1 /*
2 * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/superword.hpp"
39
40 //=============================================================================
41 //------------------------------is_loop_iv-------------------------------------
42 // Determine if a node is Counted loop induction variable.
43 // The method is declared in node.hpp.
44 const Node* Node::is_loop_iv() const {
45 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
46 this->as_Phi()->region()->is_CountedLoop() &&
47 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
48 return this;
49 } else {
50 return NULL;
51 }
52 }
53
54 //=============================================================================
55 //------------------------------dump_spec--------------------------------------
56 // Dump special per-node info
57 #ifndef PRODUCT
58 void LoopNode::dump_spec(outputStream *st) const {
59 if (is_inner_loop()) st->print( "inner " );
60 if (is_partial_peel_loop()) st->print( "partial_peel " );
61 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
62 }
63 #endif
64
65 //------------------------------is_valid_counted_loop-------------------------
66 bool LoopNode::is_valid_counted_loop() const {
67 if (is_CountedLoop()) {
68 CountedLoopNode* l = as_CountedLoop();
69 CountedLoopEndNode* le = l->loopexit();
70 if (le != NULL &&
71 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
72 Node* phi = l->phi();
73 Node* exit = le->proj_out(0 /* false */);
74 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
75 phi != NULL && phi->is_Phi() &&
76 phi->in(LoopNode::LoopBackControl) == l->incr() &&
77 le->loopnode() == l && le->stride_is_con()) {
78 return true;
79 }
80 }
81 }
82 return false;
83 }
84
85 //------------------------------get_early_ctrl---------------------------------
86 // Compute earliest legal control
87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
88 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
89 uint i;
90 Node *early;
91 if( n->in(0) ) {
92 early = n->in(0);
93 if( !early->is_CFG() ) // Might be a non-CFG multi-def
94 early = get_ctrl(early); // So treat input as a straight data input
95 i = 1;
96 } else {
97 early = get_ctrl(n->in(1));
98 i = 2;
99 }
100 uint e_d = dom_depth(early);
101 assert( early, "" );
102 for( ; i < n->req(); i++ ) {
103 Node *cin = get_ctrl(n->in(i));
104 assert( cin, "" );
105 // Keep deepest dominator depth
106 uint c_d = dom_depth(cin);
107 if( c_d > e_d ) { // Deeper guy?
108 early = cin; // Keep deepest found so far
109 e_d = c_d;
110 } else if( c_d == e_d && // Same depth?
111 early != cin ) { // If not equal, must use slower algorithm
112 // If same depth but not equal, one _must_ dominate the other
113 // and we want the deeper (i.e., dominated) guy.
114 Node *n1 = early;
115 Node *n2 = cin;
116 while( 1 ) {
117 n1 = idom(n1); // Walk up until break cycle
118 n2 = idom(n2);
119 if( n1 == cin || // Walked early up to cin
120 dom_depth(n2) < c_d )
121 break; // early is deeper; keep him
122 if( n2 == early || // Walked cin up to early
123 dom_depth(n1) < c_d ) {
124 early = cin; // cin is deeper; keep him
125 break;
126 }
127 }
128 e_d = dom_depth(early); // Reset depth register cache
129 }
130 }
131
132 // Return earliest legal location
133 assert(early == find_non_split_ctrl(early), "unexpected early control");
134
135 return early;
136 }
137
138 //------------------------------set_early_ctrl---------------------------------
139 // Set earliest legal control
140 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
141 Node *early = get_early_ctrl(n);
142
143 // Record earliest legal location
144 set_ctrl(n, early);
145 }
146
147 //------------------------------set_subtree_ctrl-------------------------------
148 // set missing _ctrl entries on new nodes
149 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
150 // Already set? Get out.
151 if( _nodes[n->_idx] ) return;
152 // Recursively set _nodes array to indicate where the Node goes
153 uint i;
154 for( i = 0; i < n->req(); ++i ) {
155 Node *m = n->in(i);
156 if( m && m != C->root() )
157 set_subtree_ctrl( m );
158 }
159
160 // Fixup self
161 set_early_ctrl( n );
162 }
163
164 //------------------------------is_counted_loop--------------------------------
165 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
166 PhaseGVN *gvn = &_igvn;
167
168 // Counted loop head must be a good RegionNode with only 3 not NULL
169 // control input edges: Self, Entry, LoopBack.
170 if (x->in(LoopNode::Self) == NULL || x->req() != 3)
171 return false;
172
173 Node *init_control = x->in(LoopNode::EntryControl);
174 Node *back_control = x->in(LoopNode::LoopBackControl);
175 if (init_control == NULL || back_control == NULL) // Partially dead
176 return false;
177 // Must also check for TOP when looking for a dead loop
178 if (init_control->is_top() || back_control->is_top())
179 return false;
180
181 // Allow funny placement of Safepoint
182 if (back_control->Opcode() == Op_SafePoint)
183 back_control = back_control->in(TypeFunc::Control);
184
185 // Controlling test for loop
186 Node *iftrue = back_control;
187 uint iftrue_op = iftrue->Opcode();
188 if (iftrue_op != Op_IfTrue &&
189 iftrue_op != Op_IfFalse)
190 // I have a weird back-control. Probably the loop-exit test is in
191 // the middle of the loop and I am looking at some trailing control-flow
192 // merge point. To fix this I would have to partially peel the loop.
193 return false; // Obscure back-control
194
195 // Get boolean guarding loop-back test
196 Node *iff = iftrue->in(0);
197 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
198 return false;
199 BoolNode *test = iff->in(1)->as_Bool();
200 BoolTest::mask bt = test->_test._test;
201 float cl_prob = iff->as_If()->_prob;
202 if (iftrue_op == Op_IfFalse) {
203 bt = BoolTest(bt).negate();
204 cl_prob = 1.0 - cl_prob;
205 }
206 // Get backedge compare
207 Node *cmp = test->in(1);
208 int cmp_op = cmp->Opcode();
209 if (cmp_op != Op_CmpI)
210 return false; // Avoid pointer & float compares
211
212 // Find the trip-counter increment & limit. Limit must be loop invariant.
213 Node *incr = cmp->in(1);
214 Node *limit = cmp->in(2);
215
216 // ---------
217 // need 'loop()' test to tell if limit is loop invariant
218 // ---------
219
220 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
221 Node *tmp = incr; // Then reverse order into the CmpI
222 incr = limit;
223 limit = tmp;
224 bt = BoolTest(bt).commute(); // And commute the exit test
225 }
226 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
227 return false;
228 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
229 return false;
230
231 Node* phi_incr = NULL;
232 // Trip-counter increment must be commutative & associative.
233 if (incr->is_Phi()) {
234 if (incr->as_Phi()->region() != x || incr->req() != 3)
235 return false; // Not simple trip counter expression
236 phi_incr = incr;
237 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
238 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
239 return false;
240 }
241
242 Node* trunc1 = NULL;
243 Node* trunc2 = NULL;
244 const TypeInt* iv_trunc_t = NULL;
245 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
246 return false; // Funny increment opcode
247 }
248 assert(incr->Opcode() == Op_AddI, "wrong increment code");
249
250 // Get merge point
251 Node *xphi = incr->in(1);
252 Node *stride = incr->in(2);
253 if (!stride->is_Con()) { // Oops, swap these
254 if (!xphi->is_Con()) // Is the other guy a constant?
255 return false; // Nope, unknown stride, bail out
256 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
257 xphi = stride;
258 stride = tmp;
259 }
260 // Stride must be constant
261 int stride_con = stride->get_int();
262 if (stride_con == 0)
263 return false; // missed some peephole opt
264
265 if (!xphi->is_Phi())
266 return false; // Too much math on the trip counter
267 if (phi_incr != NULL && phi_incr != xphi)
268 return false;
269 PhiNode *phi = xphi->as_Phi();
270
271 // Phi must be of loop header; backedge must wrap to increment
272 if (phi->region() != x)
273 return false;
274 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
275 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
276 return false;
277 }
278 Node *init_trip = phi->in(LoopNode::EntryControl);
279
280 // If iv trunc type is smaller than int, check for possible wrap.
281 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
282 assert(trunc1 != NULL, "must have found some truncation");
283
284 // Get a better type for the phi (filtered thru if's)
285 const TypeInt* phi_ft = filtered_type(phi);
286
287 // Can iv take on a value that will wrap?
288 //
289 // Ensure iv's limit is not within "stride" of the wrap value.
290 //
291 // Example for "short" type
292 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
293 // If the stride is +10, then the last value of the induction
294 // variable before the increment (phi_ft->_hi) must be
295 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
296 // ensure no truncation occurs after the increment.
297
298 if (stride_con > 0) {
299 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
300 iv_trunc_t->_lo > phi_ft->_lo) {
301 return false; // truncation may occur
302 }
303 } else if (stride_con < 0) {
304 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
305 iv_trunc_t->_hi < phi_ft->_hi) {
306 return false; // truncation may occur
307 }
308 }
309 // No possibility of wrap so truncation can be discarded
310 // Promote iv type to Int
311 } else {
312 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
313 }
314
315 // If the condition is inverted and we will be rolling
316 // through MININT to MAXINT, then bail out.
317 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
318 // Odd stride
319 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
320 // Count down loop rolls through MAXINT
321 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
322 // Count up loop rolls through MININT
323 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
324 return false; // Bail out
325 }
326
327 const TypeInt* init_t = gvn->type(init_trip)->is_int();
328 const TypeInt* limit_t = gvn->type(limit)->is_int();
329
330 if (stride_con > 0) {
331 long init_p = (long)init_t->_lo + stride_con;
332 if (init_p > (long)max_jint || init_p > (long)limit_t->_hi)
333 return false; // cyclic loop or this loop trips only once
334 } else {
335 long init_p = (long)init_t->_hi + stride_con;
336 if (init_p < (long)min_jint || init_p < (long)limit_t->_lo)
337 return false; // cyclic loop or this loop trips only once
338 }
339
340 // =================================================
341 // ---- SUCCESS! Found A Trip-Counted Loop! -----
342 //
343 assert(x->Opcode() == Op_Loop, "regular loops only");
344 C->print_method("Before CountedLoop", 3);
345
346 Node *hook = new (C, 6) Node(6);
347
348 if (LoopLimitCheck) {
349
350 // ===================================================
351 // Generate loop limit check to avoid integer overflow
352 // in cases like next (cyclic loops):
353 //
354 // for (i=0; i <= max_jint; i++) {}
355 // for (i=0; i < max_jint; i+=2) {}
356 //
357 //
358 // Limit check predicate depends on the loop test:
359 //
360 // for(;i != limit; i++) --> limit <= (max_jint)
361 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
362 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
363 //
364
365 // Check if limit is excluded to do more precise int overflow check.
366 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
367 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
368
369 // If compare points directly to the phi we need to adjust
370 // the compare so that it points to the incr. Limit have
371 // to be adjusted to keep trip count the same and the
372 // adjusted limit should be checked for int overflow.
373 if (phi_incr != NULL) {
374 stride_m += stride_con;
375 }
376
377 if (limit->is_Con()) {
378 int limit_con = limit->get_int();
379 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
380 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
381 // Bailout: it could be integer overflow.
382 return false;
383 }
384 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
385 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
386 // Limit's type may satisfy the condition, for example,
387 // when it is an array length.
388 } else {
389 // Generate loop's limit check.
390 // Loop limit check predicate should be near the loop.
391 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
392 if (!limit_check_proj) {
393 // The limit check predicate is not generated if this method trapped here before.
394 #ifdef ASSERT
395 if (TraceLoopLimitCheck) {
396 tty->print("missing loop limit check:");
397 loop->dump_head();
398 x->dump(1);
399 }
400 #endif
401 return false;
402 }
403
404 IfNode* check_iff = limit_check_proj->in(0)->as_If();
405 Node* cmp_limit;
406 Node* bol;
407
408 if (stride_con > 0) {
409 cmp_limit = new (C, 3) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
410 bol = new (C, 2) BoolNode(cmp_limit, BoolTest::le);
411 } else {
412 cmp_limit = new (C, 3) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
413 bol = new (C, 2) BoolNode(cmp_limit, BoolTest::ge);
414 }
415 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
416 bol = _igvn.register_new_node_with_optimizer(bol);
417 set_subtree_ctrl(bol);
418
419 // Replace condition in original predicate but preserve Opaque node
420 // so that previous predicates could be found.
421 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
422 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
423 Node* opq = check_iff->in(1)->in(1);
424 _igvn.hash_delete(opq);
425 opq->set_req(1, bol);
426 // Update ctrl.
427 set_ctrl(opq, check_iff->in(0));
428 set_ctrl(check_iff->in(1), check_iff->in(0));
429
430 #ifndef PRODUCT
431 // report that the loop predication has been actually performed
432 // for this loop
433 if (TraceLoopLimitCheck) {
434 tty->print_cr("Counted Loop Limit Check generated:");
435 debug_only( bol->dump(2); )
436 }
437 #endif
438 }
439
440 if (phi_incr != NULL) {
441 // If compare points directly to the phi we need to adjust
442 // the compare so that it points to the incr. Limit have
443 // to be adjusted to keep trip count the same and we
444 // should avoid int overflow.
445 //
446 // i = init; do {} while(i++ < limit);
447 // is converted to
448 // i = init; do {} while(++i < limit+1);
449 //
450 limit = gvn->transform(new (C, 3) AddINode(limit, stride));
451 }
452
453 // Now we need to canonicalize loop condition.
454 if (bt == BoolTest::ne) {
455 assert(stride_con == 1 || stride_con == -1, "simple increment only");
456 // 'ne' can be replaced with 'lt' only when init < limit.
457 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
458 bt = BoolTest::lt;
459 // 'ne' can be replaced with 'gt' only when init > limit.
460 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
461 bt = BoolTest::gt;
462 }
463
464 if (incl_limit) {
465 // The limit check guaranties that 'limit <= (max_jint - stride)' so
466 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
467 //
468 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
469 limit = gvn->transform(new (C, 3) AddINode(limit, one));
470 if (bt == BoolTest::le)
471 bt = BoolTest::lt;
472 else if (bt == BoolTest::ge)
473 bt = BoolTest::gt;
474 else
475 ShouldNotReachHere();
476 }
477 set_subtree_ctrl( limit );
478
479 } else { // LoopLimitCheck
480
481 // If compare points to incr, we are ok. Otherwise the compare
482 // can directly point to the phi; in this case adjust the compare so that
483 // it points to the incr by adjusting the limit.
484 if (cmp->in(1) == phi || cmp->in(2) == phi)
485 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
486
487 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
488 // Final value for iterator should be: trip_count * stride + init_trip.
489 Node *one_p = gvn->intcon( 1);
490 Node *one_m = gvn->intcon(-1);
491
492 Node *trip_count = NULL;
493 switch( bt ) {
494 case BoolTest::eq:
495 ShouldNotReachHere();
496 case BoolTest::ne: // Ahh, the case we desire
497 if (stride_con == 1)
498 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
499 else if (stride_con == -1)
500 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
501 else
502 ShouldNotReachHere();
503 set_subtree_ctrl(trip_count);
504 //_loop.map(trip_count->_idx,loop(limit));
505 break;
506 case BoolTest::le: // Maybe convert to '<' case
507 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
508 set_subtree_ctrl( limit );
509 hook->init_req(4, limit);
510
511 bt = BoolTest::lt;
512 // Make the new limit be in the same loop nest as the old limit
513 //_loop.map(limit->_idx,limit_loop);
514 // Fall into next case
515 case BoolTest::lt: { // Maybe convert to '!=' case
516 if (stride_con < 0) // Count down loop rolls through MAXINT
517 ShouldNotReachHere();
518 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
519 set_subtree_ctrl( range );
520 hook->init_req(0, range);
521
522 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
523 set_subtree_ctrl( bias );
524 hook->init_req(1, bias);
525
526 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
527 set_subtree_ctrl( bias1 );
528 hook->init_req(2, bias1);
529
530 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
531 set_subtree_ctrl( trip_count );
532 hook->init_req(3, trip_count);
533 break;
534 }
535
536 case BoolTest::ge: // Maybe convert to '>' case
537 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
538 set_subtree_ctrl( limit );
539 hook->init_req(4 ,limit);
540
541 bt = BoolTest::gt;
542 // Make the new limit be in the same loop nest as the old limit
543 //_loop.map(limit->_idx,limit_loop);
544 // Fall into next case
545 case BoolTest::gt: { // Maybe convert to '!=' case
546 if (stride_con > 0) // count up loop rolls through MININT
547 ShouldNotReachHere();
548 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
549 set_subtree_ctrl( range );
550 hook->init_req(0, range);
551
552 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
553 set_subtree_ctrl( bias );
554 hook->init_req(1, bias);
555
556 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
557 set_subtree_ctrl( bias1 );
558 hook->init_req(2, bias1);
559
560 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
561 set_subtree_ctrl( trip_count );
562 hook->init_req(3, trip_count);
563 break;
564 }
565 } // switch( bt )
566
567 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
568 set_subtree_ctrl( span );
569 hook->init_req(5, span);
570
571 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
572 set_subtree_ctrl( limit );
573
574 } // LoopLimitCheck
575
576 // Check for SafePoint on backedge and remove
577 Node *sfpt = x->in(LoopNode::LoopBackControl);
578 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
579 lazy_replace( sfpt, iftrue );
580 loop->_tail = iftrue;
581 }
582
583 // Build a canonical trip test.
584 // Clone code, as old values may be in use.
585 incr = incr->clone();
586 incr->set_req(1,phi);
587 incr->set_req(2,stride);
588 incr = _igvn.register_new_node_with_optimizer(incr);
589 set_early_ctrl( incr );
590 _igvn.hash_delete(phi);
591 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
592
593 // If phi type is more restrictive than Int, raise to
594 // Int to prevent (almost) infinite recursion in igvn
595 // which can only handle integer types for constants or minint..maxint.
596 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
597 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
598 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
599 nphi = _igvn.register_new_node_with_optimizer(nphi);
600 set_ctrl(nphi, get_ctrl(phi));
601 _igvn.replace_node(phi, nphi);
602 phi = nphi->as_Phi();
603 }
604 cmp = cmp->clone();
605 cmp->set_req(1,incr);
606 cmp->set_req(2,limit);
607 cmp = _igvn.register_new_node_with_optimizer(cmp);
608 set_ctrl(cmp, iff->in(0));
609
610 test = test->clone()->as_Bool();
611 (*(BoolTest*)&test->_test)._test = bt;
612 test->set_req(1,cmp);
613 _igvn.register_new_node_with_optimizer(test);
614 set_ctrl(test, iff->in(0));
615
616 // Replace the old IfNode with a new LoopEndNode
617 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
618 IfNode *le = lex->as_If();
619 uint dd = dom_depth(iff);
620 set_idom(le, le->in(0), dd); // Update dominance for loop exit
621 set_loop(le, loop);
622
623 // Get the loop-exit control
624 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
625
626 // Need to swap loop-exit and loop-back control?
627 if (iftrue_op == Op_IfFalse) {
628 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
629 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
630
631 loop->_tail = back_control = ift2;
632 set_loop(ift2, loop);
633 set_loop(iff2, get_loop(iffalse));
634
635 // Lazy update of 'get_ctrl' mechanism.
636 lazy_replace_proj( iffalse, iff2 );
637 lazy_replace_proj( iftrue, ift2 );
638
639 // Swap names
640 iffalse = iff2;
641 iftrue = ift2;
642 } else {
643 _igvn.hash_delete(iffalse);
644 _igvn.hash_delete(iftrue);
645 iffalse->set_req_X( 0, le, &_igvn );
646 iftrue ->set_req_X( 0, le, &_igvn );
647 }
648
649 set_idom(iftrue, le, dd+1);
650 set_idom(iffalse, le, dd+1);
651 assert(iff->outcnt() == 0, "should be dead now");
652 lazy_replace( iff, le ); // fix 'get_ctrl'
653
654 // Now setup a new CountedLoopNode to replace the existing LoopNode
655 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
656 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
657 // The following assert is approximately true, and defines the intention
658 // of can_be_counted_loop. It fails, however, because phase->type
659 // is not yet initialized for this loop and its parts.
660 //assert(l->can_be_counted_loop(this), "sanity");
661 _igvn.register_new_node_with_optimizer(l);
662 set_loop(l, loop);
663 loop->_head = l;
664 // Fix all data nodes placed at the old loop head.
665 // Uses the lazy-update mechanism of 'get_ctrl'.
666 lazy_replace( x, l );
667 set_idom(l, init_control, dom_depth(x));
668
669 // Check for immediately preceding SafePoint and remove
670 Node *sfpt2 = le->in(0);
671 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
672 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
673
674 // Free up intermediate goo
675 _igvn.remove_dead_node(hook);
676
677 #ifdef ASSERT
678 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
679 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
680 #endif
681 #ifndef PRODUCT
682 if (TraceLoopOpts) {
683 tty->print("Counted ");
684 loop->dump_head();
685 }
686 #endif
687
688 C->print_method("After CountedLoop", 3);
689
690 return true;
691 }
692
693 //----------------------exact_limit-------------------------------------------
694 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
695 assert(loop->_head->is_CountedLoop(), "");
696 CountedLoopNode *cl = loop->_head->as_CountedLoop();
697 assert(cl->is_valid_counted_loop(), "");
698
699 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
700 cl->limit()->Opcode() == Op_LoopLimit) {
701 // Old code has exact limit (it could be incorrect in case of int overflow).
702 // Loop limit is exact with stride == 1. And loop may already have exact limit.
703 return cl->limit();
704 }
705 Node *limit = NULL;
706 #ifdef ASSERT
707 BoolTest::mask bt = cl->loopexit()->test_trip();
708 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
709 #endif
710 if (cl->has_exact_trip_count()) {
711 // Simple case: loop has constant boundaries.
712 // Use longs to avoid integer overflow.
713 int stride_con = cl->stride_con();
714 long init_con = cl->init_trip()->get_int();
715 long limit_con = cl->limit()->get_int();
716 julong trip_cnt = cl->trip_count();
717 long final_con = init_con + trip_cnt*stride_con;
718 final_con -= stride_con;
719 int final_int = (int)final_con;
720 // The final value should be in integer range since the loop
721 // is counted and the limit was checked for overflow.
722 assert(final_con == (long)final_int, "final value should be integer");
723 limit = _igvn.intcon(final_int);
724 } else {
725 // Create new LoopLimit node to get exact limit (final iv value).
726 limit = new (C, 4) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
727 register_new_node(limit, cl->in(LoopNode::EntryControl));
728 }
729 assert(limit != NULL, "sanity");
730 return limit;
731 }
732
733 //------------------------------Ideal------------------------------------------
734 // Return a node which is more "ideal" than the current node.
735 // Attempt to convert into a counted-loop.
736 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
737 if (!can_be_counted_loop(phase)) {
738 phase->C->set_major_progress();
739 }
740 return RegionNode::Ideal(phase, can_reshape);
741 }
742
743
744 //=============================================================================
745 //------------------------------Ideal------------------------------------------
746 // Return a node which is more "ideal" than the current node.
747 // Attempt to convert into a counted-loop.
748 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
749 return RegionNode::Ideal(phase, can_reshape);
750 }
751
752 //------------------------------dump_spec--------------------------------------
753 // Dump special per-node info
754 #ifndef PRODUCT
755 void CountedLoopNode::dump_spec(outputStream *st) const {
756 LoopNode::dump_spec(st);
757 if (stride_is_con()) {
758 st->print("stride: %d ",stride_con());
759 }
760 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
761 if (is_main_loop()) st->print("main of N%d", _idx);
762 if (is_post_loop()) st->print("post of N%d", _main_idx);
763 }
764 #endif
765
766 //=============================================================================
767 int CountedLoopEndNode::stride_con() const {
768 return stride()->bottom_type()->is_int()->get_con();
769 }
770
771 //=============================================================================
772 //------------------------------Value-----------------------------------------
773 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
774 const Type* init_t = phase->type(in(Init));
775 const Type* limit_t = phase->type(in(Limit));
776 const Type* stride_t = phase->type(in(Stride));
777 // Either input is TOP ==> the result is TOP
778 if (init_t == Type::TOP) return Type::TOP;
779 if (limit_t == Type::TOP) return Type::TOP;
780 if (stride_t == Type::TOP) return Type::TOP;
781
782 int stride_con = stride_t->is_int()->get_con();
783 if (stride_con == 1)
784 return NULL; // Identity
785
786 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
787 // Use longs to avoid integer overflow.
788 long init_con = init_t->is_int()->get_con();
789 long limit_con = limit_t->is_int()->get_con();
790 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
791 long trip_count = (limit_con - init_con + stride_m)/stride_con;
792 long final_con = init_con + stride_con*trip_count;
793 int final_int = (int)final_con;
794 // The final value should be in integer range since the loop
795 // is counted and the limit was checked for overflow.
796 assert(final_con == (long)final_int, "final value should be integer");
797 return TypeInt::make(final_int);
798 }
799
800 return bottom_type(); // TypeInt::INT
801 }
802
803 //------------------------------Ideal------------------------------------------
804 // Return a node which is more "ideal" than the current node.
805 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
806 if (phase->type(in(Init)) == Type::TOP ||
807 phase->type(in(Limit)) == Type::TOP ||
808 phase->type(in(Stride)) == Type::TOP)
809 return NULL; // Dead
810
811 int stride_con = phase->type(in(Stride))->is_int()->get_con();
812 if (stride_con == 1)
813 return NULL; // Identity
814
815 if (in(Init)->is_Con() && in(Limit)->is_Con())
816 return NULL; // Value
817
818 // Delay following optimizations until all loop optimizations
819 // done to keep Ideal graph simple.
820 if (!can_reshape || phase->C->major_progress())
821 return NULL;
822
823 const TypeInt* init_t = phase->type(in(Init) )->is_int();
824 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
825 int stride_p;
826 long lim, ini;
827 julong max;
828 if (stride_con > 0) {
829 stride_p = stride_con;
830 lim = limit_t->_hi;
831 ini = init_t->_lo;
832 max = (julong)max_jint;
833 } else {
834 stride_p = -stride_con;
835 lim = init_t->_hi;
836 ini = limit_t->_lo;
837 max = (julong)min_jint;
838 }
839 julong range = lim - ini + stride_p;
840 if (range <= max) {
841 // Convert to integer expression if it is not overflow.
842 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
843 Node *range = phase->transform(new (phase->C, 3) SubINode(in(Limit), in(Init)));
844 Node *bias = phase->transform(new (phase->C, 3) AddINode(range, stride_m));
845 Node *trip = phase->transform(new (phase->C, 3) DivINode(0, bias, in(Stride)));
846 Node *span = phase->transform(new (phase->C, 3) MulINode(trip, in(Stride)));
847 return new (phase->C, 3) AddINode(span, in(Init)); // exact limit
848 }
849
850 if (is_power_of_2(stride_p) || // divisor is 2^n
851 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
852 // Convert to long expression to avoid integer overflow
853 // and let igvn optimizer convert this division.
854 //
855 Node* init = phase->transform( new (phase->C, 2) ConvI2LNode(in(Init)));
856 Node* limit = phase->transform( new (phase->C, 2) ConvI2LNode(in(Limit)));
857 Node* stride = phase->longcon(stride_con);
858 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
859
860 Node *range = phase->transform(new (phase->C, 3) SubLNode(limit, init));
861 Node *bias = phase->transform(new (phase->C, 3) AddLNode(range, stride_m));
862 Node *span;
863 if (stride_con > 0 && is_power_of_2(stride_p)) {
864 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
865 // and avoid generating rounding for division. Zero trip guard should
866 // guarantee that init < limit but sometimes the guard is missing and
867 // we can get situation when init > limit. Note, for the empty loop
868 // optimization zero trip guard is generated explicitly which leaves
869 // only RCE predicate where exact limit is used and the predicate
870 // will simply fail forcing recompilation.
871 Node* neg_stride = phase->longcon(-stride_con);
872 span = phase->transform(new (phase->C, 3) AndLNode(bias, neg_stride));
873 } else {
874 Node *trip = phase->transform(new (phase->C, 3) DivLNode(0, bias, stride));
875 span = phase->transform(new (phase->C, 3) MulLNode(trip, stride));
876 }
877 // Convert back to int
878 Node *span_int = phase->transform(new (phase->C, 2) ConvL2INode(span));
879 return new (phase->C, 3) AddINode(span_int, in(Init)); // exact limit
880 }
881
882 return NULL; // No progress
883 }
884
885 //------------------------------Identity---------------------------------------
886 // If stride == 1 return limit node.
887 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
888 int stride_con = phase->type(in(Stride))->is_int()->get_con();
889 if (stride_con == 1 || stride_con == -1)
890 return in(Limit);
891 return this;
892 }
893
894 //=============================================================================
895 //----------------------match_incr_with_optional_truncation--------------------
896 // Match increment with optional truncation:
897 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
898 // Return NULL for failure. Success returns the increment node.
899 Node* CountedLoopNode::match_incr_with_optional_truncation(
900 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
901 // Quick cutouts:
902 if (expr == NULL || expr->req() != 3) return false;
903
904 Node *t1 = NULL;
905 Node *t2 = NULL;
906 const TypeInt* trunc_t = TypeInt::INT;
907 Node* n1 = expr;
908 int n1op = n1->Opcode();
909
910 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
911 if (n1op == Op_AndI &&
912 n1->in(2)->is_Con() &&
913 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
914 // %%% This check should match any mask of 2**K-1.
915 t1 = n1;
916 n1 = t1->in(1);
917 n1op = n1->Opcode();
918 trunc_t = TypeInt::CHAR;
919 } else if (n1op == Op_RShiftI &&
920 n1->in(1) != NULL &&
921 n1->in(1)->Opcode() == Op_LShiftI &&
922 n1->in(2) == n1->in(1)->in(2) &&
923 n1->in(2)->is_Con()) {
924 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
925 // %%% This check should match any shift in [1..31].
926 if (shift == 16 || shift == 8) {
927 t1 = n1;
928 t2 = t1->in(1);
929 n1 = t2->in(1);
930 n1op = n1->Opcode();
931 if (shift == 16) {
932 trunc_t = TypeInt::SHORT;
933 } else if (shift == 8) {
934 trunc_t = TypeInt::BYTE;
935 }
936 }
937 }
938
939 // If (maybe after stripping) it is an AddI, we won:
940 if (n1op == Op_AddI) {
941 *trunc1 = t1;
942 *trunc2 = t2;
943 *trunc_type = trunc_t;
944 return n1;
945 }
946
947 // failed
948 return NULL;
949 }
950
951
952 //------------------------------filtered_type--------------------------------
953 // Return a type based on condition control flow
954 // A successful return will be a type that is restricted due
955 // to a series of dominating if-tests, such as:
956 // if (i < 10) {
957 // if (i > 0) {
958 // here: "i" type is [1..10)
959 // }
960 // }
961 // or a control flow merge
962 // if (i < 10) {
963 // do {
964 // phi( , ) -- at top of loop type is [min_int..10)
965 // i = ?
966 // } while ( i < 10)
967 //
968 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
969 assert(n && n->bottom_type()->is_int(), "must be int");
970 const TypeInt* filtered_t = NULL;
971 if (!n->is_Phi()) {
972 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
973 filtered_t = filtered_type_from_dominators(n, n_ctrl);
974
975 } else {
976 Node* phi = n->as_Phi();
977 Node* region = phi->in(0);
978 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
979 if (region && region != C->top()) {
980 for (uint i = 1; i < phi->req(); i++) {
981 Node* val = phi->in(i);
982 Node* use_c = region->in(i);
983 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
984 if (val_t != NULL) {
985 if (filtered_t == NULL) {
986 filtered_t = val_t;
987 } else {
988 filtered_t = filtered_t->meet(val_t)->is_int();
989 }
990 }
991 }
992 }
993 }
994 const TypeInt* n_t = _igvn.type(n)->is_int();
995 if (filtered_t != NULL) {
996 n_t = n_t->join(filtered_t)->is_int();
997 }
998 return n_t;
999 }
1000
1001
1002 //------------------------------filtered_type_from_dominators--------------------------------
1003 // Return a possibly more restrictive type for val based on condition control flow of dominators
1004 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1005 if (val->is_Con()) {
1006 return val->bottom_type()->is_int();
1007 }
1008 uint if_limit = 10; // Max number of dominating if's visited
1009 const TypeInt* rtn_t = NULL;
1010
1011 if (use_ctrl && use_ctrl != C->top()) {
1012 Node* val_ctrl = get_ctrl(val);
1013 uint val_dom_depth = dom_depth(val_ctrl);
1014 Node* pred = use_ctrl;
1015 uint if_cnt = 0;
1016 while (if_cnt < if_limit) {
1017 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1018 if_cnt++;
1019 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1020 if (if_t != NULL) {
1021 if (rtn_t == NULL) {
1022 rtn_t = if_t;
1023 } else {
1024 rtn_t = rtn_t->join(if_t)->is_int();
1025 }
1026 }
1027 }
1028 pred = idom(pred);
1029 if (pred == NULL || pred == C->top()) {
1030 break;
1031 }
1032 // Stop if going beyond definition block of val
1033 if (dom_depth(pred) < val_dom_depth) {
1034 break;
1035 }
1036 }
1037 }
1038 return rtn_t;
1039 }
1040
1041
1042 //------------------------------dump_spec--------------------------------------
1043 // Dump special per-node info
1044 #ifndef PRODUCT
1045 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1046 if( in(TestValue)->is_Bool() ) {
1047 BoolTest bt( test_trip()); // Added this for g++.
1048
1049 st->print("[");
1050 bt.dump_on(st);
1051 st->print("]");
1052 }
1053 st->print(" ");
1054 IfNode::dump_spec(st);
1055 }
1056 #endif
1057
1058 //=============================================================================
1059 //------------------------------is_member--------------------------------------
1060 // Is 'l' a member of 'this'?
1061 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1062 while( l->_nest > _nest ) l = l->_parent;
1063 return l == this;
1064 }
1065
1066 //------------------------------set_nest---------------------------------------
1067 // Set loop tree nesting depth. Accumulate _has_call bits.
1068 int IdealLoopTree::set_nest( uint depth ) {
1069 _nest = depth;
1070 int bits = _has_call;
1071 if( _child ) bits |= _child->set_nest(depth+1);
1072 if( bits ) _has_call = 1;
1073 if( _next ) bits |= _next ->set_nest(depth );
1074 return bits;
1075 }
1076
1077 //------------------------------split_fall_in----------------------------------
1078 // Split out multiple fall-in edges from the loop header. Move them to a
1079 // private RegionNode before the loop. This becomes the loop landing pad.
1080 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1081 PhaseIterGVN &igvn = phase->_igvn;
1082 uint i;
1083
1084 // Make a new RegionNode to be the landing pad.
1085 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
1086 phase->set_loop(landing_pad,_parent);
1087 // Gather all the fall-in control paths into the landing pad
1088 uint icnt = fall_in_cnt;
1089 uint oreq = _head->req();
1090 for( i = oreq-1; i>0; i-- )
1091 if( !phase->is_member( this, _head->in(i) ) )
1092 landing_pad->set_req(icnt--,_head->in(i));
1093
1094 // Peel off PhiNode edges as well
1095 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1096 Node *oj = _head->fast_out(j);
1097 if( oj->is_Phi() ) {
1098 PhiNode* old_phi = oj->as_Phi();
1099 assert( old_phi->region() == _head, "" );
1100 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1101 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1102 uint icnt = fall_in_cnt;
1103 for( i = oreq-1; i>0; i-- ) {
1104 if( !phase->is_member( this, _head->in(i) ) ) {
1105 p->init_req(icnt--, old_phi->in(i));
1106 // Go ahead and clean out old edges from old phi
1107 old_phi->del_req(i);
1108 }
1109 }
1110 // Search for CSE's here, because ZKM.jar does a lot of
1111 // loop hackery and we need to be a little incremental
1112 // with the CSE to avoid O(N^2) node blow-up.
1113 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1114 if( p2 ) { // Found CSE
1115 p->destruct(); // Recover useless new node
1116 p = p2; // Use old node
1117 } else {
1118 igvn.register_new_node_with_optimizer(p, old_phi);
1119 }
1120 // Make old Phi refer to new Phi.
1121 old_phi->add_req(p);
1122 // Check for the special case of making the old phi useless and
1123 // disappear it. In JavaGrande I have a case where this useless
1124 // Phi is the loop limit and prevents recognizing a CountedLoop
1125 // which in turn prevents removing an empty loop.
1126 Node *id_old_phi = old_phi->Identity( &igvn );
1127 if( id_old_phi != old_phi ) { // Found a simple identity?
1128 // Note that I cannot call 'replace_node' here, because
1129 // that will yank the edge from old_phi to the Region and
1130 // I'm mid-iteration over the Region's uses.
1131 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1132 Node* use = old_phi->last_out(i);
1133 igvn.hash_delete(use);
1134 igvn._worklist.push(use);
1135 uint uses_found = 0;
1136 for (uint j = 0; j < use->len(); j++) {
1137 if (use->in(j) == old_phi) {
1138 if (j < use->req()) use->set_req (j, id_old_phi);
1139 else use->set_prec(j, id_old_phi);
1140 uses_found++;
1141 }
1142 }
1143 i -= uses_found; // we deleted 1 or more copies of this edge
1144 }
1145 }
1146 igvn._worklist.push(old_phi);
1147 }
1148 }
1149 // Finally clean out the fall-in edges from the RegionNode
1150 for( i = oreq-1; i>0; i-- ) {
1151 if( !phase->is_member( this, _head->in(i) ) ) {
1152 _head->del_req(i);
1153 }
1154 }
1155 // Transform landing pad
1156 igvn.register_new_node_with_optimizer(landing_pad, _head);
1157 // Insert landing pad into the header
1158 _head->add_req(landing_pad);
1159 }
1160
1161 //------------------------------split_outer_loop-------------------------------
1162 // Split out the outermost loop from this shared header.
1163 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1164 PhaseIterGVN &igvn = phase->_igvn;
1165
1166 // Find index of outermost loop; it should also be my tail.
1167 uint outer_idx = 1;
1168 while( _head->in(outer_idx) != _tail ) outer_idx++;
1169
1170 // Make a LoopNode for the outermost loop.
1171 Node *ctl = _head->in(LoopNode::EntryControl);
1172 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
1173 outer = igvn.register_new_node_with_optimizer(outer, _head);
1174 phase->set_created_loop_node();
1175
1176 // Outermost loop falls into '_head' loop
1177 _head->set_req(LoopNode::EntryControl, outer);
1178 _head->del_req(outer_idx);
1179 // Split all the Phis up between '_head' loop and 'outer' loop.
1180 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1181 Node *out = _head->fast_out(j);
1182 if( out->is_Phi() ) {
1183 PhiNode *old_phi = out->as_Phi();
1184 assert( old_phi->region() == _head, "" );
1185 Node *phi = PhiNode::make_blank(outer, old_phi);
1186 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1187 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1188 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1189 // Make old Phi point to new Phi on the fall-in path
1190 igvn.hash_delete(old_phi);
1191 old_phi->set_req(LoopNode::EntryControl, phi);
1192 old_phi->del_req(outer_idx);
1193 igvn._worklist.push(old_phi);
1194 }
1195 }
1196
1197 // Use the new loop head instead of the old shared one
1198 _head = outer;
1199 phase->set_loop(_head, this);
1200 }
1201
1202 //------------------------------fix_parent-------------------------------------
1203 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1204 loop->_parent = parent;
1205 if( loop->_child ) fix_parent( loop->_child, loop );
1206 if( loop->_next ) fix_parent( loop->_next , parent );
1207 }
1208
1209 //------------------------------estimate_path_freq-----------------------------
1210 static float estimate_path_freq( Node *n ) {
1211 // Try to extract some path frequency info
1212 IfNode *iff;
1213 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1214 uint nop = n->Opcode();
1215 if( nop == Op_SafePoint ) { // Skip any safepoint
1216 n = n->in(0);
1217 continue;
1218 }
1219 if( nop == Op_CatchProj ) { // Get count from a prior call
1220 // Assume call does not always throw exceptions: means the call-site
1221 // count is also the frequency of the fall-through path.
1222 assert( n->is_CatchProj(), "" );
1223 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1224 return 0.0f; // Assume call exception path is rare
1225 Node *call = n->in(0)->in(0)->in(0);
1226 assert( call->is_Call(), "expect a call here" );
1227 const JVMState *jvms = ((CallNode*)call)->jvms();
1228 ciMethodData* methodData = jvms->method()->method_data();
1229 if (!methodData->is_mature()) return 0.0f; // No call-site data
1230 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1231 if ((data == NULL) || !data->is_CounterData()) {
1232 // no call profile available, try call's control input
1233 n = n->in(0);
1234 continue;
1235 }
1236 return data->as_CounterData()->count()/FreqCountInvocations;
1237 }
1238 // See if there's a gating IF test
1239 Node *n_c = n->in(0);
1240 if( !n_c->is_If() ) break; // No estimate available
1241 iff = n_c->as_If();
1242 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1243 // Compute how much count comes on this path
1244 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1245 // Have no count info. Skip dull uncommon-trap like branches.
1246 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1247 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1248 break;
1249 // Skip through never-taken branch; look for a real loop exit.
1250 n = iff->in(0);
1251 }
1252 return 0.0f; // No estimate available
1253 }
1254
1255 //------------------------------merge_many_backedges---------------------------
1256 // Merge all the backedges from the shared header into a private Region.
1257 // Feed that region as the one backedge to this loop.
1258 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1259 uint i;
1260
1261 // Scan for the top 2 hottest backedges
1262 float hotcnt = 0.0f;
1263 float warmcnt = 0.0f;
1264 uint hot_idx = 0;
1265 // Loop starts at 2 because slot 1 is the fall-in path
1266 for( i = 2; i < _head->req(); i++ ) {
1267 float cnt = estimate_path_freq(_head->in(i));
1268 if( cnt > hotcnt ) { // Grab hottest path
1269 warmcnt = hotcnt;
1270 hotcnt = cnt;
1271 hot_idx = i;
1272 } else if( cnt > warmcnt ) { // And 2nd hottest path
1273 warmcnt = cnt;
1274 }
1275 }
1276
1277 // See if the hottest backedge is worthy of being an inner loop
1278 // by being much hotter than the next hottest backedge.
1279 if( hotcnt <= 0.0001 ||
1280 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1281
1282 // Peel out the backedges into a private merge point; peel
1283 // them all except optionally hot_idx.
1284 PhaseIterGVN &igvn = phase->_igvn;
1285
1286 Node *hot_tail = NULL;
1287 // Make a Region for the merge point
1288 Node *r = new (phase->C, 1) RegionNode(1);
1289 for( i = 2; i < _head->req(); i++ ) {
1290 if( i != hot_idx )
1291 r->add_req( _head->in(i) );
1292 else hot_tail = _head->in(i);
1293 }
1294 igvn.register_new_node_with_optimizer(r, _head);
1295 // Plug region into end of loop _head, followed by hot_tail
1296 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1297 _head->set_req(2, r);
1298 if( hot_idx ) _head->add_req(hot_tail);
1299
1300 // Split all the Phis up between '_head' loop and the Region 'r'
1301 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1302 Node *out = _head->fast_out(j);
1303 if( out->is_Phi() ) {
1304 PhiNode* n = out->as_Phi();
1305 igvn.hash_delete(n); // Delete from hash before hacking edges
1306 Node *hot_phi = NULL;
1307 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
1308 // Check all inputs for the ones to peel out
1309 uint j = 1;
1310 for( uint i = 2; i < n->req(); i++ ) {
1311 if( i != hot_idx )
1312 phi->set_req( j++, n->in(i) );
1313 else hot_phi = n->in(i);
1314 }
1315 // Register the phi but do not transform until whole place transforms
1316 igvn.register_new_node_with_optimizer(phi, n);
1317 // Add the merge phi to the old Phi
1318 while( n->req() > 3 ) n->del_req( n->req()-1 );
1319 n->set_req(2, phi);
1320 if( hot_idx ) n->add_req(hot_phi);
1321 }
1322 }
1323
1324
1325 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1326 // of self loop tree. Turn self into a loop headed by _head and with
1327 // tail being the new merge point.
1328 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1329 phase->set_loop(_tail,ilt); // Adjust tail
1330 _tail = r; // Self's tail is new merge point
1331 phase->set_loop(r,this);
1332 ilt->_child = _child; // New guy has my children
1333 _child = ilt; // Self has new guy as only child
1334 ilt->_parent = this; // new guy has self for parent
1335 ilt->_nest = _nest; // Same nesting depth (for now)
1336
1337 // Starting with 'ilt', look for child loop trees using the same shared
1338 // header. Flatten these out; they will no longer be loops in the end.
1339 IdealLoopTree **pilt = &_child;
1340 while( ilt ) {
1341 if( ilt->_head == _head ) {
1342 uint i;
1343 for( i = 2; i < _head->req(); i++ )
1344 if( _head->in(i) == ilt->_tail )
1345 break; // Still a loop
1346 if( i == _head->req() ) { // No longer a loop
1347 // Flatten ilt. Hang ilt's "_next" list from the end of
1348 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1349 IdealLoopTree **cp = &ilt->_child;
1350 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1351 *cp = ilt->_next; // Hang next list at end of child list
1352 *pilt = ilt->_child; // Move child up to replace ilt
1353 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1354 ilt = ilt->_child; // Repeat using new ilt
1355 continue; // do not advance over ilt->_child
1356 }
1357 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1358 phase->set_loop(_head,ilt);
1359 }
1360 pilt = &ilt->_child; // Advance to next
1361 ilt = *pilt;
1362 }
1363
1364 if( _child ) fix_parent( _child, this );
1365 }
1366
1367 //------------------------------beautify_loops---------------------------------
1368 // Split shared headers and insert loop landing pads.
1369 // Insert a LoopNode to replace the RegionNode.
1370 // Return TRUE if loop tree is structurally changed.
1371 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1372 bool result = false;
1373 // Cache parts in locals for easy
1374 PhaseIterGVN &igvn = phase->_igvn;
1375
1376 igvn.hash_delete(_head); // Yank from hash before hacking edges
1377
1378 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1379 int fall_in_cnt = 0;
1380 for( uint i = 1; i < _head->req(); i++ )
1381 if( !phase->is_member( this, _head->in(i) ) )
1382 fall_in_cnt++;
1383 assert( fall_in_cnt, "at least 1 fall-in path" );
1384 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1385 split_fall_in( phase, fall_in_cnt );
1386
1387 // Swap inputs to the _head and all Phis to move the fall-in edge to
1388 // the left.
1389 fall_in_cnt = 1;
1390 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1391 fall_in_cnt++;
1392 if( fall_in_cnt > 1 ) {
1393 // Since I am just swapping inputs I do not need to update def-use info
1394 Node *tmp = _head->in(1);
1395 _head->set_req( 1, _head->in(fall_in_cnt) );
1396 _head->set_req( fall_in_cnt, tmp );
1397 // Swap also all Phis
1398 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1399 Node* phi = _head->fast_out(i);
1400 if( phi->is_Phi() ) {
1401 igvn.hash_delete(phi); // Yank from hash before hacking edges
1402 tmp = phi->in(1);
1403 phi->set_req( 1, phi->in(fall_in_cnt) );
1404 phi->set_req( fall_in_cnt, tmp );
1405 }
1406 }
1407 }
1408 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1409 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1410
1411 // If I am a shared header (multiple backedges), peel off the many
1412 // backedges into a private merge point and use the merge point as
1413 // the one true backedge.
1414 if( _head->req() > 3 ) {
1415 // Merge the many backedges into a single backedge but leave
1416 // the hottest backedge as separate edge for the following peel.
1417 merge_many_backedges( phase );
1418 result = true;
1419 }
1420
1421 // If I have one hot backedge, peel off myself loop.
1422 // I better be the outermost loop.
1423 if( _head->req() > 3 ) {
1424 split_outer_loop( phase );
1425 result = true;
1426
1427 } else if( !_head->is_Loop() && !_irreducible ) {
1428 // Make a new LoopNode to replace the old loop head
1429 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1430 l = igvn.register_new_node_with_optimizer(l, _head);
1431 phase->set_created_loop_node();
1432 // Go ahead and replace _head
1433 phase->_igvn.replace_node( _head, l );
1434 _head = l;
1435 phase->set_loop(_head, this);
1436 }
1437
1438 // Now recursively beautify nested loops
1439 if( _child ) result |= _child->beautify_loops( phase );
1440 if( _next ) result |= _next ->beautify_loops( phase );
1441 return result;
1442 }
1443
1444 //------------------------------allpaths_check_safepts----------------------------
1445 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1446 // encountered. Helper for check_safepts.
1447 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1448 assert(stack.size() == 0, "empty stack");
1449 stack.push(_tail);
1450 visited.Clear();
1451 visited.set(_tail->_idx);
1452 while (stack.size() > 0) {
1453 Node* n = stack.pop();
1454 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1455 // Terminate this path
1456 } else if (n->Opcode() == Op_SafePoint) {
1457 if (_phase->get_loop(n) != this) {
1458 if (_required_safept == NULL) _required_safept = new Node_List();
1459 _required_safept->push(n); // save the one closest to the tail
1460 }
1461 // Terminate this path
1462 } else {
1463 uint start = n->is_Region() ? 1 : 0;
1464 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1465 for (uint i = start; i < end; i++) {
1466 Node* in = n->in(i);
1467 assert(in->is_CFG(), "must be");
1468 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1469 stack.push(in);
1470 }
1471 }
1472 }
1473 }
1474 }
1475
1476 //------------------------------check_safepts----------------------------
1477 // Given dominators, try to find loops with calls that must always be
1478 // executed (call dominates loop tail). These loops do not need non-call
1479 // safepoints (ncsfpt).
1480 //
1481 // A complication is that a safepoint in a inner loop may be needed
1482 // by an outer loop. In the following, the inner loop sees it has a
1483 // call (block 3) on every path from the head (block 2) to the
1484 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1485 // in block 2, _but_ this leaves the outer loop without a safepoint.
1486 //
1487 // entry 0
1488 // |
1489 // v
1490 // outer 1,2 +->1
1491 // | |
1492 // | v
1493 // | 2<---+ ncsfpt in 2
1494 // |_/|\ |
1495 // | v |
1496 // inner 2,3 / 3 | call in 3
1497 // / | |
1498 // v +--+
1499 // exit 4
1500 //
1501 //
1502 // This method creates a list (_required_safept) of ncsfpt nodes that must
1503 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1504 // is first looked for in the lists for the outer loops of the current loop.
1505 //
1506 // The insights into the problem:
1507 // A) counted loops are okay
1508 // B) innermost loops are okay (only an inner loop can delete
1509 // a ncsfpt needed by an outer loop)
1510 // C) a loop is immune from an inner loop deleting a safepoint
1511 // if the loop has a call on the idom-path
1512 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1513 // idom-path that is not in a nested loop
1514 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1515 // loop needs to be prevented from deletion by an inner loop
1516 //
1517 // There are two analyses:
1518 // 1) The first, and cheaper one, scans the loop body from
1519 // tail to head following the idom (immediate dominator)
1520 // chain, looking for the cases (C,D,E) above.
1521 // Since inner loops are scanned before outer loops, there is summary
1522 // information about inner loops. Inner loops can be skipped over
1523 // when the tail of an inner loop is encountered.
1524 //
1525 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1526 // the idom path (which is rare), scans all predecessor control paths
1527 // from the tail to the head, terminating a path when a call or sfpt
1528 // is encountered, to find the ncsfpt's that are closest to the tail.
1529 //
1530 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1531 // Bottom up traversal
1532 IdealLoopTree* ch = _child;
1533 while (ch != NULL) {
1534 ch->check_safepts(visited, stack);
1535 ch = ch->_next;
1536 }
1537
1538 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1539 bool has_call = false; // call on dom-path
1540 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1541 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1542 // Scan the dom-path nodes from tail to head
1543 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1544 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1545 has_call = true;
1546 _has_sfpt = 1; // Then no need for a safept!
1547 break;
1548 } else if (n->Opcode() == Op_SafePoint) {
1549 if (_phase->get_loop(n) == this) {
1550 has_local_ncsfpt = true;
1551 break;
1552 }
1553 if (nonlocal_ncsfpt == NULL) {
1554 nonlocal_ncsfpt = n; // save the one closest to the tail
1555 }
1556 } else {
1557 IdealLoopTree* nlpt = _phase->get_loop(n);
1558 if (this != nlpt) {
1559 // If at an inner loop tail, see if the inner loop has already
1560 // recorded seeing a call on the dom-path (and stop.) If not,
1561 // jump to the head of the inner loop.
1562 assert(is_member(nlpt), "nested loop");
1563 Node* tail = nlpt->_tail;
1564 if (tail->in(0)->is_If()) tail = tail->in(0);
1565 if (n == tail) {
1566 // If inner loop has call on dom-path, so does outer loop
1567 if (nlpt->_has_sfpt) {
1568 has_call = true;
1569 _has_sfpt = 1;
1570 break;
1571 }
1572 // Skip to head of inner loop
1573 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1574 n = nlpt->_head;
1575 }
1576 }
1577 }
1578 }
1579 // Record safept's that this loop needs preserved when an
1580 // inner loop attempts to delete it's safepoints.
1581 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1582 if (nonlocal_ncsfpt != NULL) {
1583 if (_required_safept == NULL) _required_safept = new Node_List();
1584 _required_safept->push(nonlocal_ncsfpt);
1585 } else {
1586 // Failed to find a suitable safept on the dom-path. Now use
1587 // an all paths walk from tail to head, looking for safepoints to preserve.
1588 allpaths_check_safepts(visited, stack);
1589 }
1590 }
1591 }
1592 }
1593
1594 //---------------------------is_deleteable_safept----------------------------
1595 // Is safept not required by an outer loop?
1596 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1597 assert(sfpt->Opcode() == Op_SafePoint, "");
1598 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1599 while (lp != NULL) {
1600 Node_List* sfpts = lp->_required_safept;
1601 if (sfpts != NULL) {
1602 for (uint i = 0; i < sfpts->size(); i++) {
1603 if (sfpt == sfpts->at(i))
1604 return false;
1605 }
1606 }
1607 lp = lp->_parent;
1608 }
1609 return true;
1610 }
1611
1612 //---------------------------replace_parallel_iv-------------------------------
1613 // Replace parallel induction variable (parallel to trip counter)
1614 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1615 assert(loop->_head->is_CountedLoop(), "");
1616 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1617 if (!cl->is_valid_counted_loop())
1618 return; // skip malformed counted loop
1619 Node *incr = cl->incr();
1620 if (incr == NULL)
1621 return; // Dead loop?
1622 Node *init = cl->init_trip();
1623 Node *phi = cl->phi();
1624 int stride_con = cl->stride_con();
1625
1626 // Visit all children, looking for Phis
1627 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1628 Node *out = cl->out(i);
1629 // Look for other phis (secondary IVs). Skip dead ones
1630 if (!out->is_Phi() || out == phi || !has_node(out))
1631 continue;
1632 PhiNode* phi2 = out->as_Phi();
1633 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1634 // Look for induction variables of the form: X += constant
1635 if (phi2->region() != loop->_head ||
1636 incr2->req() != 3 ||
1637 incr2->in(1) != phi2 ||
1638 incr2 == incr ||
1639 incr2->Opcode() != Op_AddI ||
1640 !incr2->in(2)->is_Con())
1641 continue;
1642
1643 // Check for parallel induction variable (parallel to trip counter)
1644 // via an affine function. In particular, count-down loops with
1645 // count-up array indices are common. We only RCE references off
1646 // the trip-counter, so we need to convert all these to trip-counter
1647 // expressions.
1648 Node *init2 = phi2->in( LoopNode::EntryControl );
1649 int stride_con2 = incr2->in(2)->get_int();
1650
1651 // The general case here gets a little tricky. We want to find the
1652 // GCD of all possible parallel IV's and make a new IV using this
1653 // GCD for the loop. Then all possible IVs are simple multiples of
1654 // the GCD. In practice, this will cover very few extra loops.
1655 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1656 // where +/-1 is the common case, but other integer multiples are
1657 // also easy to handle.
1658 int ratio_con = stride_con2/stride_con;
1659
1660 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1661 #ifndef PRODUCT
1662 if (TraceLoopOpts) {
1663 tty->print("Parallel IV: %d ", phi2->_idx);
1664 loop->dump_head();
1665 }
1666 #endif
1667 // Convert to using the trip counter. The parallel induction
1668 // variable differs from the trip counter by a loop-invariant
1669 // amount, the difference between their respective initial values.
1670 // It is scaled by the 'ratio_con'.
1671 Node* ratio = _igvn.intcon(ratio_con);
1672 set_ctrl(ratio, C->root());
1673 Node* ratio_init = new (C, 3) MulINode(init, ratio);
1674 _igvn.register_new_node_with_optimizer(ratio_init, init);
1675 set_early_ctrl(ratio_init);
1676 Node* diff = new (C, 3) SubINode(init2, ratio_init);
1677 _igvn.register_new_node_with_optimizer(diff, init2);
1678 set_early_ctrl(diff);
1679 Node* ratio_idx = new (C, 3) MulINode(phi, ratio);
1680 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1681 set_ctrl(ratio_idx, cl);
1682 Node* add = new (C, 3) AddINode(ratio_idx, diff);
1683 _igvn.register_new_node_with_optimizer(add);
1684 set_ctrl(add, cl);
1685 _igvn.replace_node( phi2, add );
1686 // Sometimes an induction variable is unused
1687 if (add->outcnt() == 0) {
1688 _igvn.remove_dead_node(add);
1689 }
1690 --i; // deleted this phi; rescan starting with next position
1691 continue;
1692 }
1693 }
1694 }
1695
1696 //------------------------------counted_loop-----------------------------------
1697 // Convert to counted loops where possible
1698 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1699
1700 // For grins, set the inner-loop flag here
1701 if (!_child) {
1702 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1703 }
1704
1705 if (_head->is_CountedLoop() ||
1706 phase->is_counted_loop(_head, this)) {
1707 _has_sfpt = 1; // Indicate we do not need a safepoint here
1708
1709 // Look for a safepoint to remove
1710 for (Node* n = tail(); n != _head; n = phase->idom(n))
1711 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1712 phase->is_deleteable_safept(n))
1713 phase->lazy_replace(n,n->in(TypeFunc::Control));
1714
1715 // Look for induction variables
1716 phase->replace_parallel_iv(this);
1717
1718 } else if (_parent != NULL && !_irreducible) {
1719 // Not a counted loop.
1720 // Look for a safepoint on the idom-path to remove, preserving the first one
1721 bool found = false;
1722 Node* n = tail();
1723 for (; n != _head && !found; n = phase->idom(n)) {
1724 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1725 found = true; // Found one
1726 }
1727 // Skip past it and delete the others
1728 for (; n != _head; n = phase->idom(n)) {
1729 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1730 phase->is_deleteable_safept(n))
1731 phase->lazy_replace(n,n->in(TypeFunc::Control));
1732 }
1733 }
1734
1735 // Recursively
1736 if (_child) _child->counted_loop( phase );
1737 if (_next) _next ->counted_loop( phase );
1738 }
1739
1740 #ifndef PRODUCT
1741 //------------------------------dump_head--------------------------------------
1742 // Dump 1 liner for loop header info
1743 void IdealLoopTree::dump_head( ) const {
1744 for (uint i=0; i<_nest; i++)
1745 tty->print(" ");
1746 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1747 if (_irreducible) tty->print(" IRREDUCIBLE");
1748 Node* entry = _head->in(LoopNode::EntryControl);
1749 if (LoopLimitCheck) {
1750 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1751 if (predicate != NULL ) {
1752 tty->print(" limit_check");
1753 entry = entry->in(0)->in(0);
1754 }
1755 }
1756 if (UseLoopPredicate) {
1757 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1758 if (entry != NULL) {
1759 tty->print(" predicated");
1760 }
1761 }
1762 if (_head->is_CountedLoop()) {
1763 CountedLoopNode *cl = _head->as_CountedLoop();
1764 tty->print(" counted");
1765
1766 Node* init_n = cl->init_trip();
1767 if (init_n != NULL && init_n->is_Con())
1768 tty->print(" [%d,", cl->init_trip()->get_int());
1769 else
1770 tty->print(" [int,");
1771 Node* limit_n = cl->limit();
1772 if (limit_n != NULL && limit_n->is_Con())
1773 tty->print("%d),", cl->limit()->get_int());
1774 else
1775 tty->print("int),");
1776 int stride_con = cl->stride_con();
1777 if (stride_con > 0) tty->print("+");
1778 tty->print("%d", stride_con);
1779
1780 if (cl->is_pre_loop ()) tty->print(" pre" );
1781 if (cl->is_main_loop()) tty->print(" main");
1782 if (cl->is_post_loop()) tty->print(" post");
1783 }
1784 tty->cr();
1785 }
1786
1787 //------------------------------dump-------------------------------------------
1788 // Dump loops by loop tree
1789 void IdealLoopTree::dump( ) const {
1790 dump_head();
1791 if (_child) _child->dump();
1792 if (_next) _next ->dump();
1793 }
1794
1795 #endif
1796
1797 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1798 if (loop == root) {
1799 if (loop->_child != NULL) {
1800 log->begin_head("loop_tree");
1801 log->end_head();
1802 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1803 log->tail("loop_tree");
1804 assert(loop->_next == NULL, "what?");
1805 }
1806 } else {
1807 Node* head = loop->_head;
1808 log->begin_head("loop");
1809 log->print(" idx='%d' ", head->_idx);
1810 if (loop->_irreducible) log->print("irreducible='1' ");
1811 if (head->is_Loop()) {
1812 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1813 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1814 }
1815 if (head->is_CountedLoop()) {
1816 CountedLoopNode* cl = head->as_CountedLoop();
1817 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1818 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1819 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1820 }
1821 log->end_head();
1822 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1823 log->tail("loop");
1824 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1825 }
1826 }
1827
1828 //---------------------collect_potentially_useful_predicates-----------------------
1829 // Helper function to collect potentially useful predicates to prevent them from
1830 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1831 void PhaseIdealLoop::collect_potentially_useful_predicates(
1832 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1833 if (loop->_child) { // child
1834 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1835 }
1836
1837 // self (only loops that we can apply loop predication may use their predicates)
1838 if (loop->_head->is_Loop() &&
1839 !loop->_irreducible &&
1840 !loop->tail()->is_top()) {
1841 LoopNode* lpn = loop->_head->as_Loop();
1842 Node* entry = lpn->in(LoopNode::EntryControl);
1843 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1844 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1845 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1846 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1847 entry = entry->in(0)->in(0);
1848 }
1849 predicate_proj = find_predicate(entry); // Predicate
1850 if (predicate_proj != NULL ) {
1851 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1852 }
1853 }
1854
1855 if (loop->_next) { // sibling
1856 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1857 }
1858 }
1859
1860 //------------------------eliminate_useless_predicates-----------------------------
1861 // Eliminate all inserted predicates if they could not be used by loop predication.
1862 // Note: it will also eliminates loop limits check predicate since it also uses
1863 // Opaque1 node (see Parse::add_predicate()).
1864 void PhaseIdealLoop::eliminate_useless_predicates() {
1865 if (C->predicate_count() == 0)
1866 return; // no predicate left
1867
1868 Unique_Node_List useful_predicates; // to store useful predicates
1869 if (C->has_loops()) {
1870 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1871 }
1872
1873 for (int i = C->predicate_count(); i > 0; i--) {
1874 Node * n = C->predicate_opaque1_node(i-1);
1875 assert(n->Opcode() == Op_Opaque1, "must be");
1876 if (!useful_predicates.member(n)) { // not in the useful list
1877 _igvn.replace_node(n, n->in(1));
1878 }
1879 }
1880 }
1881
1882 //=============================================================================
1883 //----------------------------build_and_optimize-------------------------------
1884 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1885 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1886 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs) {
1887 ResourceMark rm;
1888
1889 int old_progress = C->major_progress();
1890 uint orig_worklist_size = _igvn._worklist.size();
1891
1892 // Reset major-progress flag for the driver's heuristics
1893 C->clear_major_progress();
1894
1895 #ifndef PRODUCT
1896 // Capture for later assert
1897 uint unique = C->unique();
1898 _loop_invokes++;
1899 _loop_work += unique;
1900 #endif
1901
1902 // True if the method has at least 1 irreducible loop
1903 _has_irreducible_loops = false;
1904
1905 _created_loop_node = false;
1906
1907 Arena *a = Thread::current()->resource_area();
1908 VectorSet visited(a);
1909 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1910 _nodes.map(C->unique(), NULL);
1911 memset(_nodes.adr(), 0, wordSize * C->unique());
1912
1913 // Pre-build the top-level outermost loop tree entry
1914 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1915 // Do not need a safepoint at the top level
1916 _ltree_root->_has_sfpt = 1;
1917
1918 // Initialize Dominators.
1919 // Checked in clone_loop_predicate() during beautify_loops().
1920 _idom_size = 0;
1921 _idom = NULL;
1922 _dom_depth = NULL;
1923 _dom_stk = NULL;
1924
1925 // Empty pre-order array
1926 allocate_preorders();
1927
1928 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1929 // IdealLoopTree entries. Data nodes are NOT walked.
1930 build_loop_tree();
1931 // Check for bailout, and return
1932 if (C->failing()) {
1933 return;
1934 }
1935
1936 // No loops after all
1937 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
1938
1939 // There should always be an outer loop containing the Root and Return nodes.
1940 // If not, we have a degenerate empty program. Bail out in this case.
1941 if (!has_node(C->root())) {
1942 if (!_verify_only) {
1943 C->clear_major_progress();
1944 C->record_method_not_compilable("empty program detected during loop optimization");
1945 }
1946 return;
1947 }
1948
1949 // Nothing to do, so get out
1950 if( !C->has_loops() && !do_split_ifs && !_verify_me && !_verify_only ) {
1951 _igvn.optimize(); // Cleanup NeverBranches
1952 return;
1953 }
1954
1955 // Set loop nesting depth
1956 _ltree_root->set_nest( 0 );
1957
1958 // Split shared headers and insert loop landing pads.
1959 // Do not bother doing this on the Root loop of course.
1960 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
1961 C->print_method("Before beautify loops", 3);
1962 if( _ltree_root->_child->beautify_loops( this ) ) {
1963 // Re-build loop tree!
1964 _ltree_root->_child = NULL;
1965 _nodes.clear();
1966 reallocate_preorders();
1967 build_loop_tree();
1968 // Check for bailout, and return
1969 if (C->failing()) {
1970 return;
1971 }
1972 // Reset loop nesting depth
1973 _ltree_root->set_nest( 0 );
1974
1975 C->print_method("After beautify loops", 3);
1976 }
1977 }
1978
1979 // Build Dominators for elision of NULL checks & loop finding.
1980 // Since nodes do not have a slot for immediate dominator, make
1981 // a persistent side array for that info indexed on node->_idx.
1982 _idom_size = C->unique();
1983 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1984 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1985 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1986 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1987
1988 Dominators();
1989
1990 if (!_verify_only) {
1991 // As a side effect, Dominators removed any unreachable CFG paths
1992 // into RegionNodes. It doesn't do this test against Root, so
1993 // we do it here.
1994 for( uint i = 1; i < C->root()->req(); i++ ) {
1995 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1996 _igvn.hash_delete(C->root());
1997 C->root()->del_req(i);
1998 _igvn._worklist.push(C->root());
1999 i--; // Rerun same iteration on compressed edges
2000 }
2001 }
2002
2003 // Given dominators, try to find inner loops with calls that must
2004 // always be executed (call dominates loop tail). These loops do
2005 // not need a separate safepoint.
2006 Node_List cisstack(a);
2007 _ltree_root->check_safepts(visited, cisstack);
2008 }
2009
2010 // Walk the DATA nodes and place into loops. Find earliest control
2011 // node. For CFG nodes, the _nodes array starts out and remains
2012 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2013 // _nodes array holds the earliest legal controlling CFG node.
2014
2015 // Allocate stack with enough space to avoid frequent realloc
2016 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
2017 Node_Stack nstack( a, stack_size );
2018
2019 visited.Clear();
2020 Node_List worklist(a);
2021 // Don't need C->root() on worklist since
2022 // it will be processed among C->top() inputs
2023 worklist.push( C->top() );
2024 visited.set( C->top()->_idx ); // Set C->top() as visited now
2025 build_loop_early( visited, worklist, nstack );
2026
2027 // Given early legal placement, try finding counted loops. This placement
2028 // is good enough to discover most loop invariants.
2029 if( !_verify_me && !_verify_only )
2030 _ltree_root->counted_loop( this );
2031
2032 // Find latest loop placement. Find ideal loop placement.
2033 visited.Clear();
2034 init_dom_lca_tags();
2035 // Need C->root() on worklist when processing outs
2036 worklist.push( C->root() );
2037 NOT_PRODUCT( C->verify_graph_edges(); )
2038 worklist.push( C->top() );
2039 build_loop_late( visited, worklist, nstack );
2040
2041 if (_verify_only) {
2042 // restore major progress flag
2043 for (int i = 0; i < old_progress; i++)
2044 C->set_major_progress();
2045 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2046 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2047 return;
2048 }
2049
2050 // Some parser-inserted loop predicates could never be used by loop
2051 // predication or they were moved away from loop during some optimizations.
2052 // For example, peeling. Eliminate them before next loop optimizations.
2053 if (UseLoopPredicate || LoopLimitCheck) {
2054 eliminate_useless_predicates();
2055 }
2056
2057 // clear out the dead code
2058 while(_deadlist.size()) {
2059 _igvn.remove_globally_dead_node(_deadlist.pop());
2060 }
2061
2062 #ifndef PRODUCT
2063 C->verify_graph_edges();
2064 if (_verify_me) { // Nested verify pass?
2065 // Check to see if the verify mode is broken
2066 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2067 return;
2068 }
2069 if(VerifyLoopOptimizations) verify();
2070 if(TraceLoopOpts && C->has_loops()) {
2071 _ltree_root->dump();
2072 }
2073 #endif
2074
2075 if (ReassociateInvariants) {
2076 // Reassociate invariants and prep for split_thru_phi
2077 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2078 IdealLoopTree* lpt = iter.current();
2079 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2080
2081 lpt->reassociate_invariants(this);
2082
2083 // Because RCE opportunities can be masked by split_thru_phi,
2084 // look for RCE candidates and inhibit split_thru_phi
2085 // on just their loop-phi's for this pass of loop opts
2086 if (SplitIfBlocks && do_split_ifs) {
2087 if (lpt->policy_range_check(this)) {
2088 lpt->_rce_candidate = 1; // = true
2089 }
2090 }
2091 }
2092 }
2093
2094 // Check for aggressive application of split-if and other transforms
2095 // that require basic-block info (like cloning through Phi's)
2096 if( SplitIfBlocks && do_split_ifs ) {
2097 visited.Clear();
2098 split_if_with_blocks( visited, nstack );
2099 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2100 }
2101
2102 // Perform loop predication before iteration splitting
2103 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2104 _ltree_root->_child->loop_predication(this);
2105 }
2106
2107 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2108 if (do_intrinsify_fill()) {
2109 C->set_major_progress();
2110 }
2111 }
2112
2113 // Perform iteration-splitting on inner loops. Split iterations to avoid
2114 // range checks or one-shot null checks.
2115
2116 // If split-if's didn't hack the graph too bad (no CFG changes)
2117 // then do loop opts.
2118 if (C->has_loops() && !C->major_progress()) {
2119 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2120 _ltree_root->_child->iteration_split( this, worklist );
2121 // No verify after peeling! GCM has hoisted code out of the loop.
2122 // After peeling, the hoisted code could sink inside the peeled area.
2123 // The peeling code does not try to recompute the best location for
2124 // all the code before the peeled area, so the verify pass will always
2125 // complain about it.
2126 }
2127 // Do verify graph edges in any case
2128 NOT_PRODUCT( C->verify_graph_edges(); );
2129
2130 if (!do_split_ifs) {
2131 // We saw major progress in Split-If to get here. We forced a
2132 // pass with unrolling and not split-if, however more split-if's
2133 // might make progress. If the unrolling didn't make progress
2134 // then the major-progress flag got cleared and we won't try
2135 // another round of Split-If. In particular the ever-common
2136 // instance-of/check-cast pattern requires at least 2 rounds of
2137 // Split-If to clear out.
2138 C->set_major_progress();
2139 }
2140
2141 // Repeat loop optimizations if new loops were seen
2142 if (created_loop_node()) {
2143 C->set_major_progress();
2144 }
2145
2146 // Keep loop predicates and perform optimizations with them
2147 // until no more loop optimizations could be done.
2148 // After that switch predicates off and do more loop optimizations.
2149 if (!C->major_progress() && (C->predicate_count() > 0)) {
2150 C->cleanup_loop_predicates(_igvn);
2151 #ifndef PRODUCT
2152 if (TraceLoopOpts) {
2153 tty->print_cr("PredicatesOff");
2154 }
2155 #endif
2156 C->set_major_progress();
2157 }
2158
2159 // Convert scalar to superword operations at the end of all loop opts.
2160 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2161 // SuperWord transform
2162 SuperWord sw(this);
2163 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2164 IdealLoopTree* lpt = iter.current();
2165 if (lpt->is_counted()) {
2166 sw.transform_loop(lpt);
2167 }
2168 }
2169 }
2170
2171 // Cleanup any modified bits
2172 _igvn.optimize();
2173
2174 // disable assert until issue with split_flow_path is resolved (6742111)
2175 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2176 // "shouldn't introduce irreducible loops");
2177
2178 if (C->log() != NULL) {
2179 log_loop_tree(_ltree_root, _ltree_root, C->log());
2180 }
2181 }
2182
2183 #ifndef PRODUCT
2184 //------------------------------print_statistics-------------------------------
2185 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2186 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2187 void PhaseIdealLoop::print_statistics() {
2188 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2189 }
2190
2191 //------------------------------verify-----------------------------------------
2192 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2193 static int fail; // debug only, so its multi-thread dont care
2194 void PhaseIdealLoop::verify() const {
2195 int old_progress = C->major_progress();
2196 ResourceMark rm;
2197 PhaseIdealLoop loop_verify( _igvn, this );
2198 VectorSet visited(Thread::current()->resource_area());
2199
2200 fail = 0;
2201 verify_compare( C->root(), &loop_verify, visited );
2202 assert( fail == 0, "verify loops failed" );
2203 // Verify loop structure is the same
2204 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2205 // Reset major-progress. It was cleared by creating a verify version of
2206 // PhaseIdealLoop.
2207 for( int i=0; i<old_progress; i++ )
2208 C->set_major_progress();
2209 }
2210
2211 //------------------------------verify_compare---------------------------------
2212 // Make sure me and the given PhaseIdealLoop agree on key data structures
2213 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2214 if( !n ) return;
2215 if( visited.test_set( n->_idx ) ) return;
2216 if( !_nodes[n->_idx] ) { // Unreachable
2217 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2218 return;
2219 }
2220
2221 uint i;
2222 for( i = 0; i < n->req(); i++ )
2223 verify_compare( n->in(i), loop_verify, visited );
2224
2225 // Check the '_nodes' block/loop structure
2226 i = n->_idx;
2227 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2228 if( _nodes[i] != loop_verify->_nodes[i] &&
2229 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2230 tty->print("Mismatched control setting for: ");
2231 n->dump();
2232 if( fail++ > 10 ) return;
2233 Node *c = get_ctrl_no_update(n);
2234 tty->print("We have it as: ");
2235 if( c->in(0) ) c->dump();
2236 else tty->print_cr("N%d",c->_idx);
2237 tty->print("Verify thinks: ");
2238 if( loop_verify->has_ctrl(n) )
2239 loop_verify->get_ctrl_no_update(n)->dump();
2240 else
2241 loop_verify->get_loop_idx(n)->dump();
2242 tty->cr();
2243 }
2244 } else { // We have a loop
2245 IdealLoopTree *us = get_loop_idx(n);
2246 if( loop_verify->has_ctrl(n) ) {
2247 tty->print("Mismatched loop setting for: ");
2248 n->dump();
2249 if( fail++ > 10 ) return;
2250 tty->print("We have it as: ");
2251 us->dump();
2252 tty->print("Verify thinks: ");
2253 loop_verify->get_ctrl_no_update(n)->dump();
2254 tty->cr();
2255 } else if (!C->major_progress()) {
2256 // Loop selection can be messed up if we did a major progress
2257 // operation, like split-if. Do not verify in that case.
2258 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2259 if( us->_head != them->_head || us->_tail != them->_tail ) {
2260 tty->print("Unequals loops for: ");
2261 n->dump();
2262 if( fail++ > 10 ) return;
2263 tty->print("We have it as: ");
2264 us->dump();
2265 tty->print("Verify thinks: ");
2266 them->dump();
2267 tty->cr();
2268 }
2269 }
2270 }
2271
2272 // Check for immediate dominators being equal
2273 if( i >= _idom_size ) {
2274 if( !n->is_CFG() ) return;
2275 tty->print("CFG Node with no idom: ");
2276 n->dump();
2277 return;
2278 }
2279 if( !n->is_CFG() ) return;
2280 if( n == C->root() ) return; // No IDOM here
2281
2282 assert(n->_idx == i, "sanity");
2283 Node *id = idom_no_update(n);
2284 if( id != loop_verify->idom_no_update(n) ) {
2285 tty->print("Unequals idoms for: ");
2286 n->dump();
2287 if( fail++ > 10 ) return;
2288 tty->print("We have it as: ");
2289 id->dump();
2290 tty->print("Verify thinks: ");
2291 loop_verify->idom_no_update(n)->dump();
2292 tty->cr();
2293 }
2294
2295 }
2296
2297 //------------------------------verify_tree------------------------------------
2298 // Verify that tree structures match. Because the CFG can change, siblings
2299 // within the loop tree can be reordered. We attempt to deal with that by
2300 // reordering the verify's loop tree if possible.
2301 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2302 assert( _parent == parent, "Badly formed loop tree" );
2303
2304 // Siblings not in same order? Attempt to re-order.
2305 if( _head != loop->_head ) {
2306 // Find _next pointer to update
2307 IdealLoopTree **pp = &loop->_parent->_child;
2308 while( *pp != loop )
2309 pp = &((*pp)->_next);
2310 // Find proper sibling to be next
2311 IdealLoopTree **nn = &loop->_next;
2312 while( (*nn) && (*nn)->_head != _head )
2313 nn = &((*nn)->_next);
2314
2315 // Check for no match.
2316 if( !(*nn) ) {
2317 // Annoyingly, irreducible loops can pick different headers
2318 // after a major_progress operation, so the rest of the loop
2319 // tree cannot be matched.
2320 if (_irreducible && Compile::current()->major_progress()) return;
2321 assert( 0, "failed to match loop tree" );
2322 }
2323
2324 // Move (*nn) to (*pp)
2325 IdealLoopTree *hit = *nn;
2326 *nn = hit->_next;
2327 hit->_next = loop;
2328 *pp = loop;
2329 loop = hit;
2330 // Now try again to verify
2331 }
2332
2333 assert( _head == loop->_head , "mismatched loop head" );
2334 Node *tail = _tail; // Inline a non-updating version of
2335 while( !tail->in(0) ) // the 'tail()' call.
2336 tail = tail->in(1);
2337 assert( tail == loop->_tail, "mismatched loop tail" );
2338
2339 // Counted loops that are guarded should be able to find their guards
2340 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2341 CountedLoopNode *cl = _head->as_CountedLoop();
2342 Node *init = cl->init_trip();
2343 Node *ctrl = cl->in(LoopNode::EntryControl);
2344 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2345 Node *iff = ctrl->in(0);
2346 assert( iff->Opcode() == Op_If, "" );
2347 Node *bol = iff->in(1);
2348 assert( bol->Opcode() == Op_Bool, "" );
2349 Node *cmp = bol->in(1);
2350 assert( cmp->Opcode() == Op_CmpI, "" );
2351 Node *add = cmp->in(1);
2352 Node *opaq;
2353 if( add->Opcode() == Op_Opaque1 ) {
2354 opaq = add;
2355 } else {
2356 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2357 assert( add == init, "" );
2358 opaq = cmp->in(2);
2359 }
2360 assert( opaq->Opcode() == Op_Opaque1, "" );
2361
2362 }
2363
2364 if (_child != NULL) _child->verify_tree(loop->_child, this);
2365 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2366 // Innermost loops need to verify loop bodies,
2367 // but only if no 'major_progress'
2368 int fail = 0;
2369 if (!Compile::current()->major_progress() && _child == NULL) {
2370 for( uint i = 0; i < _body.size(); i++ ) {
2371 Node *n = _body.at(i);
2372 if (n->outcnt() == 0) continue; // Ignore dead
2373 uint j;
2374 for( j = 0; j < loop->_body.size(); j++ )
2375 if( loop->_body.at(j) == n )
2376 break;
2377 if( j == loop->_body.size() ) { // Not found in loop body
2378 // Last ditch effort to avoid assertion: Its possible that we
2379 // have some users (so outcnt not zero) but are still dead.
2380 // Try to find from root.
2381 if (Compile::current()->root()->find(n->_idx)) {
2382 fail++;
2383 tty->print("We have that verify does not: ");
2384 n->dump();
2385 }
2386 }
2387 }
2388 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2389 Node *n = loop->_body.at(i2);
2390 if (n->outcnt() == 0) continue; // Ignore dead
2391 uint j;
2392 for( j = 0; j < _body.size(); j++ )
2393 if( _body.at(j) == n )
2394 break;
2395 if( j == _body.size() ) { // Not found in loop body
2396 // Last ditch effort to avoid assertion: Its possible that we
2397 // have some users (so outcnt not zero) but are still dead.
2398 // Try to find from root.
2399 if (Compile::current()->root()->find(n->_idx)) {
2400 fail++;
2401 tty->print("Verify has that we do not: ");
2402 n->dump();
2403 }
2404 }
2405 }
2406 assert( !fail, "loop body mismatch" );
2407 }
2408 }
2409
2410 #endif
2411
2412 //------------------------------set_idom---------------------------------------
2413 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2414 uint idx = d->_idx;
2415 if (idx >= _idom_size) {
2416 uint newsize = _idom_size<<1;
2417 while( idx >= newsize ) {
2418 newsize <<= 1;
2419 }
2420 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2421 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2422 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2423 _idom_size = newsize;
2424 }
2425 _idom[idx] = n;
2426 _dom_depth[idx] = dom_depth;
2427 }
2428
2429 //------------------------------recompute_dom_depth---------------------------------------
2430 // The dominator tree is constructed with only parent pointers.
2431 // This recomputes the depth in the tree by first tagging all
2432 // nodes as "no depth yet" marker. The next pass then runs up
2433 // the dom tree from each node marked "no depth yet", and computes
2434 // the depth on the way back down.
2435 void PhaseIdealLoop::recompute_dom_depth() {
2436 uint no_depth_marker = C->unique();
2437 uint i;
2438 // Initialize depth to "no depth yet"
2439 for (i = 0; i < _idom_size; i++) {
2440 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2441 _dom_depth[i] = no_depth_marker;
2442 }
2443 }
2444 if (_dom_stk == NULL) {
2445 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
2446 if (init_size < 10) init_size = 10;
2447 _dom_stk = new GrowableArray<uint>(init_size);
2448 }
2449 // Compute new depth for each node.
2450 for (i = 0; i < _idom_size; i++) {
2451 uint j = i;
2452 // Run up the dom tree to find a node with a depth
2453 while (_dom_depth[j] == no_depth_marker) {
2454 _dom_stk->push(j);
2455 j = _idom[j]->_idx;
2456 }
2457 // Compute the depth on the way back down this tree branch
2458 uint dd = _dom_depth[j] + 1;
2459 while (_dom_stk->length() > 0) {
2460 uint j = _dom_stk->pop();
2461 _dom_depth[j] = dd;
2462 dd++;
2463 }
2464 }
2465 }
2466
2467 //------------------------------sort-------------------------------------------
2468 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2469 // loop tree, not the root.
2470 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2471 if( !innermost ) return loop; // New innermost loop
2472
2473 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2474 assert( loop_preorder, "not yet post-walked loop" );
2475 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2476 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2477
2478 // Insert at start of list
2479 while( l ) { // Insertion sort based on pre-order
2480 if( l == loop ) return innermost; // Already on list!
2481 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2482 assert( l_preorder, "not yet post-walked l" );
2483 // Check header pre-order number to figure proper nesting
2484 if( loop_preorder > l_preorder )
2485 break; // End of insertion
2486 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2487 // Since I split shared headers, you'd think this could not happen.
2488 // BUT: I must first do the preorder numbering before I can discover I
2489 // have shared headers, so the split headers all get the same preorder
2490 // number as the RegionNode they split from.
2491 if( loop_preorder == l_preorder &&
2492 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2493 break; // Also check for shared headers (same pre#)
2494 pp = &l->_parent; // Chain up list
2495 l = *pp;
2496 }
2497 // Link into list
2498 // Point predecessor to me
2499 *pp = loop;
2500 // Point me to successor
2501 IdealLoopTree *p = loop->_parent;
2502 loop->_parent = l; // Point me to successor
2503 if( p ) sort( p, innermost ); // Insert my parents into list as well
2504 return innermost;
2505 }
2506
2507 //------------------------------build_loop_tree--------------------------------
2508 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2509 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2510 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2511 // tightest enclosing IdealLoopTree for post-walked.
2512 //
2513 // During my forward walk I do a short 1-layer lookahead to see if I can find
2514 // a loop backedge with that doesn't have any work on the backedge. This
2515 // helps me construct nested loops with shared headers better.
2516 //
2517 // Once I've done the forward recursion, I do the post-work. For each child
2518 // I check to see if there is a backedge. Backedges define a loop! I
2519 // insert an IdealLoopTree at the target of the backedge.
2520 //
2521 // During the post-work I also check to see if I have several children
2522 // belonging to different loops. If so, then this Node is a decision point
2523 // where control flow can choose to change loop nests. It is at this
2524 // decision point where I can figure out how loops are nested. At this
2525 // time I can properly order the different loop nests from my children.
2526 // Note that there may not be any backedges at the decision point!
2527 //
2528 // Since the decision point can be far removed from the backedges, I can't
2529 // order my loops at the time I discover them. Thus at the decision point
2530 // I need to inspect loop header pre-order numbers to properly nest my
2531 // loops. This means I need to sort my childrens' loops by pre-order.
2532 // The sort is of size number-of-control-children, which generally limits
2533 // it to size 2 (i.e., I just choose between my 2 target loops).
2534 void PhaseIdealLoop::build_loop_tree() {
2535 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2536 GrowableArray <Node *> bltstack(C->unique() >> 1);
2537 Node *n = C->root();
2538 bltstack.push(n);
2539 int pre_order = 1;
2540 int stack_size;
2541
2542 while ( ( stack_size = bltstack.length() ) != 0 ) {
2543 n = bltstack.top(); // Leave node on stack
2544 if ( !is_visited(n) ) {
2545 // ---- Pre-pass Work ----
2546 // Pre-walked but not post-walked nodes need a pre_order number.
2547
2548 set_preorder_visited( n, pre_order ); // set as visited
2549
2550 // ---- Scan over children ----
2551 // Scan first over control projections that lead to loop headers.
2552 // This helps us find inner-to-outer loops with shared headers better.
2553
2554 // Scan children's children for loop headers.
2555 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2556 Node* m = n->raw_out(i); // Child
2557 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2558 // Scan over children's children to find loop
2559 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2560 Node* l = m->fast_out(j);
2561 if( is_visited(l) && // Been visited?
2562 !is_postvisited(l) && // But not post-visited
2563 get_preorder(l) < pre_order ) { // And smaller pre-order
2564 // Found! Scan the DFS down this path before doing other paths
2565 bltstack.push(m);
2566 break;
2567 }
2568 }
2569 }
2570 }
2571 pre_order++;
2572 }
2573 else if ( !is_postvisited(n) ) {
2574 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2575 // such as com.sun.rsasign.am::a.
2576 // For non-recursive version, first, process current children.
2577 // On next iteration, check if additional children were added.
2578 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2579 Node* u = n->raw_out(k);
2580 if ( u->is_CFG() && !is_visited(u) ) {
2581 bltstack.push(u);
2582 }
2583 }
2584 if ( bltstack.length() == stack_size ) {
2585 // There were no additional children, post visit node now
2586 (void)bltstack.pop(); // Remove node from stack
2587 pre_order = build_loop_tree_impl( n, pre_order );
2588 // Check for bailout
2589 if (C->failing()) {
2590 return;
2591 }
2592 // Check to grow _preorders[] array for the case when
2593 // build_loop_tree_impl() adds new nodes.
2594 check_grow_preorders();
2595 }
2596 }
2597 else {
2598 (void)bltstack.pop(); // Remove post-visited node from stack
2599 }
2600 }
2601 }
2602
2603 //------------------------------build_loop_tree_impl---------------------------
2604 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2605 // ---- Post-pass Work ----
2606 // Pre-walked but not post-walked nodes need a pre_order number.
2607
2608 // Tightest enclosing loop for this Node
2609 IdealLoopTree *innermost = NULL;
2610
2611 // For all children, see if any edge is a backedge. If so, make a loop
2612 // for it. Then find the tightest enclosing loop for the self Node.
2613 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2614 Node* m = n->fast_out(i); // Child
2615 if( n == m ) continue; // Ignore control self-cycles
2616 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2617
2618 IdealLoopTree *l; // Child's loop
2619 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2620 // Found a backedge
2621 assert( get_preorder(m) < pre_order, "should be backedge" );
2622 // Check for the RootNode, which is already a LoopNode and is allowed
2623 // to have multiple "backedges".
2624 if( m == C->root()) { // Found the root?
2625 l = _ltree_root; // Root is the outermost LoopNode
2626 } else { // Else found a nested loop
2627 // Insert a LoopNode to mark this loop.
2628 l = new IdealLoopTree(this, m, n);
2629 } // End of Else found a nested loop
2630 if( !has_loop(m) ) // If 'm' does not already have a loop set
2631 set_loop(m, l); // Set loop header to loop now
2632
2633 } else { // Else not a nested loop
2634 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2635 l = get_loop(m); // Get previously determined loop
2636 // If successor is header of a loop (nest), move up-loop till it
2637 // is a member of some outer enclosing loop. Since there are no
2638 // shared headers (I've split them already) I only need to go up
2639 // at most 1 level.
2640 while( l && l->_head == m ) // Successor heads loop?
2641 l = l->_parent; // Move up 1 for me
2642 // If this loop is not properly parented, then this loop
2643 // has no exit path out, i.e. its an infinite loop.
2644 if( !l ) {
2645 // Make loop "reachable" from root so the CFG is reachable. Basically
2646 // insert a bogus loop exit that is never taken. 'm', the loop head,
2647 // points to 'n', one (of possibly many) fall-in paths. There may be
2648 // many backedges as well.
2649
2650 // Here I set the loop to be the root loop. I could have, after
2651 // inserting a bogus loop exit, restarted the recursion and found my
2652 // new loop exit. This would make the infinite loop a first-class
2653 // loop and it would then get properly optimized. What's the use of
2654 // optimizing an infinite loop?
2655 l = _ltree_root; // Oops, found infinite loop
2656
2657 if (!_verify_only) {
2658 // Insert the NeverBranch between 'm' and it's control user.
2659 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2660 _igvn.register_new_node_with_optimizer(iff);
2661 set_loop(iff, l);
2662 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2663 _igvn.register_new_node_with_optimizer(if_t);
2664 set_loop(if_t, l);
2665
2666 Node* cfg = NULL; // Find the One True Control User of m
2667 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2668 Node* x = m->fast_out(j);
2669 if (x->is_CFG() && x != m && x != iff)
2670 { cfg = x; break; }
2671 }
2672 assert(cfg != NULL, "must find the control user of m");
2673 uint k = 0; // Probably cfg->in(0)
2674 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2675 cfg->set_req( k, if_t ); // Now point to NeverBranch
2676
2677 // Now create the never-taken loop exit
2678 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2679 _igvn.register_new_node_with_optimizer(if_f);
2680 set_loop(if_f, l);
2681 // Find frame ptr for Halt. Relies on the optimizer
2682 // V-N'ing. Easier and quicker than searching through
2683 // the program structure.
2684 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2685 _igvn.register_new_node_with_optimizer(frame);
2686 // Halt & Catch Fire
2687 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2688 _igvn.register_new_node_with_optimizer(halt);
2689 set_loop(halt, l);
2690 C->root()->add_req(halt);
2691 }
2692 set_loop(C->root(), _ltree_root);
2693 }
2694 }
2695 // Weeny check for irreducible. This child was already visited (this
2696 // IS the post-work phase). Is this child's loop header post-visited
2697 // as well? If so, then I found another entry into the loop.
2698 if (!_verify_only) {
2699 while( is_postvisited(l->_head) ) {
2700 // found irreducible
2701 l->_irreducible = 1; // = true
2702 l = l->_parent;
2703 _has_irreducible_loops = true;
2704 // Check for bad CFG here to prevent crash, and bailout of compile
2705 if (l == NULL) {
2706 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2707 return pre_order;
2708 }
2709 }
2710 }
2711
2712 // This Node might be a decision point for loops. It is only if
2713 // it's children belong to several different loops. The sort call
2714 // does a trivial amount of work if there is only 1 child or all
2715 // children belong to the same loop. If however, the children
2716 // belong to different loops, the sort call will properly set the
2717 // _parent pointers to show how the loops nest.
2718 //
2719 // In any case, it returns the tightest enclosing loop.
2720 innermost = sort( l, innermost );
2721 }
2722
2723 // Def-use info will have some dead stuff; dead stuff will have no
2724 // loop decided on.
2725
2726 // Am I a loop header? If so fix up my parent's child and next ptrs.
2727 if( innermost && innermost->_head == n ) {
2728 assert( get_loop(n) == innermost, "" );
2729 IdealLoopTree *p = innermost->_parent;
2730 IdealLoopTree *l = innermost;
2731 while( p && l->_head == n ) {
2732 l->_next = p->_child; // Put self on parents 'next child'
2733 p->_child = l; // Make self as first child of parent
2734 l = p; // Now walk up the parent chain
2735 p = l->_parent;
2736 }
2737 } else {
2738 // Note that it is possible for a LoopNode to reach here, if the
2739 // backedge has been made unreachable (hence the LoopNode no longer
2740 // denotes a Loop, and will eventually be removed).
2741
2742 // Record tightest enclosing loop for self. Mark as post-visited.
2743 set_loop(n, innermost);
2744 // Also record has_call flag early on
2745 if( innermost ) {
2746 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2747 // Do not count uncommon calls
2748 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2749 Node *iff = n->in(0)->in(0);
2750 if( !iff->is_If() ||
2751 (n->in(0)->Opcode() == Op_IfFalse &&
2752 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2753 (iff->as_If()->_prob >= 0.01) )
2754 innermost->_has_call = 1;
2755 }
2756 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2757 // Disable loop optimizations if the loop has a scalar replaceable
2758 // allocation. This disabling may cause a potential performance lost
2759 // if the allocation is not eliminated for some reason.
2760 innermost->_allow_optimizations = false;
2761 innermost->_has_call = 1; // = true
2762 }
2763 }
2764 }
2765
2766 // Flag as post-visited now
2767 set_postvisited(n);
2768 return pre_order;
2769 }
2770
2771
2772 //------------------------------build_loop_early-------------------------------
2773 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2774 // First pass computes the earliest controlling node possible. This is the
2775 // controlling input with the deepest dominating depth.
2776 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2777 while (worklist.size() != 0) {
2778 // Use local variables nstack_top_n & nstack_top_i to cache values
2779 // on nstack's top.
2780 Node *nstack_top_n = worklist.pop();
2781 uint nstack_top_i = 0;
2782 //while_nstack_nonempty:
2783 while (true) {
2784 // Get parent node and next input's index from stack's top.
2785 Node *n = nstack_top_n;
2786 uint i = nstack_top_i;
2787 uint cnt = n->req(); // Count of inputs
2788 if (i == 0) { // Pre-process the node.
2789 if( has_node(n) && // Have either loop or control already?
2790 !has_ctrl(n) ) { // Have loop picked out already?
2791 // During "merge_many_backedges" we fold up several nested loops
2792 // into a single loop. This makes the members of the original
2793 // loop bodies pointing to dead loops; they need to move up
2794 // to the new UNION'd larger loop. I set the _head field of these
2795 // dead loops to NULL and the _parent field points to the owning
2796 // loop. Shades of UNION-FIND algorithm.
2797 IdealLoopTree *ilt;
2798 while( !(ilt = get_loop(n))->_head ) {
2799 // Normally I would use a set_loop here. But in this one special
2800 // case, it is legal (and expected) to change what loop a Node
2801 // belongs to.
2802 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2803 }
2804 // Remove safepoints ONLY if I've already seen I don't need one.
2805 // (the old code here would yank a 2nd safepoint after seeing a
2806 // first one, even though the 1st did not dominate in the loop body
2807 // and thus could be avoided indefinitely)
2808 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2809 is_deleteable_safept(n)) {
2810 Node *in = n->in(TypeFunc::Control);
2811 lazy_replace(n,in); // Pull safepoint now
2812 // Carry on with the recursion "as if" we are walking
2813 // only the control input
2814 if( !visited.test_set( in->_idx ) ) {
2815 worklist.push(in); // Visit this guy later, using worklist
2816 }
2817 // Get next node from nstack:
2818 // - skip n's inputs processing by setting i > cnt;
2819 // - we also will not call set_early_ctrl(n) since
2820 // has_node(n) == true (see the condition above).
2821 i = cnt + 1;
2822 }
2823 }
2824 } // if (i == 0)
2825
2826 // Visit all inputs
2827 bool done = true; // Assume all n's inputs will be processed
2828 while (i < cnt) {
2829 Node *in = n->in(i);
2830 ++i;
2831 if (in == NULL) continue;
2832 if (in->pinned() && !in->is_CFG())
2833 set_ctrl(in, in->in(0));
2834 int is_visited = visited.test_set( in->_idx );
2835 if (!has_node(in)) { // No controlling input yet?
2836 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2837 assert( !is_visited, "visit only once" );
2838 nstack.push(n, i); // Save parent node and next input's index.
2839 nstack_top_n = in; // Process current input now.
2840 nstack_top_i = 0;
2841 done = false; // Not all n's inputs processed.
2842 break; // continue while_nstack_nonempty;
2843 } else if (!is_visited) {
2844 // This guy has a location picked out for him, but has not yet
2845 // been visited. Happens to all CFG nodes, for instance.
2846 // Visit him using the worklist instead of recursion, to break
2847 // cycles. Since he has a location already we do not need to
2848 // find his location before proceeding with the current Node.
2849 worklist.push(in); // Visit this guy later, using worklist
2850 }
2851 }
2852 if (done) {
2853 // All of n's inputs have been processed, complete post-processing.
2854
2855 // Compute earliest point this Node can go.
2856 // CFG, Phi, pinned nodes already know their controlling input.
2857 if (!has_node(n)) {
2858 // Record earliest legal location
2859 set_early_ctrl( n );
2860 }
2861 if (nstack.is_empty()) {
2862 // Finished all nodes on stack.
2863 // Process next node on the worklist.
2864 break;
2865 }
2866 // Get saved parent node and next input's index.
2867 nstack_top_n = nstack.node();
2868 nstack_top_i = nstack.index();
2869 nstack.pop();
2870 }
2871 } // while (true)
2872 }
2873 }
2874
2875 //------------------------------dom_lca_internal--------------------------------
2876 // Pair-wise LCA
2877 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2878 if( !n1 ) return n2; // Handle NULL original LCA
2879 assert( n1->is_CFG(), "" );
2880 assert( n2->is_CFG(), "" );
2881 // find LCA of all uses
2882 uint d1 = dom_depth(n1);
2883 uint d2 = dom_depth(n2);
2884 while (n1 != n2) {
2885 if (d1 > d2) {
2886 n1 = idom(n1);
2887 d1 = dom_depth(n1);
2888 } else if (d1 < d2) {
2889 n2 = idom(n2);
2890 d2 = dom_depth(n2);
2891 } else {
2892 // Here d1 == d2. Due to edits of the dominator-tree, sections
2893 // of the tree might have the same depth. These sections have
2894 // to be searched more carefully.
2895
2896 // Scan up all the n1's with equal depth, looking for n2.
2897 Node *t1 = idom(n1);
2898 while (dom_depth(t1) == d1) {
2899 if (t1 == n2) return n2;
2900 t1 = idom(t1);
2901 }
2902 // Scan up all the n2's with equal depth, looking for n1.
2903 Node *t2 = idom(n2);
2904 while (dom_depth(t2) == d2) {
2905 if (t2 == n1) return n1;
2906 t2 = idom(t2);
2907 }
2908 // Move up to a new dominator-depth value as well as up the dom-tree.
2909 n1 = t1;
2910 n2 = t2;
2911 d1 = dom_depth(n1);
2912 d2 = dom_depth(n2);
2913 }
2914 }
2915 return n1;
2916 }
2917
2918 //------------------------------compute_idom-----------------------------------
2919 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2920 // IDOMs are correct.
2921 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2922 assert( region->is_Region(), "" );
2923 Node *LCA = NULL;
2924 for( uint i = 1; i < region->req(); i++ ) {
2925 if( region->in(i) != C->top() )
2926 LCA = dom_lca( LCA, region->in(i) );
2927 }
2928 return LCA;
2929 }
2930
2931 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
2932 bool had_error = false;
2933 #ifdef ASSERT
2934 if (early != C->root()) {
2935 // Make sure that there's a dominance path from use to LCA
2936 Node* d = use;
2937 while (d != LCA) {
2938 d = idom(d);
2939 if (d == C->root()) {
2940 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
2941 n->dump();
2942 use->dump();
2943 had_error = true;
2944 break;
2945 }
2946 }
2947 }
2948 #endif
2949 return had_error;
2950 }
2951
2952
2953 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
2954 // Compute LCA over list of uses
2955 bool had_error = false;
2956 Node *LCA = NULL;
2957 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2958 Node* c = n->fast_out(i);
2959 if (_nodes[c->_idx] == NULL)
2960 continue; // Skip the occasional dead node
2961 if( c->is_Phi() ) { // For Phis, we must land above on the path
2962 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2963 if( c->in(j) == n ) { // Found matching input?
2964 Node *use = c->in(0)->in(j);
2965 if (_verify_only && use->is_top()) continue;
2966 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2967 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2968 }
2969 }
2970 } else {
2971 // For CFG data-users, use is in the block just prior
2972 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2973 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2974 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2975 }
2976 }
2977 assert(!had_error, "bad dominance");
2978 return LCA;
2979 }
2980
2981 //------------------------------get_late_ctrl----------------------------------
2982 // Compute latest legal control.
2983 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2984 assert(early != NULL, "early control should not be NULL");
2985
2986 Node* LCA = compute_lca_of_uses(n, early);
2987 #ifdef ASSERT
2988 if (LCA == C->root() && LCA != early) {
2989 // def doesn't dominate uses so print some useful debugging output
2990 compute_lca_of_uses(n, early, true);
2991 }
2992 #endif
2993
2994 // if this is a load, check for anti-dependent stores
2995 // We use a conservative algorithm to identify potential interfering
2996 // instructions and for rescheduling the load. The users of the memory
2997 // input of this load are examined. Any use which is not a load and is
2998 // dominated by early is considered a potentially interfering store.
2999 // This can produce false positives.
3000 if (n->is_Load() && LCA != early) {
3001 Node_List worklist;
3002
3003 Node *mem = n->in(MemNode::Memory);
3004 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3005 Node* s = mem->fast_out(i);
3006 worklist.push(s);
3007 }
3008 while(worklist.size() != 0 && LCA != early) {
3009 Node* s = worklist.pop();
3010 if (s->is_Load()) {
3011 continue;
3012 } else if (s->is_MergeMem()) {
3013 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3014 Node* s1 = s->fast_out(i);
3015 worklist.push(s1);
3016 }
3017 } else {
3018 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3019 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3020 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3021 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3022 }
3023 }
3024 }
3025 }
3026
3027 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3028 return LCA;
3029 }
3030
3031 // true if CFG node d dominates CFG node n
3032 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3033 if (d == n)
3034 return true;
3035 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3036 uint dd = dom_depth(d);
3037 while (dom_depth(n) >= dd) {
3038 if (n == d)
3039 return true;
3040 n = idom(n);
3041 }
3042 return false;
3043 }
3044
3045 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3046 // Pair-wise LCA with tags.
3047 // Tag each index with the node 'tag' currently being processed
3048 // before advancing up the dominator chain using idom().
3049 // Later calls that find a match to 'tag' know that this path has already
3050 // been considered in the current LCA (which is input 'n1' by convention).
3051 // Since get_late_ctrl() is only called once for each node, the tag array
3052 // does not need to be cleared between calls to get_late_ctrl().
3053 // Algorithm trades a larger constant factor for better asymptotic behavior
3054 //
3055 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3056 uint d1 = dom_depth(n1);
3057 uint d2 = dom_depth(n2);
3058
3059 do {
3060 if (d1 > d2) {
3061 // current lca is deeper than n2
3062 _dom_lca_tags.map(n1->_idx, tag);
3063 n1 = idom(n1);
3064 d1 = dom_depth(n1);
3065 } else if (d1 < d2) {
3066 // n2 is deeper than current lca
3067 Node *memo = _dom_lca_tags[n2->_idx];
3068 if( memo == tag ) {
3069 return n1; // Return the current LCA
3070 }
3071 _dom_lca_tags.map(n2->_idx, tag);
3072 n2 = idom(n2);
3073 d2 = dom_depth(n2);
3074 } else {
3075 // Here d1 == d2. Due to edits of the dominator-tree, sections
3076 // of the tree might have the same depth. These sections have
3077 // to be searched more carefully.
3078
3079 // Scan up all the n1's with equal depth, looking for n2.
3080 _dom_lca_tags.map(n1->_idx, tag);
3081 Node *t1 = idom(n1);
3082 while (dom_depth(t1) == d1) {
3083 if (t1 == n2) return n2;
3084 _dom_lca_tags.map(t1->_idx, tag);
3085 t1 = idom(t1);
3086 }
3087 // Scan up all the n2's with equal depth, looking for n1.
3088 _dom_lca_tags.map(n2->_idx, tag);
3089 Node *t2 = idom(n2);
3090 while (dom_depth(t2) == d2) {
3091 if (t2 == n1) return n1;
3092 _dom_lca_tags.map(t2->_idx, tag);
3093 t2 = idom(t2);
3094 }
3095 // Move up to a new dominator-depth value as well as up the dom-tree.
3096 n1 = t1;
3097 n2 = t2;
3098 d1 = dom_depth(n1);
3099 d2 = dom_depth(n2);
3100 }
3101 } while (n1 != n2);
3102 return n1;
3103 }
3104
3105 //------------------------------init_dom_lca_tags------------------------------
3106 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3107 // Intended use does not involve any growth for the array, so it could
3108 // be of fixed size.
3109 void PhaseIdealLoop::init_dom_lca_tags() {
3110 uint limit = C->unique() + 1;
3111 _dom_lca_tags.map( limit, NULL );
3112 #ifdef ASSERT
3113 for( uint i = 0; i < limit; ++i ) {
3114 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3115 }
3116 #endif // ASSERT
3117 }
3118
3119 //------------------------------clear_dom_lca_tags------------------------------
3120 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3121 // Intended use does not involve any growth for the array, so it could
3122 // be of fixed size.
3123 void PhaseIdealLoop::clear_dom_lca_tags() {
3124 uint limit = C->unique() + 1;
3125 _dom_lca_tags.map( limit, NULL );
3126 _dom_lca_tags.clear();
3127 #ifdef ASSERT
3128 for( uint i = 0; i < limit; ++i ) {
3129 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3130 }
3131 #endif // ASSERT
3132 }
3133
3134 //------------------------------build_loop_late--------------------------------
3135 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3136 // Second pass finds latest legal placement, and ideal loop placement.
3137 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3138 while (worklist.size() != 0) {
3139 Node *n = worklist.pop();
3140 // Only visit once
3141 if (visited.test_set(n->_idx)) continue;
3142 uint cnt = n->outcnt();
3143 uint i = 0;
3144 while (true) {
3145 assert( _nodes[n->_idx], "no dead nodes" );
3146 // Visit all children
3147 if (i < cnt) {
3148 Node* use = n->raw_out(i);
3149 ++i;
3150 // Check for dead uses. Aggressively prune such junk. It might be
3151 // dead in the global sense, but still have local uses so I cannot
3152 // easily call 'remove_dead_node'.
3153 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3154 // Due to cycles, we might not hit the same fixed point in the verify
3155 // pass as we do in the regular pass. Instead, visit such phis as
3156 // simple uses of the loop head.
3157 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3158 if( !visited.test(use->_idx) )
3159 worklist.push(use);
3160 } else if( !visited.test_set(use->_idx) ) {
3161 nstack.push(n, i); // Save parent and next use's index.
3162 n = use; // Process all children of current use.
3163 cnt = use->outcnt();
3164 i = 0;
3165 }
3166 } else {
3167 // Do not visit around the backedge of loops via data edges.
3168 // push dead code onto a worklist
3169 _deadlist.push(use);
3170 }
3171 } else {
3172 // All of n's children have been processed, complete post-processing.
3173 build_loop_late_post(n);
3174 if (nstack.is_empty()) {
3175 // Finished all nodes on stack.
3176 // Process next node on the worklist.
3177 break;
3178 }
3179 // Get saved parent node and next use's index. Visit the rest of uses.
3180 n = nstack.node();
3181 cnt = n->outcnt();
3182 i = nstack.index();
3183 nstack.pop();
3184 }
3185 }
3186 }
3187 }
3188
3189 //------------------------------build_loop_late_post---------------------------
3190 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3191 // Second pass finds latest legal placement, and ideal loop placement.
3192 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3193
3194 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3195 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3196 }
3197
3198 // CFG and pinned nodes already handled
3199 if( n->in(0) ) {
3200 if( n->in(0)->is_top() ) return; // Dead?
3201
3202 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3203 // _must_ be pinned (they have to observe their control edge of course).
3204 // Unlike Stores (which modify an unallocable resource, the memory
3205 // state), Mods/Loads can float around. So free them up.
3206 bool pinned = true;
3207 switch( n->Opcode() ) {
3208 case Op_DivI:
3209 case Op_DivF:
3210 case Op_DivD:
3211 case Op_ModI:
3212 case Op_ModF:
3213 case Op_ModD:
3214 case Op_LoadB: // Same with Loads; they can sink
3215 case Op_LoadUS: // during loop optimizations.
3216 case Op_LoadD:
3217 case Op_LoadF:
3218 case Op_LoadI:
3219 case Op_LoadKlass:
3220 case Op_LoadNKlass:
3221 case Op_LoadL:
3222 case Op_LoadS:
3223 case Op_LoadP:
3224 case Op_LoadN:
3225 case Op_LoadRange:
3226 case Op_LoadD_unaligned:
3227 case Op_LoadL_unaligned:
3228 case Op_StrComp: // Does a bunch of load-like effects
3229 case Op_StrEquals:
3230 case Op_StrIndexOf:
3231 case Op_AryEq:
3232 pinned = false;
3233 }
3234 if( pinned ) {
3235 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3236 if( !chosen_loop->_child ) // Inner loop?
3237 chosen_loop->_body.push(n); // Collect inner loops
3238 return;
3239 }
3240 } else { // No slot zero
3241 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3242 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3243 return;
3244 }
3245 assert(!n->is_CFG() || n->outcnt() == 0, "");
3246 }
3247
3248 // Do I have a "safe range" I can select over?
3249 Node *early = get_ctrl(n);// Early location already computed
3250
3251 // Compute latest point this Node can go
3252 Node *LCA = get_late_ctrl( n, early );
3253 // LCA is NULL due to uses being dead
3254 if( LCA == NULL ) {
3255 #ifdef ASSERT
3256 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3257 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3258 }
3259 #endif
3260 _nodes.map(n->_idx, 0); // This node is useless
3261 _deadlist.push(n);
3262 return;
3263 }
3264 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3265
3266 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3267 Node *least = legal; // Best legal position so far
3268 while( early != legal ) { // While not at earliest legal
3269 #ifdef ASSERT
3270 if (legal->is_Start() && !early->is_Root()) {
3271 // Bad graph. Print idom path and fail.
3272 tty->print_cr( "Bad graph detected in build_loop_late");
3273 tty->print("n: ");n->dump(); tty->cr();
3274 tty->print("early: ");early->dump(); tty->cr();
3275 int ct = 0;
3276 Node *dbg_legal = LCA;
3277 while(!dbg_legal->is_Start() && ct < 100) {
3278 tty->print("idom[%d] ",ct); dbg_legal->dump(); tty->cr();
3279 ct++;
3280 dbg_legal = idom(dbg_legal);
3281 }
3282 assert(false, "Bad graph detected in build_loop_late");
3283 }
3284 #endif
3285 // Find least loop nesting depth
3286 legal = idom(legal); // Bump up the IDOM tree
3287 // Check for lower nesting depth
3288 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3289 least = legal;
3290 }
3291 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3292
3293 // Try not to place code on a loop entry projection
3294 // which can inhibit range check elimination.
3295 if (least != early) {
3296 Node* ctrl_out = least->unique_ctrl_out();
3297 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3298 least == ctrl_out->in(LoopNode::EntryControl)) {
3299 Node* least_dom = idom(least);
3300 if (get_loop(least_dom)->is_member(get_loop(least))) {
3301 least = least_dom;
3302 }
3303 }
3304 }
3305
3306 #ifdef ASSERT
3307 // If verifying, verify that 'verify_me' has a legal location
3308 // and choose it as our location.
3309 if( _verify_me ) {
3310 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3311 Node *legal = LCA;
3312 while( early != legal ) { // While not at earliest legal
3313 if( legal == v_ctrl ) break; // Check for prior good location
3314 legal = idom(legal) ;// Bump up the IDOM tree
3315 }
3316 // Check for prior good location
3317 if( legal == v_ctrl ) least = legal; // Keep prior if found
3318 }
3319 #endif
3320
3321 // Assign discovered "here or above" point
3322 least = find_non_split_ctrl(least);
3323 set_ctrl(n, least);
3324
3325 // Collect inner loop bodies
3326 IdealLoopTree *chosen_loop = get_loop(least);
3327 if( !chosen_loop->_child ) // Inner loop?
3328 chosen_loop->_body.push(n);// Collect inner loops
3329 }
3330
3331 #ifndef PRODUCT
3332 //------------------------------dump-------------------------------------------
3333 void PhaseIdealLoop::dump( ) const {
3334 ResourceMark rm;
3335 Arena* arena = Thread::current()->resource_area();
3336 Node_Stack stack(arena, C->unique() >> 2);
3337 Node_List rpo_list;
3338 VectorSet visited(arena);
3339 visited.set(C->top()->_idx);
3340 rpo( C->root(), stack, visited, rpo_list );
3341 // Dump root loop indexed by last element in PO order
3342 dump( _ltree_root, rpo_list.size(), rpo_list );
3343 }
3344
3345 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3346 loop->dump_head();
3347
3348 // Now scan for CFG nodes in the same loop
3349 for( uint j=idx; j > 0; j-- ) {
3350 Node *n = rpo_list[j-1];
3351 if( !_nodes[n->_idx] ) // Skip dead nodes
3352 continue;
3353 if( get_loop(n) != loop ) { // Wrong loop nest
3354 if( get_loop(n)->_head == n && // Found nested loop?
3355 get_loop(n)->_parent == loop )
3356 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3357 continue;
3358 }
3359
3360 // Dump controlling node
3361 for( uint x = 0; x < loop->_nest; x++ )
3362 tty->print(" ");
3363 tty->print("C");
3364 if( n == C->root() ) {
3365 n->dump();
3366 } else {
3367 Node* cached_idom = idom_no_update(n);
3368 Node *computed_idom = n->in(0);
3369 if( n->is_Region() ) {
3370 computed_idom = compute_idom(n);
3371 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3372 // any MultiBranch ctrl node), so apply a similar transform to
3373 // the cached idom returned from idom_no_update.
3374 cached_idom = find_non_split_ctrl(cached_idom);
3375 }
3376 tty->print(" ID:%d",computed_idom->_idx);
3377 n->dump();
3378 if( cached_idom != computed_idom ) {
3379 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3380 computed_idom->_idx, cached_idom->_idx);
3381 }
3382 }
3383 // Dump nodes it controls
3384 for( uint k = 0; k < _nodes.Size(); k++ ) {
3385 // (k < C->unique() && get_ctrl(find(k)) == n)
3386 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3387 Node *m = C->root()->find(k);
3388 if( m && m->outcnt() > 0 ) {
3389 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3390 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3391 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3392 }
3393 for( uint j = 0; j < loop->_nest; j++ )
3394 tty->print(" ");
3395 tty->print(" ");
3396 m->dump();
3397 }
3398 }
3399 }
3400 }
3401 }
3402
3403 // Collect a R-P-O for the whole CFG.
3404 // Result list is in post-order (scan backwards for RPO)
3405 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3406 stk.push(start, 0);
3407 visited.set(start->_idx);
3408
3409 while (stk.is_nonempty()) {
3410 Node* m = stk.node();
3411 uint idx = stk.index();
3412 if (idx < m->outcnt()) {
3413 stk.set_index(idx + 1);
3414 Node* n = m->raw_out(idx);
3415 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3416 stk.push(n, 0);
3417 }
3418 } else {
3419 rpo_list.push(m);
3420 stk.pop();
3421 }
3422 }
3423 }
3424 #endif
3425
3426
3427 //=============================================================================
3428 //------------------------------LoopTreeIterator-----------------------------------
3429
3430 // Advance to next loop tree using a preorder, left-to-right traversal.
3431 void LoopTreeIterator::next() {
3432 assert(!done(), "must not be done.");
3433 if (_curnt->_child != NULL) {
3434 _curnt = _curnt->_child;
3435 } else if (_curnt->_next != NULL) {
3436 _curnt = _curnt->_next;
3437 } else {
3438 while (_curnt != _root && _curnt->_next == NULL) {
3439 _curnt = _curnt->_parent;
3440 }
3441 if (_curnt == _root) {
3442 _curnt = NULL;
3443 assert(done(), "must be done.");
3444 } else {
3445 assert(_curnt->_next != NULL, "must be more to do");
3446 _curnt = _curnt->_next;
3447 }
3448 }
3449 }