1 /*
2 * Copyright 1998-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_loopnode.cpp.incl"
27
28 //=============================================================================
29 //------------------------------is_loop_iv-------------------------------------
30 // Determine if a node is Counted loop induction variable.
31 // The method is declared in node.hpp.
32 const Node* Node::is_loop_iv() const {
33 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
34 this->as_Phi()->region()->is_CountedLoop() &&
35 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
36 return this;
37 } else {
38 return NULL;
39 }
40 }
41
42 //=============================================================================
43 //------------------------------dump_spec--------------------------------------
44 // Dump special per-node info
45 #ifndef PRODUCT
46 void LoopNode::dump_spec(outputStream *st) const {
47 if( is_inner_loop () ) st->print( "inner " );
48 if( is_partial_peel_loop () ) st->print( "partial_peel " );
49 if( partial_peel_has_failed () ) st->print( "partial_peel_failed " );
50 }
51 #endif
52
53 //------------------------------get_early_ctrl---------------------------------
54 // Compute earliest legal control
55 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
56 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
57 uint i;
58 Node *early;
59 if( n->in(0) ) {
60 early = n->in(0);
61 if( !early->is_CFG() ) // Might be a non-CFG multi-def
62 early = get_ctrl(early); // So treat input as a straight data input
63 i = 1;
64 } else {
65 early = get_ctrl(n->in(1));
66 i = 2;
67 }
68 uint e_d = dom_depth(early);
69 assert( early, "" );
70 for( ; i < n->req(); i++ ) {
71 Node *cin = get_ctrl(n->in(i));
72 assert( cin, "" );
73 // Keep deepest dominator depth
74 uint c_d = dom_depth(cin);
75 if( c_d > e_d ) { // Deeper guy?
76 early = cin; // Keep deepest found so far
77 e_d = c_d;
78 } else if( c_d == e_d && // Same depth?
79 early != cin ) { // If not equal, must use slower algorithm
80 // If same depth but not equal, one _must_ dominate the other
81 // and we want the deeper (i.e., dominated) guy.
82 Node *n1 = early;
83 Node *n2 = cin;
84 while( 1 ) {
85 n1 = idom(n1); // Walk up until break cycle
86 n2 = idom(n2);
87 if( n1 == cin || // Walked early up to cin
88 dom_depth(n2) < c_d )
89 break; // early is deeper; keep him
90 if( n2 == early || // Walked cin up to early
91 dom_depth(n1) < c_d ) {
92 early = cin; // cin is deeper; keep him
93 break;
94 }
95 }
96 e_d = dom_depth(early); // Reset depth register cache
97 }
98 }
99
100 // Return earliest legal location
101 assert(early == find_non_split_ctrl(early), "unexpected early control");
102
103 return early;
104 }
105
106 //------------------------------set_early_ctrl---------------------------------
107 // Set earliest legal control
108 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
109 Node *early = get_early_ctrl(n);
110
111 // Record earliest legal location
112 set_ctrl(n, early);
113 }
114
115 //------------------------------set_subtree_ctrl-------------------------------
116 // set missing _ctrl entries on new nodes
117 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
118 // Already set? Get out.
119 if( _nodes[n->_idx] ) return;
120 // Recursively set _nodes array to indicate where the Node goes
121 uint i;
122 for( i = 0; i < n->req(); ++i ) {
123 Node *m = n->in(i);
124 if( m && m != C->root() )
125 set_subtree_ctrl( m );
126 }
127
128 // Fixup self
129 set_early_ctrl( n );
130 }
131
132 //------------------------------is_counted_loop--------------------------------
133 Node *PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
134 PhaseGVN *gvn = &_igvn;
135
136 // Counted loop head must be a good RegionNode with only 3 not NULL
137 // control input edges: Self, Entry, LoopBack.
138 if ( x->in(LoopNode::Self) == NULL || x->req() != 3 )
139 return NULL;
140
141 Node *init_control = x->in(LoopNode::EntryControl);
142 Node *back_control = x->in(LoopNode::LoopBackControl);
143 if( init_control == NULL || back_control == NULL ) // Partially dead
144 return NULL;
145 // Must also check for TOP when looking for a dead loop
146 if( init_control->is_top() || back_control->is_top() )
147 return NULL;
148
149 // Allow funny placement of Safepoint
150 if( back_control->Opcode() == Op_SafePoint )
151 back_control = back_control->in(TypeFunc::Control);
152
153 // Controlling test for loop
154 Node *iftrue = back_control;
155 uint iftrue_op = iftrue->Opcode();
156 if( iftrue_op != Op_IfTrue &&
157 iftrue_op != Op_IfFalse )
158 // I have a weird back-control. Probably the loop-exit test is in
159 // the middle of the loop and I am looking at some trailing control-flow
160 // merge point. To fix this I would have to partially peel the loop.
161 return NULL; // Obscure back-control
162
163 // Get boolean guarding loop-back test
164 Node *iff = iftrue->in(0);
165 if( get_loop(iff) != loop || !iff->in(1)->is_Bool() ) return NULL;
166 BoolNode *test = iff->in(1)->as_Bool();
167 BoolTest::mask bt = test->_test._test;
168 float cl_prob = iff->as_If()->_prob;
169 if( iftrue_op == Op_IfFalse ) {
170 bt = BoolTest(bt).negate();
171 cl_prob = 1.0 - cl_prob;
172 }
173 // Get backedge compare
174 Node *cmp = test->in(1);
175 int cmp_op = cmp->Opcode();
176 if( cmp_op != Op_CmpI )
177 return NULL; // Avoid pointer & float compares
178
179 // Find the trip-counter increment & limit. Limit must be loop invariant.
180 Node *incr = cmp->in(1);
181 Node *limit = cmp->in(2);
182
183 // ---------
184 // need 'loop()' test to tell if limit is loop invariant
185 // ---------
186
187 if( !is_member( loop, get_ctrl(incr) ) ) { // Swapped trip counter and limit?
188 Node *tmp = incr; // Then reverse order into the CmpI
189 incr = limit;
190 limit = tmp;
191 bt = BoolTest(bt).commute(); // And commute the exit test
192 }
193 if( is_member( loop, get_ctrl(limit) ) ) // Limit must loop-invariant
194 return NULL;
195
196 // Trip-counter increment must be commutative & associative.
197 uint incr_op = incr->Opcode();
198 if( incr_op == Op_Phi && incr->req() == 3 ) {
199 incr = incr->in(2); // Assume incr is on backedge of Phi
200 incr_op = incr->Opcode();
201 }
202 Node* trunc1 = NULL;
203 Node* trunc2 = NULL;
204 const TypeInt* iv_trunc_t = NULL;
205 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
206 return NULL; // Funny increment opcode
207 }
208
209 // Get merge point
210 Node *xphi = incr->in(1);
211 Node *stride = incr->in(2);
212 if( !stride->is_Con() ) { // Oops, swap these
213 if( !xphi->is_Con() ) // Is the other guy a constant?
214 return NULL; // Nope, unknown stride, bail out
215 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
216 xphi = stride;
217 stride = tmp;
218 }
219 //if( loop(xphi) != l) return NULL;// Merge point is in inner loop??
220 if( !xphi->is_Phi() ) return NULL; // Too much math on the trip counter
221 PhiNode *phi = xphi->as_Phi();
222
223 // Stride must be constant
224 const Type *stride_t = stride->bottom_type();
225 int stride_con = stride_t->is_int()->get_con();
226 assert( stride_con, "missed some peephole opt" );
227
228 // Phi must be of loop header; backedge must wrap to increment
229 if( phi->region() != x ) return NULL;
230 if( trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
231 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1 ) {
232 return NULL;
233 }
234 Node *init_trip = phi->in(LoopNode::EntryControl);
235 //if (!init_trip->is_Con()) return NULL; // avoid rolling over MAXINT/MININT
236
237 // If iv trunc type is smaller than int, check for possible wrap.
238 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
239 assert(trunc1 != NULL, "must have found some truncation");
240
241 // Get a better type for the phi (filtered thru if's)
242 const TypeInt* phi_ft = filtered_type(phi);
243
244 // Can iv take on a value that will wrap?
245 //
246 // Ensure iv's limit is not within "stride" of the wrap value.
247 //
248 // Example for "short" type
249 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
250 // If the stride is +10, then the last value of the induction
251 // variable before the increment (phi_ft->_hi) must be
252 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
253 // ensure no truncation occurs after the increment.
254
255 if (stride_con > 0) {
256 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
257 iv_trunc_t->_lo > phi_ft->_lo) {
258 return NULL; // truncation may occur
259 }
260 } else if (stride_con < 0) {
261 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
262 iv_trunc_t->_hi < phi_ft->_hi) {
263 return NULL; // truncation may occur
264 }
265 }
266 // No possibility of wrap so truncation can be discarded
267 // Promote iv type to Int
268 } else {
269 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
270 }
271
272 // =================================================
273 // ---- SUCCESS! Found A Trip-Counted Loop! -----
274 //
275 // Canonicalize the condition on the test. If we can exactly determine
276 // the trip-counter exit value, then set limit to that value and use
277 // a '!=' test. Otherwise use conditon '<' for count-up loops and
278 // '>' for count-down loops. If the condition is inverted and we will
279 // be rolling through MININT to MAXINT, then bail out.
280
281 C->print_method("Before CountedLoop", 3);
282
283 // Check for SafePoint on backedge and remove
284 Node *sfpt = x->in(LoopNode::LoopBackControl);
285 if( sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
286 lazy_replace( sfpt, iftrue );
287 loop->_tail = iftrue;
288 }
289
290
291 // If compare points to incr, we are ok. Otherwise the compare
292 // can directly point to the phi; in this case adjust the compare so that
293 // it points to the incr by adusting the limit.
294 if( cmp->in(1) == phi || cmp->in(2) == phi )
295 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
296
297 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
298 // Final value for iterator should be: trip_count * stride + init_trip.
299 const Type *limit_t = limit->bottom_type();
300 const Type *init_t = init_trip->bottom_type();
301 Node *one_p = gvn->intcon( 1);
302 Node *one_m = gvn->intcon(-1);
303
304 Node *trip_count = NULL;
305 Node *hook = new (C, 6) Node(6);
306 switch( bt ) {
307 case BoolTest::eq:
308 return NULL; // Bail out, but this loop trips at most twice!
309 case BoolTest::ne: // Ahh, the case we desire
310 if( stride_con == 1 )
311 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
312 else if( stride_con == -1 )
313 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
314 else
315 return NULL; // Odd stride; must prove we hit limit exactly
316 set_subtree_ctrl( trip_count );
317 //_loop.map(trip_count->_idx,loop(limit));
318 break;
319 case BoolTest::le: // Maybe convert to '<' case
320 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
321 set_subtree_ctrl( limit );
322 hook->init_req(4, limit);
323
324 bt = BoolTest::lt;
325 // Make the new limit be in the same loop nest as the old limit
326 //_loop.map(limit->_idx,limit_loop);
327 // Fall into next case
328 case BoolTest::lt: { // Maybe convert to '!=' case
329 if( stride_con < 0 ) return NULL; // Count down loop rolls through MAXINT
330 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
331 set_subtree_ctrl( range );
332 hook->init_req(0, range);
333
334 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
335 set_subtree_ctrl( bias );
336 hook->init_req(1, bias);
337
338 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
339 set_subtree_ctrl( bias1 );
340 hook->init_req(2, bias1);
341
342 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
343 set_subtree_ctrl( trip_count );
344 hook->init_req(3, trip_count);
345 break;
346 }
347
348 case BoolTest::ge: // Maybe convert to '>' case
349 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
350 set_subtree_ctrl( limit );
351 hook->init_req(4 ,limit);
352
353 bt = BoolTest::gt;
354 // Make the new limit be in the same loop nest as the old limit
355 //_loop.map(limit->_idx,limit_loop);
356 // Fall into next case
357 case BoolTest::gt: { // Maybe convert to '!=' case
358 if( stride_con > 0 ) return NULL; // count up loop rolls through MININT
359 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
360 set_subtree_ctrl( range );
361 hook->init_req(0, range);
362
363 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
364 set_subtree_ctrl( bias );
365 hook->init_req(1, bias);
366
367 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
368 set_subtree_ctrl( bias1 );
369 hook->init_req(2, bias1);
370
371 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
372 set_subtree_ctrl( trip_count );
373 hook->init_req(3, trip_count);
374 break;
375 }
376 }
377
378 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
379 set_subtree_ctrl( span );
380 hook->init_req(5, span);
381
382 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
383 set_subtree_ctrl( limit );
384
385 // Build a canonical trip test.
386 // Clone code, as old values may be in use.
387 incr = incr->clone();
388 incr->set_req(1,phi);
389 incr->set_req(2,stride);
390 incr = _igvn.register_new_node_with_optimizer(incr);
391 set_early_ctrl( incr );
392 _igvn.hash_delete(phi);
393 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
394
395 // If phi type is more restrictive than Int, raise to
396 // Int to prevent (almost) infinite recursion in igvn
397 // which can only handle integer types for constants or minint..maxint.
398 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
399 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
400 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
401 nphi = _igvn.register_new_node_with_optimizer(nphi);
402 set_ctrl(nphi, get_ctrl(phi));
403 _igvn.subsume_node(phi, nphi);
404 phi = nphi->as_Phi();
405 }
406 cmp = cmp->clone();
407 cmp->set_req(1,incr);
408 cmp->set_req(2,limit);
409 cmp = _igvn.register_new_node_with_optimizer(cmp);
410 set_ctrl(cmp, iff->in(0));
411
412 Node *tmp = test->clone();
413 assert( tmp->is_Bool(), "" );
414 test = (BoolNode*)tmp;
415 (*(BoolTest*)&test->_test)._test = bt; //BoolTest::ne;
416 test->set_req(1,cmp);
417 _igvn.register_new_node_with_optimizer(test);
418 set_ctrl(test, iff->in(0));
419 // If the exit test is dead, STOP!
420 if( test == NULL ) return NULL;
421 _igvn.hash_delete(iff);
422 iff->set_req_X( 1, test, &_igvn );
423
424 // Replace the old IfNode with a new LoopEndNode
425 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), iff->in(1), cl_prob, iff->as_If()->_fcnt ));
426 IfNode *le = lex->as_If();
427 uint dd = dom_depth(iff);
428 set_idom(le, le->in(0), dd); // Update dominance for loop exit
429 set_loop(le, loop);
430
431 // Get the loop-exit control
432 Node *if_f = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
433
434 // Need to swap loop-exit and loop-back control?
435 if( iftrue_op == Op_IfFalse ) {
436 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
437 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
438
439 loop->_tail = back_control = ift2;
440 set_loop(ift2, loop);
441 set_loop(iff2, get_loop(if_f));
442
443 // Lazy update of 'get_ctrl' mechanism.
444 lazy_replace_proj( if_f , iff2 );
445 lazy_replace_proj( iftrue, ift2 );
446
447 // Swap names
448 if_f = iff2;
449 iftrue = ift2;
450 } else {
451 _igvn.hash_delete(if_f );
452 _igvn.hash_delete(iftrue);
453 if_f ->set_req_X( 0, le, &_igvn );
454 iftrue->set_req_X( 0, le, &_igvn );
455 }
456
457 set_idom(iftrue, le, dd+1);
458 set_idom(if_f, le, dd+1);
459
460 // Now setup a new CountedLoopNode to replace the existing LoopNode
461 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
462 // The following assert is approximately true, and defines the intention
463 // of can_be_counted_loop. It fails, however, because phase->type
464 // is not yet initialized for this loop and its parts.
465 //assert(l->can_be_counted_loop(this), "sanity");
466 _igvn.register_new_node_with_optimizer(l);
467 set_loop(l, loop);
468 loop->_head = l;
469 // Fix all data nodes placed at the old loop head.
470 // Uses the lazy-update mechanism of 'get_ctrl'.
471 lazy_replace( x, l );
472 set_idom(l, init_control, dom_depth(x));
473
474 // Check for immediately preceeding SafePoint and remove
475 Node *sfpt2 = le->in(0);
476 if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
477 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
478
479 // Free up intermediate goo
480 _igvn.remove_dead_node(hook);
481
482 C->print_method("After CountedLoop", 3);
483
484 // Return trip counter
485 return trip_count;
486 }
487
488
489 //------------------------------Ideal------------------------------------------
490 // Return a node which is more "ideal" than the current node.
491 // Attempt to convert into a counted-loop.
492 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
493 if (!can_be_counted_loop(phase)) {
494 phase->C->set_major_progress();
495 }
496 return RegionNode::Ideal(phase, can_reshape);
497 }
498
499
500 //=============================================================================
501 //------------------------------Ideal------------------------------------------
502 // Return a node which is more "ideal" than the current node.
503 // Attempt to convert into a counted-loop.
504 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
505 return RegionNode::Ideal(phase, can_reshape);
506 }
507
508 //------------------------------dump_spec--------------------------------------
509 // Dump special per-node info
510 #ifndef PRODUCT
511 void CountedLoopNode::dump_spec(outputStream *st) const {
512 LoopNode::dump_spec(st);
513 if( stride_is_con() ) {
514 st->print("stride: %d ",stride_con());
515 } else {
516 st->print("stride: not constant ");
517 }
518 if( is_pre_loop () ) st->print("pre of N%d" , _main_idx );
519 if( is_main_loop() ) st->print("main of N%d", _idx );
520 if( is_post_loop() ) st->print("post of N%d", _main_idx );
521 }
522 #endif
523
524 //=============================================================================
525 int CountedLoopEndNode::stride_con() const {
526 return stride()->bottom_type()->is_int()->get_con();
527 }
528
529
530 //----------------------match_incr_with_optional_truncation--------------------
531 // Match increment with optional truncation:
532 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
533 // Return NULL for failure. Success returns the increment node.
534 Node* CountedLoopNode::match_incr_with_optional_truncation(
535 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
536 // Quick cutouts:
537 if (expr == NULL || expr->req() != 3) return false;
538
539 Node *t1 = NULL;
540 Node *t2 = NULL;
541 const TypeInt* trunc_t = TypeInt::INT;
542 Node* n1 = expr;
543 int n1op = n1->Opcode();
544
545 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
546 if (n1op == Op_AndI &&
547 n1->in(2)->is_Con() &&
548 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
549 // %%% This check should match any mask of 2**K-1.
550 t1 = n1;
551 n1 = t1->in(1);
552 n1op = n1->Opcode();
553 trunc_t = TypeInt::CHAR;
554 } else if (n1op == Op_RShiftI &&
555 n1->in(1) != NULL &&
556 n1->in(1)->Opcode() == Op_LShiftI &&
557 n1->in(2) == n1->in(1)->in(2) &&
558 n1->in(2)->is_Con()) {
559 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
560 // %%% This check should match any shift in [1..31].
561 if (shift == 16 || shift == 8) {
562 t1 = n1;
563 t2 = t1->in(1);
564 n1 = t2->in(1);
565 n1op = n1->Opcode();
566 if (shift == 16) {
567 trunc_t = TypeInt::SHORT;
568 } else if (shift == 8) {
569 trunc_t = TypeInt::BYTE;
570 }
571 }
572 }
573
574 // If (maybe after stripping) it is an AddI, we won:
575 if (n1op == Op_AddI) {
576 *trunc1 = t1;
577 *trunc2 = t2;
578 *trunc_type = trunc_t;
579 return n1;
580 }
581
582 // failed
583 return NULL;
584 }
585
586
587 //------------------------------filtered_type--------------------------------
588 // Return a type based on condition control flow
589 // A successful return will be a type that is restricted due
590 // to a series of dominating if-tests, such as:
591 // if (i < 10) {
592 // if (i > 0) {
593 // here: "i" type is [1..10)
594 // }
595 // }
596 // or a control flow merge
597 // if (i < 10) {
598 // do {
599 // phi( , ) -- at top of loop type is [min_int..10)
600 // i = ?
601 // } while ( i < 10)
602 //
603 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
604 assert(n && n->bottom_type()->is_int(), "must be int");
605 const TypeInt* filtered_t = NULL;
606 if (!n->is_Phi()) {
607 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
608 filtered_t = filtered_type_from_dominators(n, n_ctrl);
609
610 } else {
611 Node* phi = n->as_Phi();
612 Node* region = phi->in(0);
613 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
614 if (region && region != C->top()) {
615 for (uint i = 1; i < phi->req(); i++) {
616 Node* val = phi->in(i);
617 Node* use_c = region->in(i);
618 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
619 if (val_t != NULL) {
620 if (filtered_t == NULL) {
621 filtered_t = val_t;
622 } else {
623 filtered_t = filtered_t->meet(val_t)->is_int();
624 }
625 }
626 }
627 }
628 }
629 const TypeInt* n_t = _igvn.type(n)->is_int();
630 if (filtered_t != NULL) {
631 n_t = n_t->join(filtered_t)->is_int();
632 }
633 return n_t;
634 }
635
636
637 //------------------------------filtered_type_from_dominators--------------------------------
638 // Return a possibly more restrictive type for val based on condition control flow of dominators
639 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
640 if (val->is_Con()) {
641 return val->bottom_type()->is_int();
642 }
643 uint if_limit = 10; // Max number of dominating if's visited
644 const TypeInt* rtn_t = NULL;
645
646 if (use_ctrl && use_ctrl != C->top()) {
647 Node* val_ctrl = get_ctrl(val);
648 uint val_dom_depth = dom_depth(val_ctrl);
649 Node* pred = use_ctrl;
650 uint if_cnt = 0;
651 while (if_cnt < if_limit) {
652 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
653 if_cnt++;
654 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
655 if (if_t != NULL) {
656 if (rtn_t == NULL) {
657 rtn_t = if_t;
658 } else {
659 rtn_t = rtn_t->join(if_t)->is_int();
660 }
661 }
662 }
663 pred = idom(pred);
664 if (pred == NULL || pred == C->top()) {
665 break;
666 }
667 // Stop if going beyond definition block of val
668 if (dom_depth(pred) < val_dom_depth) {
669 break;
670 }
671 }
672 }
673 return rtn_t;
674 }
675
676
677 //------------------------------dump_spec--------------------------------------
678 // Dump special per-node info
679 #ifndef PRODUCT
680 void CountedLoopEndNode::dump_spec(outputStream *st) const {
681 if( in(TestValue)->is_Bool() ) {
682 BoolTest bt( test_trip()); // Added this for g++.
683
684 st->print("[");
685 bt.dump_on(st);
686 st->print("]");
687 }
688 st->print(" ");
689 IfNode::dump_spec(st);
690 }
691 #endif
692
693 //=============================================================================
694 //------------------------------is_member--------------------------------------
695 // Is 'l' a member of 'this'?
696 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
697 while( l->_nest > _nest ) l = l->_parent;
698 return l == this;
699 }
700
701 //------------------------------set_nest---------------------------------------
702 // Set loop tree nesting depth. Accumulate _has_call bits.
703 int IdealLoopTree::set_nest( uint depth ) {
704 _nest = depth;
705 int bits = _has_call;
706 if( _child ) bits |= _child->set_nest(depth+1);
707 if( bits ) _has_call = 1;
708 if( _next ) bits |= _next ->set_nest(depth );
709 return bits;
710 }
711
712 //------------------------------split_fall_in----------------------------------
713 // Split out multiple fall-in edges from the loop header. Move them to a
714 // private RegionNode before the loop. This becomes the loop landing pad.
715 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
716 PhaseIterGVN &igvn = phase->_igvn;
717 uint i;
718
719 // Make a new RegionNode to be the landing pad.
720 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
721 phase->set_loop(landing_pad,_parent);
722 // Gather all the fall-in control paths into the landing pad
723 uint icnt = fall_in_cnt;
724 uint oreq = _head->req();
725 for( i = oreq-1; i>0; i-- )
726 if( !phase->is_member( this, _head->in(i) ) )
727 landing_pad->set_req(icnt--,_head->in(i));
728
729 // Peel off PhiNode edges as well
730 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
731 Node *oj = _head->fast_out(j);
732 if( oj->is_Phi() ) {
733 PhiNode* old_phi = oj->as_Phi();
734 assert( old_phi->region() == _head, "" );
735 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
736 Node *p = PhiNode::make_blank(landing_pad, old_phi);
737 uint icnt = fall_in_cnt;
738 for( i = oreq-1; i>0; i-- ) {
739 if( !phase->is_member( this, _head->in(i) ) ) {
740 p->init_req(icnt--, old_phi->in(i));
741 // Go ahead and clean out old edges from old phi
742 old_phi->del_req(i);
743 }
744 }
745 // Search for CSE's here, because ZKM.jar does a lot of
746 // loop hackery and we need to be a little incremental
747 // with the CSE to avoid O(N^2) node blow-up.
748 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
749 if( p2 ) { // Found CSE
750 p->destruct(); // Recover useless new node
751 p = p2; // Use old node
752 } else {
753 igvn.register_new_node_with_optimizer(p, old_phi);
754 }
755 // Make old Phi refer to new Phi.
756 old_phi->add_req(p);
757 // Check for the special case of making the old phi useless and
758 // disappear it. In JavaGrande I have a case where this useless
759 // Phi is the loop limit and prevents recognizing a CountedLoop
760 // which in turn prevents removing an empty loop.
761 Node *id_old_phi = old_phi->Identity( &igvn );
762 if( id_old_phi != old_phi ) { // Found a simple identity?
763 // Note that I cannot call 'subsume_node' here, because
764 // that will yank the edge from old_phi to the Region and
765 // I'm mid-iteration over the Region's uses.
766 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
767 Node* use = old_phi->last_out(i);
768 igvn.hash_delete(use);
769 igvn._worklist.push(use);
770 uint uses_found = 0;
771 for (uint j = 0; j < use->len(); j++) {
772 if (use->in(j) == old_phi) {
773 if (j < use->req()) use->set_req (j, id_old_phi);
774 else use->set_prec(j, id_old_phi);
775 uses_found++;
776 }
777 }
778 i -= uses_found; // we deleted 1 or more copies of this edge
779 }
780 }
781 igvn._worklist.push(old_phi);
782 }
783 }
784 // Finally clean out the fall-in edges from the RegionNode
785 for( i = oreq-1; i>0; i-- ) {
786 if( !phase->is_member( this, _head->in(i) ) ) {
787 _head->del_req(i);
788 }
789 }
790 // Transform landing pad
791 igvn.register_new_node_with_optimizer(landing_pad, _head);
792 // Insert landing pad into the header
793 _head->add_req(landing_pad);
794 }
795
796 //------------------------------split_outer_loop-------------------------------
797 // Split out the outermost loop from this shared header.
798 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
799 PhaseIterGVN &igvn = phase->_igvn;
800
801 // Find index of outermost loop; it should also be my tail.
802 uint outer_idx = 1;
803 while( _head->in(outer_idx) != _tail ) outer_idx++;
804
805 // Make a LoopNode for the outermost loop.
806 Node *ctl = _head->in(LoopNode::EntryControl);
807 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
808 outer = igvn.register_new_node_with_optimizer(outer, _head);
809 phase->set_created_loop_node();
810 // Outermost loop falls into '_head' loop
811 _head->set_req(LoopNode::EntryControl, outer);
812 _head->del_req(outer_idx);
813 // Split all the Phis up between '_head' loop and 'outer' loop.
814 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
815 Node *out = _head->fast_out(j);
816 if( out->is_Phi() ) {
817 PhiNode *old_phi = out->as_Phi();
818 assert( old_phi->region() == _head, "" );
819 Node *phi = PhiNode::make_blank(outer, old_phi);
820 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
821 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
822 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
823 // Make old Phi point to new Phi on the fall-in path
824 igvn.hash_delete(old_phi);
825 old_phi->set_req(LoopNode::EntryControl, phi);
826 old_phi->del_req(outer_idx);
827 igvn._worklist.push(old_phi);
828 }
829 }
830
831 // Use the new loop head instead of the old shared one
832 _head = outer;
833 phase->set_loop(_head, this);
834 }
835
836 //------------------------------fix_parent-------------------------------------
837 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
838 loop->_parent = parent;
839 if( loop->_child ) fix_parent( loop->_child, loop );
840 if( loop->_next ) fix_parent( loop->_next , parent );
841 }
842
843 //------------------------------estimate_path_freq-----------------------------
844 static float estimate_path_freq( Node *n ) {
845 // Try to extract some path frequency info
846 IfNode *iff;
847 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
848 uint nop = n->Opcode();
849 if( nop == Op_SafePoint ) { // Skip any safepoint
850 n = n->in(0);
851 continue;
852 }
853 if( nop == Op_CatchProj ) { // Get count from a prior call
854 // Assume call does not always throw exceptions: means the call-site
855 // count is also the frequency of the fall-through path.
856 assert( n->is_CatchProj(), "" );
857 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
858 return 0.0f; // Assume call exception path is rare
859 Node *call = n->in(0)->in(0)->in(0);
860 assert( call->is_Call(), "expect a call here" );
861 const JVMState *jvms = ((CallNode*)call)->jvms();
862 ciMethodData* methodData = jvms->method()->method_data();
863 if (!methodData->is_mature()) return 0.0f; // No call-site data
864 ciProfileData* data = methodData->bci_to_data(jvms->bci());
865 if ((data == NULL) || !data->is_CounterData()) {
866 // no call profile available, try call's control input
867 n = n->in(0);
868 continue;
869 }
870 return data->as_CounterData()->count()/FreqCountInvocations;
871 }
872 // See if there's a gating IF test
873 Node *n_c = n->in(0);
874 if( !n_c->is_If() ) break; // No estimate available
875 iff = n_c->as_If();
876 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
877 // Compute how much count comes on this path
878 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
879 // Have no count info. Skip dull uncommon-trap like branches.
880 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
881 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
882 break;
883 // Skip through never-taken branch; look for a real loop exit.
884 n = iff->in(0);
885 }
886 return 0.0f; // No estimate available
887 }
888
889 //------------------------------merge_many_backedges---------------------------
890 // Merge all the backedges from the shared header into a private Region.
891 // Feed that region as the one backedge to this loop.
892 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
893 uint i;
894
895 // Scan for the top 2 hottest backedges
896 float hotcnt = 0.0f;
897 float warmcnt = 0.0f;
898 uint hot_idx = 0;
899 // Loop starts at 2 because slot 1 is the fall-in path
900 for( i = 2; i < _head->req(); i++ ) {
901 float cnt = estimate_path_freq(_head->in(i));
902 if( cnt > hotcnt ) { // Grab hottest path
903 warmcnt = hotcnt;
904 hotcnt = cnt;
905 hot_idx = i;
906 } else if( cnt > warmcnt ) { // And 2nd hottest path
907 warmcnt = cnt;
908 }
909 }
910
911 // See if the hottest backedge is worthy of being an inner loop
912 // by being much hotter than the next hottest backedge.
913 if( hotcnt <= 0.0001 ||
914 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
915
916 // Peel out the backedges into a private merge point; peel
917 // them all except optionally hot_idx.
918 PhaseIterGVN &igvn = phase->_igvn;
919
920 Node *hot_tail = NULL;
921 // Make a Region for the merge point
922 Node *r = new (phase->C, 1) RegionNode(1);
923 for( i = 2; i < _head->req(); i++ ) {
924 if( i != hot_idx )
925 r->add_req( _head->in(i) );
926 else hot_tail = _head->in(i);
927 }
928 igvn.register_new_node_with_optimizer(r, _head);
929 // Plug region into end of loop _head, followed by hot_tail
930 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
931 _head->set_req(2, r);
932 if( hot_idx ) _head->add_req(hot_tail);
933
934 // Split all the Phis up between '_head' loop and the Region 'r'
935 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
936 Node *out = _head->fast_out(j);
937 if( out->is_Phi() ) {
938 PhiNode* n = out->as_Phi();
939 igvn.hash_delete(n); // Delete from hash before hacking edges
940 Node *hot_phi = NULL;
941 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
942 // Check all inputs for the ones to peel out
943 uint j = 1;
944 for( uint i = 2; i < n->req(); i++ ) {
945 if( i != hot_idx )
946 phi->set_req( j++, n->in(i) );
947 else hot_phi = n->in(i);
948 }
949 // Register the phi but do not transform until whole place transforms
950 igvn.register_new_node_with_optimizer(phi, n);
951 // Add the merge phi to the old Phi
952 while( n->req() > 3 ) n->del_req( n->req()-1 );
953 n->set_req(2, phi);
954 if( hot_idx ) n->add_req(hot_phi);
955 }
956 }
957
958
959 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
960 // of self loop tree. Turn self into a loop headed by _head and with
961 // tail being the new merge point.
962 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
963 phase->set_loop(_tail,ilt); // Adjust tail
964 _tail = r; // Self's tail is new merge point
965 phase->set_loop(r,this);
966 ilt->_child = _child; // New guy has my children
967 _child = ilt; // Self has new guy as only child
968 ilt->_parent = this; // new guy has self for parent
969 ilt->_nest = _nest; // Same nesting depth (for now)
970
971 // Starting with 'ilt', look for child loop trees using the same shared
972 // header. Flatten these out; they will no longer be loops in the end.
973 IdealLoopTree **pilt = &_child;
974 while( ilt ) {
975 if( ilt->_head == _head ) {
976 uint i;
977 for( i = 2; i < _head->req(); i++ )
978 if( _head->in(i) == ilt->_tail )
979 break; // Still a loop
980 if( i == _head->req() ) { // No longer a loop
981 // Flatten ilt. Hang ilt's "_next" list from the end of
982 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
983 IdealLoopTree **cp = &ilt->_child;
984 while( *cp ) cp = &(*cp)->_next; // Find end of child list
985 *cp = ilt->_next; // Hang next list at end of child list
986 *pilt = ilt->_child; // Move child up to replace ilt
987 ilt->_head = NULL; // Flag as a loop UNIONED into parent
988 ilt = ilt->_child; // Repeat using new ilt
989 continue; // do not advance over ilt->_child
990 }
991 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
992 phase->set_loop(_head,ilt);
993 }
994 pilt = &ilt->_child; // Advance to next
995 ilt = *pilt;
996 }
997
998 if( _child ) fix_parent( _child, this );
999 }
1000
1001 //------------------------------beautify_loops---------------------------------
1002 // Split shared headers and insert loop landing pads.
1003 // Insert a LoopNode to replace the RegionNode.
1004 // Return TRUE if loop tree is structurally changed.
1005 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1006 bool result = false;
1007 // Cache parts in locals for easy
1008 PhaseIterGVN &igvn = phase->_igvn;
1009
1010 phase->C->print_method("Before beautify loops", 3);
1011
1012 igvn.hash_delete(_head); // Yank from hash before hacking edges
1013
1014 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1015 int fall_in_cnt = 0;
1016 for( uint i = 1; i < _head->req(); i++ )
1017 if( !phase->is_member( this, _head->in(i) ) )
1018 fall_in_cnt++;
1019 assert( fall_in_cnt, "at least 1 fall-in path" );
1020 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1021 split_fall_in( phase, fall_in_cnt );
1022
1023 // Swap inputs to the _head and all Phis to move the fall-in edge to
1024 // the left.
1025 fall_in_cnt = 1;
1026 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1027 fall_in_cnt++;
1028 if( fall_in_cnt > 1 ) {
1029 // Since I am just swapping inputs I do not need to update def-use info
1030 Node *tmp = _head->in(1);
1031 _head->set_req( 1, _head->in(fall_in_cnt) );
1032 _head->set_req( fall_in_cnt, tmp );
1033 // Swap also all Phis
1034 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1035 Node* phi = _head->fast_out(i);
1036 if( phi->is_Phi() ) {
1037 igvn.hash_delete(phi); // Yank from hash before hacking edges
1038 tmp = phi->in(1);
1039 phi->set_req( 1, phi->in(fall_in_cnt) );
1040 phi->set_req( fall_in_cnt, tmp );
1041 }
1042 }
1043 }
1044 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1045 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1046
1047 // If I am a shared header (multiple backedges), peel off the many
1048 // backedges into a private merge point and use the merge point as
1049 // the one true backedge.
1050 if( _head->req() > 3 ) {
1051 // Merge the many backedges into a single backedge.
1052 merge_many_backedges( phase );
1053 result = true;
1054 }
1055
1056 // If I am a shared header (multiple backedges), peel off myself loop.
1057 // I better be the outermost loop.
1058 if( _head->req() > 3 ) {
1059 split_outer_loop( phase );
1060 result = true;
1061
1062 } else if( !_head->is_Loop() && !_irreducible ) {
1063 // Make a new LoopNode to replace the old loop head
1064 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1065 l = igvn.register_new_node_with_optimizer(l, _head);
1066 phase->set_created_loop_node();
1067 // Go ahead and replace _head
1068 phase->_igvn.subsume_node( _head, l );
1069 _head = l;
1070 phase->set_loop(_head, this);
1071 for (DUIterator_Fast imax, i = l->fast_outs(imax); i < imax; i++)
1072 phase->_igvn.add_users_to_worklist(l->fast_out(i));
1073 }
1074
1075 // Now recursively beautify nested loops
1076 if( _child ) result |= _child->beautify_loops( phase );
1077 if( _next ) result |= _next ->beautify_loops( phase );
1078 return result;
1079 }
1080
1081 //------------------------------allpaths_check_safepts----------------------------
1082 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1083 // encountered. Helper for check_safepts.
1084 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1085 assert(stack.size() == 0, "empty stack");
1086 stack.push(_tail);
1087 visited.Clear();
1088 visited.set(_tail->_idx);
1089 while (stack.size() > 0) {
1090 Node* n = stack.pop();
1091 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1092 // Terminate this path
1093 } else if (n->Opcode() == Op_SafePoint) {
1094 if (_phase->get_loop(n) != this) {
1095 if (_required_safept == NULL) _required_safept = new Node_List();
1096 _required_safept->push(n); // save the one closest to the tail
1097 }
1098 // Terminate this path
1099 } else {
1100 uint start = n->is_Region() ? 1 : 0;
1101 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1102 for (uint i = start; i < end; i++) {
1103 Node* in = n->in(i);
1104 assert(in->is_CFG(), "must be");
1105 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1106 stack.push(in);
1107 }
1108 }
1109 }
1110 }
1111 }
1112
1113 //------------------------------check_safepts----------------------------
1114 // Given dominators, try to find loops with calls that must always be
1115 // executed (call dominates loop tail). These loops do not need non-call
1116 // safepoints (ncsfpt).
1117 //
1118 // A complication is that a safepoint in a inner loop may be needed
1119 // by an outer loop. In the following, the inner loop sees it has a
1120 // call (block 3) on every path from the head (block 2) to the
1121 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1122 // in block 2, _but_ this leaves the outer loop without a safepoint.
1123 //
1124 // entry 0
1125 // |
1126 // v
1127 // outer 1,2 +->1
1128 // | |
1129 // | v
1130 // | 2<---+ ncsfpt in 2
1131 // |_/|\ |
1132 // | v |
1133 // inner 2,3 / 3 | call in 3
1134 // / | |
1135 // v +--+
1136 // exit 4
1137 //
1138 //
1139 // This method creates a list (_required_safept) of ncsfpt nodes that must
1140 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1141 // is first looked for in the lists for the outer loops of the current loop.
1142 //
1143 // The insights into the problem:
1144 // A) counted loops are okay
1145 // B) innermost loops are okay (only an inner loop can delete
1146 // a ncsfpt needed by an outer loop)
1147 // C) a loop is immune from an inner loop deleting a safepoint
1148 // if the loop has a call on the idom-path
1149 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1150 // idom-path that is not in a nested loop
1151 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1152 // loop needs to be prevented from deletion by an inner loop
1153 //
1154 // There are two analyses:
1155 // 1) The first, and cheaper one, scans the loop body from
1156 // tail to head following the idom (immediate dominator)
1157 // chain, looking for the cases (C,D,E) above.
1158 // Since inner loops are scanned before outer loops, there is summary
1159 // information about inner loops. Inner loops can be skipped over
1160 // when the tail of an inner loop is encountered.
1161 //
1162 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1163 // the idom path (which is rare), scans all predecessor control paths
1164 // from the tail to the head, terminating a path when a call or sfpt
1165 // is encountered, to find the ncsfpt's that are closest to the tail.
1166 //
1167 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1168 // Bottom up traversal
1169 IdealLoopTree* ch = _child;
1170 while (ch != NULL) {
1171 ch->check_safepts(visited, stack);
1172 ch = ch->_next;
1173 }
1174
1175 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1176 bool has_call = false; // call on dom-path
1177 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1178 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1179 // Scan the dom-path nodes from tail to head
1180 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1181 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1182 has_call = true;
1183 _has_sfpt = 1; // Then no need for a safept!
1184 break;
1185 } else if (n->Opcode() == Op_SafePoint) {
1186 if (_phase->get_loop(n) == this) {
1187 has_local_ncsfpt = true;
1188 break;
1189 }
1190 if (nonlocal_ncsfpt == NULL) {
1191 nonlocal_ncsfpt = n; // save the one closest to the tail
1192 }
1193 } else {
1194 IdealLoopTree* nlpt = _phase->get_loop(n);
1195 if (this != nlpt) {
1196 // If at an inner loop tail, see if the inner loop has already
1197 // recorded seeing a call on the dom-path (and stop.) If not,
1198 // jump to the head of the inner loop.
1199 assert(is_member(nlpt), "nested loop");
1200 Node* tail = nlpt->_tail;
1201 if (tail->in(0)->is_If()) tail = tail->in(0);
1202 if (n == tail) {
1203 // If inner loop has call on dom-path, so does outer loop
1204 if (nlpt->_has_sfpt) {
1205 has_call = true;
1206 _has_sfpt = 1;
1207 break;
1208 }
1209 // Skip to head of inner loop
1210 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1211 n = nlpt->_head;
1212 }
1213 }
1214 }
1215 }
1216 // Record safept's that this loop needs preserved when an
1217 // inner loop attempts to delete it's safepoints.
1218 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1219 if (nonlocal_ncsfpt != NULL) {
1220 if (_required_safept == NULL) _required_safept = new Node_List();
1221 _required_safept->push(nonlocal_ncsfpt);
1222 } else {
1223 // Failed to find a suitable safept on the dom-path. Now use
1224 // an all paths walk from tail to head, looking for safepoints to preserve.
1225 allpaths_check_safepts(visited, stack);
1226 }
1227 }
1228 }
1229 }
1230
1231 //---------------------------is_deleteable_safept----------------------------
1232 // Is safept not required by an outer loop?
1233 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1234 assert(sfpt->Opcode() == Op_SafePoint, "");
1235 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1236 while (lp != NULL) {
1237 Node_List* sfpts = lp->_required_safept;
1238 if (sfpts != NULL) {
1239 for (uint i = 0; i < sfpts->size(); i++) {
1240 if (sfpt == sfpts->at(i))
1241 return false;
1242 }
1243 }
1244 lp = lp->_parent;
1245 }
1246 return true;
1247 }
1248
1249 //------------------------------counted_loop-----------------------------------
1250 // Convert to counted loops where possible
1251 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1252
1253 // For grins, set the inner-loop flag here
1254 if( !_child ) {
1255 if( _head->is_Loop() ) _head->as_Loop()->set_inner_loop();
1256 }
1257
1258 if( _head->is_CountedLoop() ||
1259 phase->is_counted_loop( _head, this ) ) {
1260 _has_sfpt = 1; // Indicate we do not need a safepoint here
1261
1262 // Look for a safepoint to remove
1263 for (Node* n = tail(); n != _head; n = phase->idom(n))
1264 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1265 phase->is_deleteable_safept(n))
1266 phase->lazy_replace(n,n->in(TypeFunc::Control));
1267
1268 CountedLoopNode *cl = _head->as_CountedLoop();
1269 Node *incr = cl->incr();
1270 if( !incr ) return; // Dead loop?
1271 Node *init = cl->init_trip();
1272 Node *phi = cl->phi();
1273 // protect against stride not being a constant
1274 if( !cl->stride_is_con() ) return;
1275 int stride_con = cl->stride_con();
1276
1277 // Look for induction variables
1278
1279 // Visit all children, looking for Phis
1280 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1281 Node *out = cl->out(i);
1282 if (!out->is_Phi()) continue; // Looking for phis
1283 PhiNode* phi2 = out->as_Phi();
1284 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1285 // Look for induction variables of the form: X += constant
1286 if( phi2->region() != _head ||
1287 incr2->req() != 3 ||
1288 incr2->in(1) != phi2 ||
1289 incr2 == incr ||
1290 incr2->Opcode() != Op_AddI ||
1291 !incr2->in(2)->is_Con() )
1292 continue;
1293
1294 // Check for parallel induction variable (parallel to trip counter)
1295 // via an affine function. In particular, count-down loops with
1296 // count-up array indices are common. We only RCE references off
1297 // the trip-counter, so we need to convert all these to trip-counter
1298 // expressions.
1299 Node *init2 = phi2->in( LoopNode::EntryControl );
1300 int stride_con2 = incr2->in(2)->get_int();
1301
1302 // The general case here gets a little tricky. We want to find the
1303 // GCD of all possible parallel IV's and make a new IV using this
1304 // GCD for the loop. Then all possible IVs are simple multiples of
1305 // the GCD. In practice, this will cover very few extra loops.
1306 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1307 // where +/-1 is the common case, but other integer multiples are
1308 // also easy to handle.
1309 int ratio_con = stride_con2/stride_con;
1310
1311 if( ratio_con * stride_con == stride_con2 ) { // Check for exact
1312 // Convert to using the trip counter. The parallel induction
1313 // variable differs from the trip counter by a loop-invariant
1314 // amount, the difference between their respective initial values.
1315 // It is scaled by the 'ratio_con'.
1316 Compile* C = phase->C;
1317 Node* ratio = phase->_igvn.intcon(ratio_con);
1318 phase->set_ctrl(ratio, C->root());
1319 Node* ratio_init = new (C, 3) MulINode(init, ratio);
1320 phase->_igvn.register_new_node_with_optimizer(ratio_init, init);
1321 phase->set_early_ctrl(ratio_init);
1322 Node* diff = new (C, 3) SubINode(init2, ratio_init);
1323 phase->_igvn.register_new_node_with_optimizer(diff, init2);
1324 phase->set_early_ctrl(diff);
1325 Node* ratio_idx = new (C, 3) MulINode(phi, ratio);
1326 phase->_igvn.register_new_node_with_optimizer(ratio_idx, phi);
1327 phase->set_ctrl(ratio_idx, cl);
1328 Node* add = new (C, 3) AddINode(ratio_idx, diff);
1329 phase->_igvn.register_new_node_with_optimizer(add);
1330 phase->set_ctrl(add, cl);
1331 phase->_igvn.hash_delete( phi2 );
1332 phase->_igvn.subsume_node( phi2, add );
1333 // Sometimes an induction variable is unused
1334 if (add->outcnt() == 0) {
1335 phase->_igvn.remove_dead_node(add);
1336 }
1337 --i; // deleted this phi; rescan starting with next position
1338 continue;
1339 }
1340 }
1341 } else if (_parent != NULL && !_irreducible) {
1342 // Not a counted loop.
1343 // Look for a safepoint on the idom-path to remove, preserving the first one
1344 bool found = false;
1345 Node* n = tail();
1346 for (; n != _head && !found; n = phase->idom(n)) {
1347 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1348 found = true; // Found one
1349 }
1350 // Skip past it and delete the others
1351 for (; n != _head; n = phase->idom(n)) {
1352 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1353 phase->is_deleteable_safept(n))
1354 phase->lazy_replace(n,n->in(TypeFunc::Control));
1355 }
1356 }
1357
1358 // Recursively
1359 if( _child ) _child->counted_loop( phase );
1360 if( _next ) _next ->counted_loop( phase );
1361 }
1362
1363 #ifndef PRODUCT
1364 //------------------------------dump_head--------------------------------------
1365 // Dump 1 liner for loop header info
1366 void IdealLoopTree::dump_head( ) const {
1367 for( uint i=0; i<_nest; i++ )
1368 tty->print(" ");
1369 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1370 if( _irreducible ) tty->print(" IRREDUCIBLE");
1371 if( _head->is_CountedLoop() ) {
1372 CountedLoopNode *cl = _head->as_CountedLoop();
1373 tty->print(" counted");
1374 if( cl->is_pre_loop () ) tty->print(" pre" );
1375 if( cl->is_main_loop() ) tty->print(" main");
1376 if( cl->is_post_loop() ) tty->print(" post");
1377 }
1378 tty->cr();
1379 }
1380
1381 //------------------------------dump-------------------------------------------
1382 // Dump loops by loop tree
1383 void IdealLoopTree::dump( ) const {
1384 dump_head();
1385 if( _child ) _child->dump();
1386 if( _next ) _next ->dump();
1387 }
1388
1389 #endif
1390
1391 //=============================================================================
1392 //------------------------------PhaseIdealLoop---------------------------------
1393 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1394 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1395 PhaseIdealLoop::PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs )
1396 : PhaseTransform(Ideal_Loop),
1397 _igvn(igvn),
1398 _dom_lca_tags(C->comp_arena()) {
1399 // Reset major-progress flag for the driver's heuristics
1400 C->clear_major_progress();
1401
1402 #ifndef PRODUCT
1403 // Capture for later assert
1404 uint unique = C->unique();
1405 _loop_invokes++;
1406 _loop_work += unique;
1407 #endif
1408
1409 // True if the method has at least 1 irreducible loop
1410 _has_irreducible_loops = false;
1411
1412 _created_loop_node = false;
1413
1414 Arena *a = Thread::current()->resource_area();
1415 VectorSet visited(a);
1416 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1417 _nodes.map(C->unique(), NULL);
1418 memset(_nodes.adr(), 0, wordSize * C->unique());
1419
1420 // Pre-build the top-level outermost loop tree entry
1421 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1422 // Do not need a safepoint at the top level
1423 _ltree_root->_has_sfpt = 1;
1424
1425 // Empty pre-order array
1426 allocate_preorders();
1427
1428 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1429 // IdealLoopTree entries. Data nodes are NOT walked.
1430 build_loop_tree();
1431 // Check for bailout, and return
1432 if (C->failing()) {
1433 return;
1434 }
1435
1436 // No loops after all
1437 if( !_ltree_root->_child ) C->set_has_loops(false);
1438
1439 // There should always be an outer loop containing the Root and Return nodes.
1440 // If not, we have a degenerate empty program. Bail out in this case.
1441 if (!has_node(C->root())) {
1442 C->clear_major_progress();
1443 C->record_method_not_compilable("empty program detected during loop optimization");
1444 return;
1445 }
1446
1447 // Nothing to do, so get out
1448 if( !C->has_loops() && !do_split_ifs && !verify_me) {
1449 _igvn.optimize(); // Cleanup NeverBranches
1450 return;
1451 }
1452
1453 // Set loop nesting depth
1454 _ltree_root->set_nest( 0 );
1455
1456 // Split shared headers and insert loop landing pads.
1457 // Do not bother doing this on the Root loop of course.
1458 if( !verify_me && _ltree_root->_child ) {
1459 if( _ltree_root->_child->beautify_loops( this ) ) {
1460 // Re-build loop tree!
1461 _ltree_root->_child = NULL;
1462 _nodes.clear();
1463 reallocate_preorders();
1464 build_loop_tree();
1465 // Check for bailout, and return
1466 if (C->failing()) {
1467 return;
1468 }
1469 // Reset loop nesting depth
1470 _ltree_root->set_nest( 0 );
1471
1472 C->print_method("After beautify loops", 3);
1473 }
1474 }
1475
1476 // Build Dominators for elision of NULL checks & loop finding.
1477 // Since nodes do not have a slot for immediate dominator, make
1478 // a persistant side array for that info indexed on node->_idx.
1479 _idom_size = C->unique();
1480 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1481 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1482 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1483 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1484
1485 Dominators();
1486
1487 // As a side effect, Dominators removed any unreachable CFG paths
1488 // into RegionNodes. It doesn't do this test against Root, so
1489 // we do it here.
1490 for( uint i = 1; i < C->root()->req(); i++ ) {
1491 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1492 _igvn.hash_delete(C->root());
1493 C->root()->del_req(i);
1494 _igvn._worklist.push(C->root());
1495 i--; // Rerun same iteration on compressed edges
1496 }
1497 }
1498
1499 // Given dominators, try to find inner loops with calls that must
1500 // always be executed (call dominates loop tail). These loops do
1501 // not need a seperate safepoint.
1502 Node_List cisstack(a);
1503 _ltree_root->check_safepts(visited, cisstack);
1504
1505 // Walk the DATA nodes and place into loops. Find earliest control
1506 // node. For CFG nodes, the _nodes array starts out and remains
1507 // holding the associated IdealLoopTree pointer. For DATA nodes, the
1508 // _nodes array holds the earliest legal controlling CFG node.
1509
1510 // Allocate stack with enough space to avoid frequent realloc
1511 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
1512 Node_Stack nstack( a, stack_size );
1513
1514 visited.Clear();
1515 Node_List worklist(a);
1516 // Don't need C->root() on worklist since
1517 // it will be processed among C->top() inputs
1518 worklist.push( C->top() );
1519 visited.set( C->top()->_idx ); // Set C->top() as visited now
1520 build_loop_early( visited, worklist, nstack, verify_me );
1521
1522 // Given early legal placement, try finding counted loops. This placement
1523 // is good enough to discover most loop invariants.
1524 if( !verify_me )
1525 _ltree_root->counted_loop( this );
1526
1527 // Find latest loop placement. Find ideal loop placement.
1528 visited.Clear();
1529 init_dom_lca_tags();
1530 // Need C->root() on worklist when processing outs
1531 worklist.push( C->root() );
1532 NOT_PRODUCT( C->verify_graph_edges(); )
1533 worklist.push( C->top() );
1534 build_loop_late( visited, worklist, nstack, verify_me );
1535
1536 // clear out the dead code
1537 while(_deadlist.size()) {
1538 igvn.remove_globally_dead_node(_deadlist.pop());
1539 }
1540
1541 #ifndef PRODUCT
1542 C->verify_graph_edges();
1543 if( verify_me ) { // Nested verify pass?
1544 // Check to see if the verify mode is broken
1545 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
1546 return;
1547 }
1548 if( VerifyLoopOptimizations ) verify();
1549 #endif
1550
1551 if (ReassociateInvariants) {
1552 // Reassociate invariants and prep for split_thru_phi
1553 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1554 IdealLoopTree* lpt = iter.current();
1555 if (!lpt->is_counted() || !lpt->is_inner()) continue;
1556
1557 lpt->reassociate_invariants(this);
1558
1559 // Because RCE opportunities can be masked by split_thru_phi,
1560 // look for RCE candidates and inhibit split_thru_phi
1561 // on just their loop-phi's for this pass of loop opts
1562 if( SplitIfBlocks && do_split_ifs ) {
1563 if (lpt->policy_range_check(this)) {
1564 lpt->_rce_candidate = 1; // = true
1565 }
1566 }
1567 }
1568 }
1569
1570 // Check for aggressive application of split-if and other transforms
1571 // that require basic-block info (like cloning through Phi's)
1572 if( SplitIfBlocks && do_split_ifs ) {
1573 visited.Clear();
1574 split_if_with_blocks( visited, nstack );
1575 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
1576 }
1577
1578 // Perform iteration-splitting on inner loops. Split iterations to avoid
1579 // range checks or one-shot null checks.
1580
1581 // If split-if's didn't hack the graph too bad (no CFG changes)
1582 // then do loop opts.
1583 if( C->has_loops() && !C->major_progress() ) {
1584 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
1585 _ltree_root->_child->iteration_split( this, worklist );
1586 // No verify after peeling! GCM has hoisted code out of the loop.
1587 // After peeling, the hoisted code could sink inside the peeled area.
1588 // The peeling code does not try to recompute the best location for
1589 // all the code before the peeled area, so the verify pass will always
1590 // complain about it.
1591 }
1592 // Do verify graph edges in any case
1593 NOT_PRODUCT( C->verify_graph_edges(); );
1594
1595 if( !do_split_ifs ) {
1596 // We saw major progress in Split-If to get here. We forced a
1597 // pass with unrolling and not split-if, however more split-if's
1598 // might make progress. If the unrolling didn't make progress
1599 // then the major-progress flag got cleared and we won't try
1600 // another round of Split-If. In particular the ever-common
1601 // instance-of/check-cast pattern requires at least 2 rounds of
1602 // Split-If to clear out.
1603 C->set_major_progress();
1604 }
1605
1606 // Repeat loop optimizations if new loops were seen
1607 if (created_loop_node()) {
1608 C->set_major_progress();
1609 }
1610
1611 // Convert scalar to superword operations
1612
1613 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
1614 // SuperWord transform
1615 SuperWord sw(this);
1616 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1617 IdealLoopTree* lpt = iter.current();
1618 if (lpt->is_counted()) {
1619 sw.transform_loop(lpt);
1620 }
1621 }
1622 }
1623
1624 // Cleanup any modified bits
1625 _igvn.optimize();
1626
1627 // Do not repeat loop optimizations if irreducible loops are present
1628 // by claiming no-progress.
1629 if( _has_irreducible_loops )
1630 C->clear_major_progress();
1631 }
1632
1633 #ifndef PRODUCT
1634 //------------------------------print_statistics-------------------------------
1635 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
1636 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
1637 void PhaseIdealLoop::print_statistics() {
1638 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
1639 }
1640
1641 //------------------------------verify-----------------------------------------
1642 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
1643 static int fail; // debug only, so its multi-thread dont care
1644 void PhaseIdealLoop::verify() const {
1645 int old_progress = C->major_progress();
1646 ResourceMark rm;
1647 PhaseIdealLoop loop_verify( _igvn, this, false );
1648 VectorSet visited(Thread::current()->resource_area());
1649
1650 fail = 0;
1651 verify_compare( C->root(), &loop_verify, visited );
1652 assert( fail == 0, "verify loops failed" );
1653 // Verify loop structure is the same
1654 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
1655 // Reset major-progress. It was cleared by creating a verify version of
1656 // PhaseIdealLoop.
1657 for( int i=0; i<old_progress; i++ )
1658 C->set_major_progress();
1659 }
1660
1661 //------------------------------verify_compare---------------------------------
1662 // Make sure me and the given PhaseIdealLoop agree on key data structures
1663 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
1664 if( !n ) return;
1665 if( visited.test_set( n->_idx ) ) return;
1666 if( !_nodes[n->_idx] ) { // Unreachable
1667 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
1668 return;
1669 }
1670
1671 uint i;
1672 for( i = 0; i < n->req(); i++ )
1673 verify_compare( n->in(i), loop_verify, visited );
1674
1675 // Check the '_nodes' block/loop structure
1676 i = n->_idx;
1677 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
1678 if( _nodes[i] != loop_verify->_nodes[i] &&
1679 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
1680 tty->print("Mismatched control setting for: ");
1681 n->dump();
1682 if( fail++ > 10 ) return;
1683 Node *c = get_ctrl_no_update(n);
1684 tty->print("We have it as: ");
1685 if( c->in(0) ) c->dump();
1686 else tty->print_cr("N%d",c->_idx);
1687 tty->print("Verify thinks: ");
1688 if( loop_verify->has_ctrl(n) )
1689 loop_verify->get_ctrl_no_update(n)->dump();
1690 else
1691 loop_verify->get_loop_idx(n)->dump();
1692 tty->cr();
1693 }
1694 } else { // We have a loop
1695 IdealLoopTree *us = get_loop_idx(n);
1696 if( loop_verify->has_ctrl(n) ) {
1697 tty->print("Mismatched loop setting for: ");
1698 n->dump();
1699 if( fail++ > 10 ) return;
1700 tty->print("We have it as: ");
1701 us->dump();
1702 tty->print("Verify thinks: ");
1703 loop_verify->get_ctrl_no_update(n)->dump();
1704 tty->cr();
1705 } else if (!C->major_progress()) {
1706 // Loop selection can be messed up if we did a major progress
1707 // operation, like split-if. Do not verify in that case.
1708 IdealLoopTree *them = loop_verify->get_loop_idx(n);
1709 if( us->_head != them->_head || us->_tail != them->_tail ) {
1710 tty->print("Unequals loops for: ");
1711 n->dump();
1712 if( fail++ > 10 ) return;
1713 tty->print("We have it as: ");
1714 us->dump();
1715 tty->print("Verify thinks: ");
1716 them->dump();
1717 tty->cr();
1718 }
1719 }
1720 }
1721
1722 // Check for immediate dominators being equal
1723 if( i >= _idom_size ) {
1724 if( !n->is_CFG() ) return;
1725 tty->print("CFG Node with no idom: ");
1726 n->dump();
1727 return;
1728 }
1729 if( !n->is_CFG() ) return;
1730 if( n == C->root() ) return; // No IDOM here
1731
1732 assert(n->_idx == i, "sanity");
1733 Node *id = idom_no_update(n);
1734 if( id != loop_verify->idom_no_update(n) ) {
1735 tty->print("Unequals idoms for: ");
1736 n->dump();
1737 if( fail++ > 10 ) return;
1738 tty->print("We have it as: ");
1739 id->dump();
1740 tty->print("Verify thinks: ");
1741 loop_verify->idom_no_update(n)->dump();
1742 tty->cr();
1743 }
1744
1745 }
1746
1747 //------------------------------verify_tree------------------------------------
1748 // Verify that tree structures match. Because the CFG can change, siblings
1749 // within the loop tree can be reordered. We attempt to deal with that by
1750 // reordering the verify's loop tree if possible.
1751 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
1752 assert( _parent == parent, "Badly formed loop tree" );
1753
1754 // Siblings not in same order? Attempt to re-order.
1755 if( _head != loop->_head ) {
1756 // Find _next pointer to update
1757 IdealLoopTree **pp = &loop->_parent->_child;
1758 while( *pp != loop )
1759 pp = &((*pp)->_next);
1760 // Find proper sibling to be next
1761 IdealLoopTree **nn = &loop->_next;
1762 while( (*nn) && (*nn)->_head != _head )
1763 nn = &((*nn)->_next);
1764
1765 // Check for no match.
1766 if( !(*nn) ) {
1767 // Annoyingly, irreducible loops can pick different headers
1768 // after a major_progress operation, so the rest of the loop
1769 // tree cannot be matched.
1770 if (_irreducible && Compile::current()->major_progress()) return;
1771 assert( 0, "failed to match loop tree" );
1772 }
1773
1774 // Move (*nn) to (*pp)
1775 IdealLoopTree *hit = *nn;
1776 *nn = hit->_next;
1777 hit->_next = loop;
1778 *pp = loop;
1779 loop = hit;
1780 // Now try again to verify
1781 }
1782
1783 assert( _head == loop->_head , "mismatched loop head" );
1784 Node *tail = _tail; // Inline a non-updating version of
1785 while( !tail->in(0) ) // the 'tail()' call.
1786 tail = tail->in(1);
1787 assert( tail == loop->_tail, "mismatched loop tail" );
1788
1789 // Counted loops that are guarded should be able to find their guards
1790 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
1791 CountedLoopNode *cl = _head->as_CountedLoop();
1792 Node *init = cl->init_trip();
1793 Node *ctrl = cl->in(LoopNode::EntryControl);
1794 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1795 Node *iff = ctrl->in(0);
1796 assert( iff->Opcode() == Op_If, "" );
1797 Node *bol = iff->in(1);
1798 assert( bol->Opcode() == Op_Bool, "" );
1799 Node *cmp = bol->in(1);
1800 assert( cmp->Opcode() == Op_CmpI, "" );
1801 Node *add = cmp->in(1);
1802 Node *opaq;
1803 if( add->Opcode() == Op_Opaque1 ) {
1804 opaq = add;
1805 } else {
1806 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
1807 assert( add == init, "" );
1808 opaq = cmp->in(2);
1809 }
1810 assert( opaq->Opcode() == Op_Opaque1, "" );
1811
1812 }
1813
1814 if (_child != NULL) _child->verify_tree(loop->_child, this);
1815 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
1816 // Innermost loops need to verify loop bodies,
1817 // but only if no 'major_progress'
1818 int fail = 0;
1819 if (!Compile::current()->major_progress() && _child == NULL) {
1820 for( uint i = 0; i < _body.size(); i++ ) {
1821 Node *n = _body.at(i);
1822 if (n->outcnt() == 0) continue; // Ignore dead
1823 uint j;
1824 for( j = 0; j < loop->_body.size(); j++ )
1825 if( loop->_body.at(j) == n )
1826 break;
1827 if( j == loop->_body.size() ) { // Not found in loop body
1828 // Last ditch effort to avoid assertion: Its possible that we
1829 // have some users (so outcnt not zero) but are still dead.
1830 // Try to find from root.
1831 if (Compile::current()->root()->find(n->_idx)) {
1832 fail++;
1833 tty->print("We have that verify does not: ");
1834 n->dump();
1835 }
1836 }
1837 }
1838 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
1839 Node *n = loop->_body.at(i2);
1840 if (n->outcnt() == 0) continue; // Ignore dead
1841 uint j;
1842 for( j = 0; j < _body.size(); j++ )
1843 if( _body.at(j) == n )
1844 break;
1845 if( j == _body.size() ) { // Not found in loop body
1846 // Last ditch effort to avoid assertion: Its possible that we
1847 // have some users (so outcnt not zero) but are still dead.
1848 // Try to find from root.
1849 if (Compile::current()->root()->find(n->_idx)) {
1850 fail++;
1851 tty->print("Verify has that we do not: ");
1852 n->dump();
1853 }
1854 }
1855 }
1856 assert( !fail, "loop body mismatch" );
1857 }
1858 }
1859
1860 #endif
1861
1862 //------------------------------set_idom---------------------------------------
1863 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
1864 uint idx = d->_idx;
1865 if (idx >= _idom_size) {
1866 uint newsize = _idom_size<<1;
1867 while( idx >= newsize ) {
1868 newsize <<= 1;
1869 }
1870 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
1871 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
1872 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
1873 _idom_size = newsize;
1874 }
1875 _idom[idx] = n;
1876 _dom_depth[idx] = dom_depth;
1877 }
1878
1879 //------------------------------recompute_dom_depth---------------------------------------
1880 // The dominator tree is constructed with only parent pointers.
1881 // This recomputes the depth in the tree by first tagging all
1882 // nodes as "no depth yet" marker. The next pass then runs up
1883 // the dom tree from each node marked "no depth yet", and computes
1884 // the depth on the way back down.
1885 void PhaseIdealLoop::recompute_dom_depth() {
1886 uint no_depth_marker = C->unique();
1887 uint i;
1888 // Initialize depth to "no depth yet"
1889 for (i = 0; i < _idom_size; i++) {
1890 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
1891 _dom_depth[i] = no_depth_marker;
1892 }
1893 }
1894 if (_dom_stk == NULL) {
1895 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
1896 if (init_size < 10) init_size = 10;
1897 _dom_stk = new (C->node_arena()) GrowableArray<uint>(C->node_arena(), init_size, 0, 0);
1898 }
1899 // Compute new depth for each node.
1900 for (i = 0; i < _idom_size; i++) {
1901 uint j = i;
1902 // Run up the dom tree to find a node with a depth
1903 while (_dom_depth[j] == no_depth_marker) {
1904 _dom_stk->push(j);
1905 j = _idom[j]->_idx;
1906 }
1907 // Compute the depth on the way back down this tree branch
1908 uint dd = _dom_depth[j] + 1;
1909 while (_dom_stk->length() > 0) {
1910 uint j = _dom_stk->pop();
1911 _dom_depth[j] = dd;
1912 dd++;
1913 }
1914 }
1915 }
1916
1917 //------------------------------sort-------------------------------------------
1918 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
1919 // loop tree, not the root.
1920 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
1921 if( !innermost ) return loop; // New innermost loop
1922
1923 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
1924 assert( loop_preorder, "not yet post-walked loop" );
1925 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
1926 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
1927
1928 // Insert at start of list
1929 while( l ) { // Insertion sort based on pre-order
1930 if( l == loop ) return innermost; // Already on list!
1931 int l_preorder = get_preorder(l->_head); // Cache pre-order number
1932 assert( l_preorder, "not yet post-walked l" );
1933 // Check header pre-order number to figure proper nesting
1934 if( loop_preorder > l_preorder )
1935 break; // End of insertion
1936 // If headers tie (e.g., shared headers) check tail pre-order numbers.
1937 // Since I split shared headers, you'd think this could not happen.
1938 // BUT: I must first do the preorder numbering before I can discover I
1939 // have shared headers, so the split headers all get the same preorder
1940 // number as the RegionNode they split from.
1941 if( loop_preorder == l_preorder &&
1942 get_preorder(loop->_tail) < get_preorder(l->_tail) )
1943 break; // Also check for shared headers (same pre#)
1944 pp = &l->_parent; // Chain up list
1945 l = *pp;
1946 }
1947 // Link into list
1948 // Point predecessor to me
1949 *pp = loop;
1950 // Point me to successor
1951 IdealLoopTree *p = loop->_parent;
1952 loop->_parent = l; // Point me to successor
1953 if( p ) sort( p, innermost ); // Insert my parents into list as well
1954 return innermost;
1955 }
1956
1957 //------------------------------build_loop_tree--------------------------------
1958 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
1959 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
1960 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
1961 // tightest enclosing IdealLoopTree for post-walked.
1962 //
1963 // During my forward walk I do a short 1-layer lookahead to see if I can find
1964 // a loop backedge with that doesn't have any work on the backedge. This
1965 // helps me construct nested loops with shared headers better.
1966 //
1967 // Once I've done the forward recursion, I do the post-work. For each child
1968 // I check to see if there is a backedge. Backedges define a loop! I
1969 // insert an IdealLoopTree at the target of the backedge.
1970 //
1971 // During the post-work I also check to see if I have several children
1972 // belonging to different loops. If so, then this Node is a decision point
1973 // where control flow can choose to change loop nests. It is at this
1974 // decision point where I can figure out how loops are nested. At this
1975 // time I can properly order the different loop nests from my children.
1976 // Note that there may not be any backedges at the decision point!
1977 //
1978 // Since the decision point can be far removed from the backedges, I can't
1979 // order my loops at the time I discover them. Thus at the decision point
1980 // I need to inspect loop header pre-order numbers to properly nest my
1981 // loops. This means I need to sort my childrens' loops by pre-order.
1982 // The sort is of size number-of-control-children, which generally limits
1983 // it to size 2 (i.e., I just choose between my 2 target loops).
1984 void PhaseIdealLoop::build_loop_tree() {
1985 // Allocate stack of size C->unique()/2 to avoid frequent realloc
1986 GrowableArray <Node *> bltstack(C->unique() >> 1);
1987 Node *n = C->root();
1988 bltstack.push(n);
1989 int pre_order = 1;
1990 int stack_size;
1991
1992 while ( ( stack_size = bltstack.length() ) != 0 ) {
1993 n = bltstack.top(); // Leave node on stack
1994 if ( !is_visited(n) ) {
1995 // ---- Pre-pass Work ----
1996 // Pre-walked but not post-walked nodes need a pre_order number.
1997
1998 set_preorder_visited( n, pre_order ); // set as visited
1999
2000 // ---- Scan over children ----
2001 // Scan first over control projections that lead to loop headers.
2002 // This helps us find inner-to-outer loops with shared headers better.
2003
2004 // Scan children's children for loop headers.
2005 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2006 Node* m = n->raw_out(i); // Child
2007 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2008 // Scan over children's children to find loop
2009 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2010 Node* l = m->fast_out(j);
2011 if( is_visited(l) && // Been visited?
2012 !is_postvisited(l) && // But not post-visited
2013 get_preorder(l) < pre_order ) { // And smaller pre-order
2014 // Found! Scan the DFS down this path before doing other paths
2015 bltstack.push(m);
2016 break;
2017 }
2018 }
2019 }
2020 }
2021 pre_order++;
2022 }
2023 else if ( !is_postvisited(n) ) {
2024 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2025 // such as com.sun.rsasign.am::a.
2026 // For non-recursive version, first, process current children.
2027 // On next iteration, check if additional children were added.
2028 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2029 Node* u = n->raw_out(k);
2030 if ( u->is_CFG() && !is_visited(u) ) {
2031 bltstack.push(u);
2032 }
2033 }
2034 if ( bltstack.length() == stack_size ) {
2035 // There were no additional children, post visit node now
2036 (void)bltstack.pop(); // Remove node from stack
2037 pre_order = build_loop_tree_impl( n, pre_order );
2038 // Check for bailout
2039 if (C->failing()) {
2040 return;
2041 }
2042 // Check to grow _preorders[] array for the case when
2043 // build_loop_tree_impl() adds new nodes.
2044 check_grow_preorders();
2045 }
2046 }
2047 else {
2048 (void)bltstack.pop(); // Remove post-visited node from stack
2049 }
2050 }
2051 }
2052
2053 //------------------------------build_loop_tree_impl---------------------------
2054 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2055 // ---- Post-pass Work ----
2056 // Pre-walked but not post-walked nodes need a pre_order number.
2057
2058 // Tightest enclosing loop for this Node
2059 IdealLoopTree *innermost = NULL;
2060
2061 // For all children, see if any edge is a backedge. If so, make a loop
2062 // for it. Then find the tightest enclosing loop for the self Node.
2063 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2064 Node* m = n->fast_out(i); // Child
2065 if( n == m ) continue; // Ignore control self-cycles
2066 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2067
2068 IdealLoopTree *l; // Child's loop
2069 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2070 // Found a backedge
2071 assert( get_preorder(m) < pre_order, "should be backedge" );
2072 // Check for the RootNode, which is already a LoopNode and is allowed
2073 // to have multiple "backedges".
2074 if( m == C->root()) { // Found the root?
2075 l = _ltree_root; // Root is the outermost LoopNode
2076 } else { // Else found a nested loop
2077 // Insert a LoopNode to mark this loop.
2078 l = new IdealLoopTree(this, m, n);
2079 } // End of Else found a nested loop
2080 if( !has_loop(m) ) // If 'm' does not already have a loop set
2081 set_loop(m, l); // Set loop header to loop now
2082
2083 } else { // Else not a nested loop
2084 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2085 l = get_loop(m); // Get previously determined loop
2086 // If successor is header of a loop (nest), move up-loop till it
2087 // is a member of some outer enclosing loop. Since there are no
2088 // shared headers (I've split them already) I only need to go up
2089 // at most 1 level.
2090 while( l && l->_head == m ) // Successor heads loop?
2091 l = l->_parent; // Move up 1 for me
2092 // If this loop is not properly parented, then this loop
2093 // has no exit path out, i.e. its an infinite loop.
2094 if( !l ) {
2095 // Make loop "reachable" from root so the CFG is reachable. Basically
2096 // insert a bogus loop exit that is never taken. 'm', the loop head,
2097 // points to 'n', one (of possibly many) fall-in paths. There may be
2098 // many backedges as well.
2099
2100 // Here I set the loop to be the root loop. I could have, after
2101 // inserting a bogus loop exit, restarted the recursion and found my
2102 // new loop exit. This would make the infinite loop a first-class
2103 // loop and it would then get properly optimized. What's the use of
2104 // optimizing an infinite loop?
2105 l = _ltree_root; // Oops, found infinite loop
2106
2107 // Insert the NeverBranch between 'm' and it's control user.
2108 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2109 _igvn.register_new_node_with_optimizer(iff);
2110 set_loop(iff, l);
2111 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2112 _igvn.register_new_node_with_optimizer(if_t);
2113 set_loop(if_t, l);
2114
2115 Node* cfg = NULL; // Find the One True Control User of m
2116 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2117 Node* x = m->fast_out(j);
2118 if (x->is_CFG() && x != m && x != iff)
2119 { cfg = x; break; }
2120 }
2121 assert(cfg != NULL, "must find the control user of m");
2122 uint k = 0; // Probably cfg->in(0)
2123 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2124 cfg->set_req( k, if_t ); // Now point to NeverBranch
2125
2126 // Now create the never-taken loop exit
2127 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2128 _igvn.register_new_node_with_optimizer(if_f);
2129 set_loop(if_f, l);
2130 // Find frame ptr for Halt. Relies on the optimizer
2131 // V-N'ing. Easier and quicker than searching through
2132 // the program structure.
2133 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2134 _igvn.register_new_node_with_optimizer(frame);
2135 // Halt & Catch Fire
2136 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2137 _igvn.register_new_node_with_optimizer(halt);
2138 set_loop(halt, l);
2139 C->root()->add_req(halt);
2140 set_loop(C->root(), _ltree_root);
2141 }
2142 }
2143 // Weeny check for irreducible. This child was already visited (this
2144 // IS the post-work phase). Is this child's loop header post-visited
2145 // as well? If so, then I found another entry into the loop.
2146 while( is_postvisited(l->_head) ) {
2147 // found irreducible
2148 l->_irreducible = 1; // = true
2149 l = l->_parent;
2150 _has_irreducible_loops = true;
2151 // Check for bad CFG here to prevent crash, and bailout of compile
2152 if (l == NULL) {
2153 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2154 return pre_order;
2155 }
2156 }
2157
2158 // This Node might be a decision point for loops. It is only if
2159 // it's children belong to several different loops. The sort call
2160 // does a trivial amount of work if there is only 1 child or all
2161 // children belong to the same loop. If however, the children
2162 // belong to different loops, the sort call will properly set the
2163 // _parent pointers to show how the loops nest.
2164 //
2165 // In any case, it returns the tightest enclosing loop.
2166 innermost = sort( l, innermost );
2167 }
2168
2169 // Def-use info will have some dead stuff; dead stuff will have no
2170 // loop decided on.
2171
2172 // Am I a loop header? If so fix up my parent's child and next ptrs.
2173 if( innermost && innermost->_head == n ) {
2174 assert( get_loop(n) == innermost, "" );
2175 IdealLoopTree *p = innermost->_parent;
2176 IdealLoopTree *l = innermost;
2177 while( p && l->_head == n ) {
2178 l->_next = p->_child; // Put self on parents 'next child'
2179 p->_child = l; // Make self as first child of parent
2180 l = p; // Now walk up the parent chain
2181 p = l->_parent;
2182 }
2183 } else {
2184 // Note that it is possible for a LoopNode to reach here, if the
2185 // backedge has been made unreachable (hence the LoopNode no longer
2186 // denotes a Loop, and will eventually be removed).
2187
2188 // Record tightest enclosing loop for self. Mark as post-visited.
2189 set_loop(n, innermost);
2190 // Also record has_call flag early on
2191 if( innermost ) {
2192 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2193 // Do not count uncommon calls
2194 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2195 Node *iff = n->in(0)->in(0);
2196 if( !iff->is_If() ||
2197 (n->in(0)->Opcode() == Op_IfFalse &&
2198 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2199 (iff->as_If()->_prob >= 0.01) )
2200 innermost->_has_call = 1;
2201 }
2202 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2203 // Disable loop optimizations if the loop has a scalar replaceable
2204 // allocation. This disabling may cause a potential performance lost
2205 // if the allocation is not eliminated for some reason.
2206 innermost->_allow_optimizations = false;
2207 innermost->_has_call = 1; // = true
2208 }
2209 }
2210 }
2211
2212 // Flag as post-visited now
2213 set_postvisited(n);
2214 return pre_order;
2215 }
2216
2217
2218 //------------------------------build_loop_early-------------------------------
2219 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2220 // First pass computes the earliest controlling node possible. This is the
2221 // controlling input with the deepest dominating depth.
2222 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) {
2223 while (worklist.size() != 0) {
2224 // Use local variables nstack_top_n & nstack_top_i to cache values
2225 // on nstack's top.
2226 Node *nstack_top_n = worklist.pop();
2227 uint nstack_top_i = 0;
2228 //while_nstack_nonempty:
2229 while (true) {
2230 // Get parent node and next input's index from stack's top.
2231 Node *n = nstack_top_n;
2232 uint i = nstack_top_i;
2233 uint cnt = n->req(); // Count of inputs
2234 if (i == 0) { // Pre-process the node.
2235 if( has_node(n) && // Have either loop or control already?
2236 !has_ctrl(n) ) { // Have loop picked out already?
2237 // During "merge_many_backedges" we fold up several nested loops
2238 // into a single loop. This makes the members of the original
2239 // loop bodies pointing to dead loops; they need to move up
2240 // to the new UNION'd larger loop. I set the _head field of these
2241 // dead loops to NULL and the _parent field points to the owning
2242 // loop. Shades of UNION-FIND algorithm.
2243 IdealLoopTree *ilt;
2244 while( !(ilt = get_loop(n))->_head ) {
2245 // Normally I would use a set_loop here. But in this one special
2246 // case, it is legal (and expected) to change what loop a Node
2247 // belongs to.
2248 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2249 }
2250 // Remove safepoints ONLY if I've already seen I don't need one.
2251 // (the old code here would yank a 2nd safepoint after seeing a
2252 // first one, even though the 1st did not dominate in the loop body
2253 // and thus could be avoided indefinitely)
2254 if( !verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2255 is_deleteable_safept(n)) {
2256 Node *in = n->in(TypeFunc::Control);
2257 lazy_replace(n,in); // Pull safepoint now
2258 // Carry on with the recursion "as if" we are walking
2259 // only the control input
2260 if( !visited.test_set( in->_idx ) ) {
2261 worklist.push(in); // Visit this guy later, using worklist
2262 }
2263 // Get next node from nstack:
2264 // - skip n's inputs processing by setting i > cnt;
2265 // - we also will not call set_early_ctrl(n) since
2266 // has_node(n) == true (see the condition above).
2267 i = cnt + 1;
2268 }
2269 }
2270 } // if (i == 0)
2271
2272 // Visit all inputs
2273 bool done = true; // Assume all n's inputs will be processed
2274 while (i < cnt) {
2275 Node *in = n->in(i);
2276 ++i;
2277 if (in == NULL) continue;
2278 if (in->pinned() && !in->is_CFG())
2279 set_ctrl(in, in->in(0));
2280 int is_visited = visited.test_set( in->_idx );
2281 if (!has_node(in)) { // No controlling input yet?
2282 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2283 assert( !is_visited, "visit only once" );
2284 nstack.push(n, i); // Save parent node and next input's index.
2285 nstack_top_n = in; // Process current input now.
2286 nstack_top_i = 0;
2287 done = false; // Not all n's inputs processed.
2288 break; // continue while_nstack_nonempty;
2289 } else if (!is_visited) {
2290 // This guy has a location picked out for him, but has not yet
2291 // been visited. Happens to all CFG nodes, for instance.
2292 // Visit him using the worklist instead of recursion, to break
2293 // cycles. Since he has a location already we do not need to
2294 // find his location before proceeding with the current Node.
2295 worklist.push(in); // Visit this guy later, using worklist
2296 }
2297 }
2298 if (done) {
2299 // All of n's inputs have been processed, complete post-processing.
2300
2301 // Compute earilest point this Node can go.
2302 // CFG, Phi, pinned nodes already know their controlling input.
2303 if (!has_node(n)) {
2304 // Record earliest legal location
2305 set_early_ctrl( n );
2306 }
2307 if (nstack.is_empty()) {
2308 // Finished all nodes on stack.
2309 // Process next node on the worklist.
2310 break;
2311 }
2312 // Get saved parent node and next input's index.
2313 nstack_top_n = nstack.node();
2314 nstack_top_i = nstack.index();
2315 nstack.pop();
2316 }
2317 } // while (true)
2318 }
2319 }
2320
2321 //------------------------------dom_lca_internal--------------------------------
2322 // Pair-wise LCA
2323 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2324 if( !n1 ) return n2; // Handle NULL original LCA
2325 assert( n1->is_CFG(), "" );
2326 assert( n2->is_CFG(), "" );
2327 // find LCA of all uses
2328 uint d1 = dom_depth(n1);
2329 uint d2 = dom_depth(n2);
2330 while (n1 != n2) {
2331 if (d1 > d2) {
2332 n1 = idom(n1);
2333 d1 = dom_depth(n1);
2334 } else if (d1 < d2) {
2335 n2 = idom(n2);
2336 d2 = dom_depth(n2);
2337 } else {
2338 // Here d1 == d2. Due to edits of the dominator-tree, sections
2339 // of the tree might have the same depth. These sections have
2340 // to be searched more carefully.
2341
2342 // Scan up all the n1's with equal depth, looking for n2.
2343 Node *t1 = idom(n1);
2344 while (dom_depth(t1) == d1) {
2345 if (t1 == n2) return n2;
2346 t1 = idom(t1);
2347 }
2348 // Scan up all the n2's with equal depth, looking for n1.
2349 Node *t2 = idom(n2);
2350 while (dom_depth(t2) == d2) {
2351 if (t2 == n1) return n1;
2352 t2 = idom(t2);
2353 }
2354 // Move up to a new dominator-depth value as well as up the dom-tree.
2355 n1 = t1;
2356 n2 = t2;
2357 d1 = dom_depth(n1);
2358 d2 = dom_depth(n2);
2359 }
2360 }
2361 return n1;
2362 }
2363
2364 //------------------------------compute_idom-----------------------------------
2365 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2366 // IDOMs are correct.
2367 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2368 assert( region->is_Region(), "" );
2369 Node *LCA = NULL;
2370 for( uint i = 1; i < region->req(); i++ ) {
2371 if( region->in(i) != C->top() )
2372 LCA = dom_lca( LCA, region->in(i) );
2373 }
2374 return LCA;
2375 }
2376
2377 //------------------------------get_late_ctrl----------------------------------
2378 // Compute latest legal control.
2379 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2380 assert(early != NULL, "early control should not be NULL");
2381
2382 // Compute LCA over list of uses
2383 Node *LCA = NULL;
2384 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2385 Node* c = n->fast_out(i);
2386 if (_nodes[c->_idx] == NULL)
2387 continue; // Skip the occasional dead node
2388 if( c->is_Phi() ) { // For Phis, we must land above on the path
2389 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2390 if( c->in(j) == n ) { // Found matching input?
2391 Node *use = c->in(0)->in(j);
2392 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2393 }
2394 }
2395 } else {
2396 // For CFG data-users, use is in the block just prior
2397 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2398 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2399 }
2400 }
2401
2402 // if this is a load, check for anti-dependent stores
2403 // We use a conservative algorithm to identify potential interfering
2404 // instructions and for rescheduling the load. The users of the memory
2405 // input of this load are examined. Any use which is not a load and is
2406 // dominated by early is considered a potentially interfering store.
2407 // This can produce false positives.
2408 if (n->is_Load() && LCA != early) {
2409 Node_List worklist;
2410
2411 Node *mem = n->in(MemNode::Memory);
2412 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2413 Node* s = mem->fast_out(i);
2414 worklist.push(s);
2415 }
2416 while(worklist.size() != 0 && LCA != early) {
2417 Node* s = worklist.pop();
2418 if (s->is_Load()) {
2419 continue;
2420 } else if (s->is_MergeMem()) {
2421 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
2422 Node* s1 = s->fast_out(i);
2423 worklist.push(s1);
2424 }
2425 } else {
2426 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
2427 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
2428 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
2429 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
2430 }
2431 }
2432 }
2433 }
2434
2435 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
2436 return LCA;
2437 }
2438
2439 // true if CFG node d dominates CFG node n
2440 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
2441 if (d == n)
2442 return true;
2443 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
2444 uint dd = dom_depth(d);
2445 while (dom_depth(n) >= dd) {
2446 if (n == d)
2447 return true;
2448 n = idom(n);
2449 }
2450 return false;
2451 }
2452
2453 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
2454 // Pair-wise LCA with tags.
2455 // Tag each index with the node 'tag' currently being processed
2456 // before advancing up the dominator chain using idom().
2457 // Later calls that find a match to 'tag' know that this path has already
2458 // been considered in the current LCA (which is input 'n1' by convention).
2459 // Since get_late_ctrl() is only called once for each node, the tag array
2460 // does not need to be cleared between calls to get_late_ctrl().
2461 // Algorithm trades a larger constant factor for better asymptotic behavior
2462 //
2463 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
2464 uint d1 = dom_depth(n1);
2465 uint d2 = dom_depth(n2);
2466
2467 do {
2468 if (d1 > d2) {
2469 // current lca is deeper than n2
2470 _dom_lca_tags.map(n1->_idx, tag);
2471 n1 = idom(n1);
2472 d1 = dom_depth(n1);
2473 } else if (d1 < d2) {
2474 // n2 is deeper than current lca
2475 Node *memo = _dom_lca_tags[n2->_idx];
2476 if( memo == tag ) {
2477 return n1; // Return the current LCA
2478 }
2479 _dom_lca_tags.map(n2->_idx, tag);
2480 n2 = idom(n2);
2481 d2 = dom_depth(n2);
2482 } else {
2483 // Here d1 == d2. Due to edits of the dominator-tree, sections
2484 // of the tree might have the same depth. These sections have
2485 // to be searched more carefully.
2486
2487 // Scan up all the n1's with equal depth, looking for n2.
2488 _dom_lca_tags.map(n1->_idx, tag);
2489 Node *t1 = idom(n1);
2490 while (dom_depth(t1) == d1) {
2491 if (t1 == n2) return n2;
2492 _dom_lca_tags.map(t1->_idx, tag);
2493 t1 = idom(t1);
2494 }
2495 // Scan up all the n2's with equal depth, looking for n1.
2496 _dom_lca_tags.map(n2->_idx, tag);
2497 Node *t2 = idom(n2);
2498 while (dom_depth(t2) == d2) {
2499 if (t2 == n1) return n1;
2500 _dom_lca_tags.map(t2->_idx, tag);
2501 t2 = idom(t2);
2502 }
2503 // Move up to a new dominator-depth value as well as up the dom-tree.
2504 n1 = t1;
2505 n2 = t2;
2506 d1 = dom_depth(n1);
2507 d2 = dom_depth(n2);
2508 }
2509 } while (n1 != n2);
2510 return n1;
2511 }
2512
2513 //------------------------------init_dom_lca_tags------------------------------
2514 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2515 // Intended use does not involve any growth for the array, so it could
2516 // be of fixed size.
2517 void PhaseIdealLoop::init_dom_lca_tags() {
2518 uint limit = C->unique() + 1;
2519 _dom_lca_tags.map( limit, NULL );
2520 #ifdef ASSERT
2521 for( uint i = 0; i < limit; ++i ) {
2522 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2523 }
2524 #endif // ASSERT
2525 }
2526
2527 //------------------------------clear_dom_lca_tags------------------------------
2528 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2529 // Intended use does not involve any growth for the array, so it could
2530 // be of fixed size.
2531 void PhaseIdealLoop::clear_dom_lca_tags() {
2532 uint limit = C->unique() + 1;
2533 _dom_lca_tags.map( limit, NULL );
2534 _dom_lca_tags.clear();
2535 #ifdef ASSERT
2536 for( uint i = 0; i < limit; ++i ) {
2537 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2538 }
2539 #endif // ASSERT
2540 }
2541
2542 //------------------------------build_loop_late--------------------------------
2543 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2544 // Second pass finds latest legal placement, and ideal loop placement.
2545 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) {
2546 while (worklist.size() != 0) {
2547 Node *n = worklist.pop();
2548 // Only visit once
2549 if (visited.test_set(n->_idx)) continue;
2550 uint cnt = n->outcnt();
2551 uint i = 0;
2552 while (true) {
2553 assert( _nodes[n->_idx], "no dead nodes" );
2554 // Visit all children
2555 if (i < cnt) {
2556 Node* use = n->raw_out(i);
2557 ++i;
2558 // Check for dead uses. Aggressively prune such junk. It might be
2559 // dead in the global sense, but still have local uses so I cannot
2560 // easily call 'remove_dead_node'.
2561 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
2562 // Due to cycles, we might not hit the same fixed point in the verify
2563 // pass as we do in the regular pass. Instead, visit such phis as
2564 // simple uses of the loop head.
2565 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
2566 if( !visited.test(use->_idx) )
2567 worklist.push(use);
2568 } else if( !visited.test_set(use->_idx) ) {
2569 nstack.push(n, i); // Save parent and next use's index.
2570 n = use; // Process all children of current use.
2571 cnt = use->outcnt();
2572 i = 0;
2573 }
2574 } else {
2575 // Do not visit around the backedge of loops via data edges.
2576 // push dead code onto a worklist
2577 _deadlist.push(use);
2578 }
2579 } else {
2580 // All of n's children have been processed, complete post-processing.
2581 build_loop_late_post(n, verify_me);
2582 if (nstack.is_empty()) {
2583 // Finished all nodes on stack.
2584 // Process next node on the worklist.
2585 break;
2586 }
2587 // Get saved parent node and next use's index. Visit the rest of uses.
2588 n = nstack.node();
2589 cnt = n->outcnt();
2590 i = nstack.index();
2591 nstack.pop();
2592 }
2593 }
2594 }
2595 }
2596
2597 //------------------------------build_loop_late_post---------------------------
2598 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2599 // Second pass finds latest legal placement, and ideal loop placement.
2600 void PhaseIdealLoop::build_loop_late_post( Node *n, const PhaseIdealLoop *verify_me ) {
2601
2602 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress()) {
2603 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
2604 }
2605
2606 // CFG and pinned nodes already handled
2607 if( n->in(0) ) {
2608 if( n->in(0)->is_top() ) return; // Dead?
2609
2610 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
2611 // _must_ be pinned (they have to observe their control edge of course).
2612 // Unlike Stores (which modify an unallocable resource, the memory
2613 // state), Mods/Loads can float around. So free them up.
2614 bool pinned = true;
2615 switch( n->Opcode() ) {
2616 case Op_DivI:
2617 case Op_DivF:
2618 case Op_DivD:
2619 case Op_ModI:
2620 case Op_ModF:
2621 case Op_ModD:
2622 case Op_LoadB: // Same with Loads; they can sink
2623 case Op_LoadC: // during loop optimizations.
2624 case Op_LoadD:
2625 case Op_LoadF:
2626 case Op_LoadI:
2627 case Op_LoadKlass:
2628 case Op_LoadL:
2629 case Op_LoadS:
2630 case Op_LoadP:
2631 case Op_LoadRange:
2632 case Op_LoadD_unaligned:
2633 case Op_LoadL_unaligned:
2634 case Op_StrComp: // Does a bunch of load-like effects
2635 case Op_AryEq:
2636 pinned = false;
2637 }
2638 if( pinned ) {
2639 IdealLoopTree *choosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
2640 if( !choosen_loop->_child ) // Inner loop?
2641 choosen_loop->_body.push(n); // Collect inner loops
2642 return;
2643 }
2644 } else { // No slot zero
2645 if( n->is_CFG() ) { // CFG with no slot 0 is dead
2646 _nodes.map(n->_idx,0); // No block setting, it's globally dead
2647 return;
2648 }
2649 assert(!n->is_CFG() || n->outcnt() == 0, "");
2650 }
2651
2652 // Do I have a "safe range" I can select over?
2653 Node *early = get_ctrl(n);// Early location already computed
2654
2655 // Compute latest point this Node can go
2656 Node *LCA = get_late_ctrl( n, early );
2657 // LCA is NULL due to uses being dead
2658 if( LCA == NULL ) {
2659 #ifdef ASSERT
2660 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
2661 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
2662 }
2663 #endif
2664 _nodes.map(n->_idx, 0); // This node is useless
2665 _deadlist.push(n);
2666 return;
2667 }
2668 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
2669
2670 Node *legal = LCA; // Walk 'legal' up the IDOM chain
2671 Node *least = legal; // Best legal position so far
2672 while( early != legal ) { // While not at earliest legal
2673 // Find least loop nesting depth
2674 legal = idom(legal); // Bump up the IDOM tree
2675 // Check for lower nesting depth
2676 if( get_loop(legal)->_nest < get_loop(least)->_nest )
2677 least = legal;
2678 }
2679
2680 // Try not to place code on a loop entry projection
2681 // which can inhibit range check elimination.
2682 if (least != early) {
2683 Node* ctrl_out = least->unique_ctrl_out();
2684 if (ctrl_out && ctrl_out->is_CountedLoop() &&
2685 least == ctrl_out->in(LoopNode::EntryControl)) {
2686 Node* least_dom = idom(least);
2687 if (get_loop(least_dom)->is_member(get_loop(least))) {
2688 least = least_dom;
2689 }
2690 }
2691 }
2692
2693 #ifdef ASSERT
2694 // If verifying, verify that 'verify_me' has a legal location
2695 // and choose it as our location.
2696 if( verify_me ) {
2697 Node *v_ctrl = verify_me->get_ctrl_no_update(n);
2698 Node *legal = LCA;
2699 while( early != legal ) { // While not at earliest legal
2700 if( legal == v_ctrl ) break; // Check for prior good location
2701 legal = idom(legal) ;// Bump up the IDOM tree
2702 }
2703 // Check for prior good location
2704 if( legal == v_ctrl ) least = legal; // Keep prior if found
2705 }
2706 #endif
2707
2708 // Assign discovered "here or above" point
2709 least = find_non_split_ctrl(least);
2710 set_ctrl(n, least);
2711
2712 // Collect inner loop bodies
2713 IdealLoopTree *choosen_loop = get_loop(least);
2714 if( !choosen_loop->_child ) // Inner loop?
2715 choosen_loop->_body.push(n);// Collect inner loops
2716 }
2717
2718 #ifndef PRODUCT
2719 //------------------------------dump-------------------------------------------
2720 void PhaseIdealLoop::dump( ) const {
2721 ResourceMark rm;
2722 Arena* arena = Thread::current()->resource_area();
2723 Node_Stack stack(arena, C->unique() >> 2);
2724 Node_List rpo_list;
2725 VectorSet visited(arena);
2726 visited.set(C->top()->_idx);
2727 rpo( C->root(), stack, visited, rpo_list );
2728 // Dump root loop indexed by last element in PO order
2729 dump( _ltree_root, rpo_list.size(), rpo_list );
2730 }
2731
2732 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
2733
2734 // Indent by loop nesting depth
2735 for( uint x = 0; x < loop->_nest; x++ )
2736 tty->print(" ");
2737 tty->print_cr("---- Loop N%d-N%d ----", loop->_head->_idx,loop->_tail->_idx);
2738
2739 // Now scan for CFG nodes in the same loop
2740 for( uint j=idx; j > 0; j-- ) {
2741 Node *n = rpo_list[j-1];
2742 if( !_nodes[n->_idx] ) // Skip dead nodes
2743 continue;
2744 if( get_loop(n) != loop ) { // Wrong loop nest
2745 if( get_loop(n)->_head == n && // Found nested loop?
2746 get_loop(n)->_parent == loop )
2747 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
2748 continue;
2749 }
2750
2751 // Dump controlling node
2752 for( uint x = 0; x < loop->_nest; x++ )
2753 tty->print(" ");
2754 tty->print("C");
2755 if( n == C->root() ) {
2756 n->dump();
2757 } else {
2758 Node* cached_idom = idom_no_update(n);
2759 Node *computed_idom = n->in(0);
2760 if( n->is_Region() ) {
2761 computed_idom = compute_idom(n);
2762 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
2763 // any MultiBranch ctrl node), so apply a similar transform to
2764 // the cached idom returned from idom_no_update.
2765 cached_idom = find_non_split_ctrl(cached_idom);
2766 }
2767 tty->print(" ID:%d",computed_idom->_idx);
2768 n->dump();
2769 if( cached_idom != computed_idom ) {
2770 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
2771 computed_idom->_idx, cached_idom->_idx);
2772 }
2773 }
2774 // Dump nodes it controls
2775 for( uint k = 0; k < _nodes.Size(); k++ ) {
2776 // (k < C->unique() && get_ctrl(find(k)) == n)
2777 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
2778 Node *m = C->root()->find(k);
2779 if( m && m->outcnt() > 0 ) {
2780 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
2781 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
2782 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
2783 }
2784 for( uint j = 0; j < loop->_nest; j++ )
2785 tty->print(" ");
2786 tty->print(" ");
2787 m->dump();
2788 }
2789 }
2790 }
2791 }
2792 }
2793
2794 // Collect a R-P-O for the whole CFG.
2795 // Result list is in post-order (scan backwards for RPO)
2796 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
2797 stk.push(start, 0);
2798 visited.set(start->_idx);
2799
2800 while (stk.is_nonempty()) {
2801 Node* m = stk.node();
2802 uint idx = stk.index();
2803 if (idx < m->outcnt()) {
2804 stk.set_index(idx + 1);
2805 Node* n = m->raw_out(idx);
2806 if (n->is_CFG() && !visited.test_set(n->_idx)) {
2807 stk.push(n, 0);
2808 }
2809 } else {
2810 rpo_list.push(m);
2811 stk.pop();
2812 }
2813 }
2814 }
2815 #endif
2816
2817
2818 //=============================================================================
2819 //------------------------------LoopTreeIterator-----------------------------------
2820
2821 // Advance to next loop tree using a preorder, left-to-right traversal.
2822 void LoopTreeIterator::next() {
2823 assert(!done(), "must not be done.");
2824 if (_curnt->_child != NULL) {
2825 _curnt = _curnt->_child;
2826 } else if (_curnt->_next != NULL) {
2827 _curnt = _curnt->_next;
2828 } else {
2829 while (_curnt != _root && _curnt->_next == NULL) {
2830 _curnt = _curnt->_parent;
2831 }
2832 if (_curnt == _root) {
2833 _curnt = NULL;
2834 assert(done(), "must be done.");
2835 } else {
2836 assert(_curnt->_next != NULL, "must be more to do");
2837 _curnt = _curnt->_next;
2838 }
2839 }
2840 }