dump();
3729 if( cached_idom != computed_idom ) {
3730 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3731 computed_idom->_idx, cached_idom->_idx);
3732 }
3733 }
3734 // Dump nodes it controls
3735 for( uint k = 0; k < _nodes.Size(); k++ ) {
3736 // (k < C->unique() && get_ctrl(find(k)) == n)
3737 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3738 Node *m = C->root()->find(k);
3739 if( m && m->outcnt() > 0 ) {
3740 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3741 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3742 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3743 }
3744 for( uint j = 0; j < loop->_nest; j++ )
3745 tty->print(" ");
3746 tty->print(" ");
3747 m->dump();
3748 }
3749 }
3750 }
3751 }
3752 }
3753
3754 // Collect a R-P-O for the whole CFG.
3755 // Result list is in post-order (scan backwards for RPO)
3756 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3757 stk.push(start, 0);
3758 visited.set(start->_idx);
3759
3760 while (stk.is_nonempty()) {
3761 Node* m = stk.node();
3762 uint idx = stk.index();
3763 if (idx < m->outcnt()) {
3764 stk.set_index(idx + 1);
3765 Node* n = m->raw_out(idx);
3766 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3767 stk.push(n, 0);
3768 }
3769 } else {
3770 rpo_list.push(m);
3771 stk.pop();
3772 }
3773 }
3774 }
3775 #endif
3776
3777
3778 //=============================================================================
3779 //------------------------------LoopTreeIterator-----------------------------------
3780
3781 // Advance to next loop tree using a preorder, left-to-right traversal.
3782 void LoopTreeIterator::next() {
3783 assert(!done(), "must not be done.");
3784 if (_curnt->_child != NULL) {
3785 _curnt = _curnt->_child;
3786 } else if (_curnt->_next != NULL) {
3787 _curnt = _curnt->_next;
3788 } else {
3789 while (_curnt != _root && _curnt->_next == NULL) {
3790 _curnt = _curnt->_parent;
3791 }
3792 if (_curnt == _root) {
3793 _curnt = NULL;
3794 assert(done(), "must be done.");
3795 } else {
3796 assert(_curnt->_next != NULL, "must be more to do");
3797 _curnt = _curnt->_next;
3798 }
3799 }
3800 }