1 #ifdef USE_PRAGMA_IDENT_SRC
2 #pragma ident "@(#)loopnode.cpp 1.262 07/10/23 13:12:50 JVM"
3 #endif
4 /*
5 * Copyright 1998-2007 Sun Microsystems, Inc. All Rights Reserved.
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7 *
8 * This code is free software; you can redistribute it and/or modify it
9 * under the terms of the GNU General Public License version 2 only, as
10 * published by the Free Software Foundation.
11 *
12 * This code is distributed in the hope that it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 * version 2 for more details (a copy is included in the LICENSE file that
16 * accompanied this code).
17 *
18 * You should have received a copy of the GNU General Public License version
19 * 2 along with this work; if not, write to the Free Software Foundation,
20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
21 *
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
24 * have any questions.
25 *
26 */
27
28 #include "incls/_precompiled.incl"
29 #include "incls/_loopnode.cpp.incl"
30
31 //=============================================================================
32 //------------------------------is_loop_iv-------------------------------------
33 // Determine if a node is Counted loop induction variable.
34 // The method is declared in node.hpp.
35 const Node* Node::is_loop_iv() const {
36 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
37 this->as_Phi()->region()->is_CountedLoop() &&
38 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
39 return this;
40 } else {
41 return NULL;
42 }
43 }
44
45 //=============================================================================
46 //------------------------------dump_spec--------------------------------------
47 // Dump special per-node info
48 #ifndef PRODUCT
49 void LoopNode::dump_spec(outputStream *st) const {
50 if( is_inner_loop () ) st->print( "inner " );
51 if( is_partial_peel_loop () ) st->print( "partial_peel " );
52 if( partial_peel_has_failed () ) st->print( "partial_peel_failed " );
53 }
54 #endif
55
56 //------------------------------get_early_ctrl---------------------------------
57 // Compute earliest legal control
58 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
59 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
60 uint i;
61 Node *early;
62 if( n->in(0) ) {
63 early = n->in(0);
64 if( !early->is_CFG() ) // Might be a non-CFG multi-def
65 early = get_ctrl(early); // So treat input as a straight data input
66 i = 1;
67 } else {
68 early = get_ctrl(n->in(1));
69 i = 2;
70 }
71 uint e_d = dom_depth(early);
72 assert( early, "" );
73 for( ; i < n->req(); i++ ) {
74 Node *cin = get_ctrl(n->in(i));
75 assert( cin, "" );
76 // Keep deepest dominator depth
77 uint c_d = dom_depth(cin);
78 if( c_d > e_d ) { // Deeper guy?
79 early = cin; // Keep deepest found so far
80 e_d = c_d;
81 } else if( c_d == e_d && // Same depth?
82 early != cin ) { // If not equal, must use slower algorithm
83 // If same depth but not equal, one _must_ dominate the other
84 // and we want the deeper (i.e., dominated) guy.
85 Node *n1 = early;
86 Node *n2 = cin;
87 while( 1 ) {
88 n1 = idom(n1); // Walk up until break cycle
89 n2 = idom(n2);
90 if( n1 == cin || // Walked early up to cin
91 dom_depth(n2) < c_d )
92 break; // early is deeper; keep him
93 if( n2 == early || // Walked cin up to early
94 dom_depth(n1) < c_d ) {
95 early = cin; // cin is deeper; keep him
96 break;
97 }
98 }
99 e_d = dom_depth(early); // Reset depth register cache
100 }
101 }
102
103 // Return earliest legal location
104 assert(early == find_non_split_ctrl(early), "unexpected early control");
105
106 return early;
107 }
108
109 //------------------------------set_early_ctrl---------------------------------
110 // Set earliest legal control
111 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
112 Node *early = get_early_ctrl(n);
113
114 // Record earliest legal location
115 set_ctrl(n, early);
116 }
117
118 //------------------------------set_subtree_ctrl-------------------------------
119 // set missing _ctrl entries on new nodes
120 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
121 // Already set? Get out.
122 if( _nodes[n->_idx] ) return;
123 // Recursively set _nodes array to indicate where the Node goes
124 uint i;
125 for( i = 0; i < n->req(); ++i ) {
126 Node *m = n->in(i);
127 if( m && m != C->root() )
128 set_subtree_ctrl( m );
129 }
130
131 // Fixup self
132 set_early_ctrl( n );
133 }
134
135 //------------------------------is_counted_loop--------------------------------
136 Node *PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
137 PhaseGVN *gvn = &_igvn;
138
139 // Counted loop head must be a good RegionNode with only 3 not NULL
140 // control input edges: Self, Entry, LoopBack.
141 if ( x->in(LoopNode::Self) == NULL || x->req() != 3 )
142 return NULL;
143
144 Node *init_control = x->in(LoopNode::EntryControl);
145 Node *back_control = x->in(LoopNode::LoopBackControl);
146 if( init_control == NULL || back_control == NULL ) // Partially dead
147 return NULL;
148 // Must also check for TOP when looking for a dead loop
149 if( init_control->is_top() || back_control->is_top() )
150 return NULL;
151
152 // Allow funny placement of Safepoint
153 if( back_control->Opcode() == Op_SafePoint )
154 back_control = back_control->in(TypeFunc::Control);
155
156 // Controlling test for loop
157 Node *iftrue = back_control;
158 uint iftrue_op = iftrue->Opcode();
159 if( iftrue_op != Op_IfTrue &&
160 iftrue_op != Op_IfFalse )
161 // I have a weird back-control. Probably the loop-exit test is in
162 // the middle of the loop and I am looking at some trailing control-flow
163 // merge point. To fix this I would have to partially peel the loop.
164 return NULL; // Obscure back-control
165
166 // Get boolean guarding loop-back test
167 Node *iff = iftrue->in(0);
168 if( get_loop(iff) != loop || !iff->in(1)->is_Bool() ) return NULL;
169 BoolNode *test = iff->in(1)->as_Bool();
170 BoolTest::mask bt = test->_test._test;
171 float cl_prob = iff->as_If()->_prob;
172 if( iftrue_op == Op_IfFalse ) {
173 bt = BoolTest(bt).negate();
174 cl_prob = 1.0 - cl_prob;
175 }
176 // Get backedge compare
177 Node *cmp = test->in(1);
178 int cmp_op = cmp->Opcode();
179 if( cmp_op != Op_CmpI )
180 return NULL; // Avoid pointer & float compares
181
182 // Find the trip-counter increment & limit. Limit must be loop invariant.
183 Node *incr = cmp->in(1);
184 Node *limit = cmp->in(2);
185
186 // ---------
187 // need 'loop()' test to tell if limit is loop invariant
188 // ---------
189
190 if( !is_member( loop, get_ctrl(incr) ) ) { // Swapped trip counter and limit?
191 Node *tmp = incr; // Then reverse order into the CmpI
192 incr = limit;
193 limit = tmp;
194 bt = BoolTest(bt).commute(); // And commute the exit test
195 }
196 if( is_member( loop, get_ctrl(limit) ) ) // Limit must loop-invariant
197 return NULL;
198
199 // Trip-counter increment must be commutative & associative.
200 uint incr_op = incr->Opcode();
201 if( incr_op == Op_Phi && incr->req() == 3 ) {
202 incr = incr->in(2); // Assume incr is on backedge of Phi
203 incr_op = incr->Opcode();
204 }
205 Node* trunc1 = NULL;
206 Node* trunc2 = NULL;
207 const TypeInt* iv_trunc_t = NULL;
208 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
209 return NULL; // Funny increment opcode
210 }
211
212 // Get merge point
213 Node *xphi = incr->in(1);
214 Node *stride = incr->in(2);
215 if( !stride->is_Con() ) { // Oops, swap these
216 if( !xphi->is_Con() ) // Is the other guy a constant?
217 return NULL; // Nope, unknown stride, bail out
218 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
219 xphi = stride;
220 stride = tmp;
221 }
222 //if( loop(xphi) != l) return NULL;// Merge point is in inner loop??
223 if( !xphi->is_Phi() ) return NULL; // Too much math on the trip counter
224 PhiNode *phi = xphi->as_Phi();
225
226 // Stride must be constant
227 const Type *stride_t = stride->bottom_type();
228 int stride_con = stride_t->is_int()->get_con();
229 assert( stride_con, "missed some peephole opt" );
230
231 // Phi must be of loop header; backedge must wrap to increment
232 if( phi->region() != x ) return NULL;
233 if( trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
234 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1 ) {
235 return NULL;
236 }
237 Node *init_trip = phi->in(LoopNode::EntryControl);
238 //if (!init_trip->is_Con()) return NULL; // avoid rolling over MAXINT/MININT
239
240 // If iv trunc type is smaller than int, check for possible wrap.
241 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
242 assert(trunc1 != NULL, "must have found some truncation");
243
244 // Get a better type for the phi (filtered thru if's)
245 const TypeInt* phi_ft = filtered_type(phi);
246
247 // Can iv take on a value that will wrap?
248 //
249 // Ensure iv's limit is not within "stride" of the wrap value.
250 //
251 // Example for "short" type
252 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
253 // If the stride is +10, then the last value of the induction
254 // variable before the increment (phi_ft->_hi) must be
255 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
256 // ensure no truncation occurs after the increment.
257
258 if (stride_con > 0) {
259 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
260 iv_trunc_t->_lo > phi_ft->_lo) {
261 return NULL; // truncation may occur
262 }
263 } else if (stride_con < 0) {
264 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
265 iv_trunc_t->_hi < phi_ft->_hi) {
266 return NULL; // truncation may occur
267 }
268 }
269 // No possibility of wrap so truncation can be discarded
270 // Promote iv type to Int
271 } else {
272 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
273 }
274
275 // =================================================
276 // ---- SUCCESS! Found A Trip-Counted Loop! -----
277 //
278 // Canonicalize the condition on the test. If we can exactly determine
279 // the trip-counter exit value, then set limit to that value and use
280 // a '!=' test. Otherwise use conditon '<' for count-up loops and
281 // '>' for count-down loops. If the condition is inverted and we will
282 // be rolling through MININT to MAXINT, then bail out.
283
284 C->print_method("Before CountedLoop", 3);
285
286 // Check for SafePoint on backedge and remove
287 Node *sfpt = x->in(LoopNode::LoopBackControl);
288 if( sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
289 lazy_replace( sfpt, iftrue );
290 loop->_tail = iftrue;
291 }
292
293
294 // If compare points to incr, we are ok. Otherwise the compare
295 // can directly point to the phi; in this case adjust the compare so that
296 // it points to the incr by adusting the limit.
297 if( cmp->in(1) == phi || cmp->in(2) == phi )
298 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
299
300 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
301 // Final value for iterator should be: trip_count * stride + init_trip.
302 const Type *limit_t = limit->bottom_type();
303 const Type *init_t = init_trip->bottom_type();
304 Node *one_p = gvn->intcon( 1);
305 Node *one_m = gvn->intcon(-1);
306
307 Node *trip_count = NULL;
308 Node *hook = new (C, 6) Node(6);
309 switch( bt ) {
310 case BoolTest::eq:
311 return NULL; // Bail out, but this loop trips at most twice!
312 case BoolTest::ne: // Ahh, the case we desire
313 if( stride_con == 1 )
314 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
315 else if( stride_con == -1 )
316 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
317 else
318 return NULL; // Odd stride; must prove we hit limit exactly
319 set_subtree_ctrl( trip_count );
320 //_loop.map(trip_count->_idx,loop(limit));
321 break;
322 case BoolTest::le: // Maybe convert to '<' case
323 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
324 set_subtree_ctrl( limit );
325 hook->init_req(4, limit);
326
327 bt = BoolTest::lt;
328 // Make the new limit be in the same loop nest as the old limit
329 //_loop.map(limit->_idx,limit_loop);
330 // Fall into next case
331 case BoolTest::lt: { // Maybe convert to '!=' case
332 if( stride_con < 0 ) return NULL; // Count down loop rolls through MAXINT
333 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
334 set_subtree_ctrl( range );
335 hook->init_req(0, range);
336
337 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
338 set_subtree_ctrl( bias );
339 hook->init_req(1, bias);
340
341 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
342 set_subtree_ctrl( bias1 );
343 hook->init_req(2, bias1);
344
345 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
346 set_subtree_ctrl( trip_count );
347 hook->init_req(3, trip_count);
348 break;
349 }
350
351 case BoolTest::ge: // Maybe convert to '>' case
352 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
353 set_subtree_ctrl( limit );
354 hook->init_req(4 ,limit);
355
356 bt = BoolTest::gt;
357 // Make the new limit be in the same loop nest as the old limit
358 //_loop.map(limit->_idx,limit_loop);
359 // Fall into next case
360 case BoolTest::gt: { // Maybe convert to '!=' case
361 if( stride_con > 0 ) return NULL; // count up loop rolls through MININT
362 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
363 set_subtree_ctrl( range );
364 hook->init_req(0, range);
365
366 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
367 set_subtree_ctrl( bias );
368 hook->init_req(1, bias);
369
370 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
371 set_subtree_ctrl( bias1 );
372 hook->init_req(2, bias1);
373
374 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
375 set_subtree_ctrl( trip_count );
376 hook->init_req(3, trip_count);
377 break;
378 }
379 }
380
381 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
382 set_subtree_ctrl( span );
383 hook->init_req(5, span);
384
385 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
386 set_subtree_ctrl( limit );
387
388 // Build a canonical trip test.
389 // Clone code, as old values may be in use.
390 incr = incr->clone();
391 incr->set_req(1,phi);
392 incr->set_req(2,stride);
393 incr = _igvn.register_new_node_with_optimizer(incr);
394 set_early_ctrl( incr );
395 _igvn.hash_delete(phi);
396 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
397
398 // If phi type is more restrictive than Int, raise to
399 // Int to prevent (almost) infinite recursion in igvn
400 // which can only handle integer types for constants or minint..maxint.
401 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
402 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
403 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
404 nphi = _igvn.register_new_node_with_optimizer(nphi);
405 set_ctrl(nphi, get_ctrl(phi));
406 _igvn.subsume_node(phi, nphi);
407 phi = nphi->as_Phi();
408 }
409 cmp = cmp->clone();
410 cmp->set_req(1,incr);
411 cmp->set_req(2,limit);
412 cmp = _igvn.register_new_node_with_optimizer(cmp);
413 set_ctrl(cmp, iff->in(0));
414
415 Node *tmp = test->clone();
416 assert( tmp->is_Bool(), "" );
417 test = (BoolNode*)tmp;
418 (*(BoolTest*)&test->_test)._test = bt; //BoolTest::ne;
419 test->set_req(1,cmp);
420 _igvn.register_new_node_with_optimizer(test);
421 set_ctrl(test, iff->in(0));
422 // If the exit test is dead, STOP!
423 if( test == NULL ) return NULL;
424 _igvn.hash_delete(iff);
425 iff->set_req_X( 1, test, &_igvn );
426
427 // Replace the old IfNode with a new LoopEndNode
428 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), iff->in(1), cl_prob, iff->as_If()->_fcnt ));
429 IfNode *le = lex->as_If();
430 uint dd = dom_depth(iff);
431 set_idom(le, le->in(0), dd); // Update dominance for loop exit
432 set_loop(le, loop);
433
434 // Get the loop-exit control
435 Node *if_f = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
436
437 // Need to swap loop-exit and loop-back control?
438 if( iftrue_op == Op_IfFalse ) {
439 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
440 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
441
442 loop->_tail = back_control = ift2;
443 set_loop(ift2, loop);
444 set_loop(iff2, get_loop(if_f));
445
446 // Lazy update of 'get_ctrl' mechanism.
447 lazy_replace_proj( if_f , iff2 );
448 lazy_replace_proj( iftrue, ift2 );
449
450 // Swap names
451 if_f = iff2;
452 iftrue = ift2;
453 } else {
454 _igvn.hash_delete(if_f );
455 _igvn.hash_delete(iftrue);
456 if_f ->set_req_X( 0, le, &_igvn );
457 iftrue->set_req_X( 0, le, &_igvn );
458 }
459
460 set_idom(iftrue, le, dd+1);
461 set_idom(if_f, le, dd+1);
462
463 // Now setup a new CountedLoopNode to replace the existing LoopNode
464 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
465 // The following assert is approximately true, and defines the intention
466 // of can_be_counted_loop. It fails, however, because phase->type
467 // is not yet initialized for this loop and its parts.
468 //assert(l->can_be_counted_loop(this), "sanity");
469 _igvn.register_new_node_with_optimizer(l);
470 set_loop(l, loop);
471 loop->_head = l;
472 // Fix all data nodes placed at the old loop head.
473 // Uses the lazy-update mechanism of 'get_ctrl'.
474 lazy_replace( x, l );
475 set_idom(l, init_control, dom_depth(x));
476
477 // Check for immediately preceeding SafePoint and remove
478 Node *sfpt2 = le->in(0);
479 if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
480 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
481
482 // Free up intermediate goo
483 _igvn.remove_dead_node(hook);
484
485 C->print_method("After CountedLoop", 3);
486
487 // Return trip counter
488 return trip_count;
489 }
490
491
492 //------------------------------Ideal------------------------------------------
493 // Return a node which is more "ideal" than the current node.
494 // Attempt to convert into a counted-loop.
495 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
496 if (!can_be_counted_loop(phase)) {
497 phase->C->set_major_progress();
498 }
499 return RegionNode::Ideal(phase, can_reshape);
500 }
501
502
503 //=============================================================================
504 //------------------------------Ideal------------------------------------------
505 // Return a node which is more "ideal" than the current node.
506 // Attempt to convert into a counted-loop.
507 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
508 return RegionNode::Ideal(phase, can_reshape);
509 }
510
511 //------------------------------dump_spec--------------------------------------
512 // Dump special per-node info
513 #ifndef PRODUCT
514 void CountedLoopNode::dump_spec(outputStream *st) const {
515 LoopNode::dump_spec(st);
516 if( stride_is_con() ) {
517 st->print("stride: %d ",stride_con());
518 } else {
519 st->print("stride: not constant ");
520 }
521 if( is_pre_loop () ) st->print("pre of N%d" , _main_idx );
522 if( is_main_loop() ) st->print("main of N%d", _idx );
523 if( is_post_loop() ) st->print("post of N%d", _main_idx );
524 }
525 #endif
526
527 //=============================================================================
528 int CountedLoopEndNode::stride_con() const {
529 return stride()->bottom_type()->is_int()->get_con();
530 }
531
532
533 //----------------------match_incr_with_optional_truncation--------------------
534 // Match increment with optional truncation:
535 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
536 // Return NULL for failure. Success returns the increment node.
537 Node* CountedLoopNode::match_incr_with_optional_truncation(
538 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
539 // Quick cutouts:
540 if (expr == NULL || expr->req() != 3) return false;
541
542 Node *t1 = NULL;
543 Node *t2 = NULL;
544 const TypeInt* trunc_t = TypeInt::INT;
545 Node* n1 = expr;
546 int n1op = n1->Opcode();
547
548 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
549 if (n1op == Op_AndI &&
550 n1->in(2)->is_Con() &&
551 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
552 // %%% This check should match any mask of 2**K-1.
553 t1 = n1;
554 n1 = t1->in(1);
555 n1op = n1->Opcode();
556 trunc_t = TypeInt::CHAR;
557 } else if (n1op == Op_RShiftI &&
558 n1->in(1) != NULL &&
559 n1->in(1)->Opcode() == Op_LShiftI &&
560 n1->in(2) == n1->in(1)->in(2) &&
561 n1->in(2)->is_Con()) {
562 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
563 // %%% This check should match any shift in [1..31].
564 if (shift == 16 || shift == 8) {
565 t1 = n1;
566 t2 = t1->in(1);
567 n1 = t2->in(1);
568 n1op = n1->Opcode();
569 if (shift == 16) {
570 trunc_t = TypeInt::SHORT;
571 } else if (shift == 8) {
572 trunc_t = TypeInt::BYTE;
573 }
574 }
575 }
576
577 // If (maybe after stripping) it is an AddI, we won:
578 if (n1op == Op_AddI) {
579 *trunc1 = t1;
580 *trunc2 = t2;
581 *trunc_type = trunc_t;
582 return n1;
583 }
584
585 // failed
586 return NULL;
587 }
588
589
590 //------------------------------filtered_type--------------------------------
591 // Return a type based on condition control flow
592 // A successful return will be a type that is restricted due
593 // to a series of dominating if-tests, such as:
594 // if (i < 10) {
595 // if (i > 0) {
596 // here: "i" type is [1..10)
597 // }
598 // }
599 // or a control flow merge
600 // if (i < 10) {
601 // do {
602 // phi( , ) -- at top of loop type is [min_int..10)
603 // i = ?
604 // } while ( i < 10)
605 //
606 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
607 assert(n && n->bottom_type()->is_int(), "must be int");
608 const TypeInt* filtered_t = NULL;
609 if (!n->is_Phi()) {
610 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
611 filtered_t = filtered_type_from_dominators(n, n_ctrl);
612
613 } else {
614 Node* phi = n->as_Phi();
615 Node* region = phi->in(0);
616 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
617 if (region && region != C->top()) {
618 for (uint i = 1; i < phi->req(); i++) {
619 Node* val = phi->in(i);
620 Node* use_c = region->in(i);
621 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
622 if (val_t != NULL) {
623 if (filtered_t == NULL) {
624 filtered_t = val_t;
625 } else {
626 filtered_t = filtered_t->meet(val_t)->is_int();
627 }
628 }
629 }
630 }
631 }
632 const TypeInt* n_t = _igvn.type(n)->is_int();
633 if (filtered_t != NULL) {
634 n_t = n_t->join(filtered_t)->is_int();
635 }
636 return n_t;
637 }
638
639
640 //------------------------------filtered_type_from_dominators--------------------------------
641 // Return a possibly more restrictive type for val based on condition control flow of dominators
642 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
643 if (val->is_Con()) {
644 return val->bottom_type()->is_int();
645 }
646 uint if_limit = 10; // Max number of dominating if's visited
647 const TypeInt* rtn_t = NULL;
648
649 if (use_ctrl && use_ctrl != C->top()) {
650 Node* val_ctrl = get_ctrl(val);
651 uint val_dom_depth = dom_depth(val_ctrl);
652 Node* pred = use_ctrl;
653 uint if_cnt = 0;
654 while (if_cnt < if_limit) {
655 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
656 if_cnt++;
657 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
658 if (if_t != NULL) {
659 if (rtn_t == NULL) {
660 rtn_t = if_t;
661 } else {
662 rtn_t = rtn_t->join(if_t)->is_int();
663 }
664 }
665 }
666 pred = idom(pred);
667 if (pred == NULL || pred == C->top()) {
668 break;
669 }
670 // Stop if going beyond definition block of val
671 if (dom_depth(pred) < val_dom_depth) {
672 break;
673 }
674 }
675 }
676 return rtn_t;
677 }
678
679
680 //------------------------------dump_spec--------------------------------------
681 // Dump special per-node info
682 #ifndef PRODUCT
683 void CountedLoopEndNode::dump_spec(outputStream *st) const {
684 if( in(TestValue)->is_Bool() ) {
685 BoolTest bt( test_trip()); // Added this for g++.
686
687 st->print("[");
688 bt.dump_on(st);
689 st->print("]");
690 }
691 st->print(" ");
692 IfNode::dump_spec(st);
693 }
694 #endif
695
696 //=============================================================================
697 //------------------------------is_member--------------------------------------
698 // Is 'l' a member of 'this'?
699 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
700 while( l->_nest > _nest ) l = l->_parent;
701 return l == this;
702 }
703
704 //------------------------------set_nest---------------------------------------
705 // Set loop tree nesting depth. Accumulate _has_call bits.
706 int IdealLoopTree::set_nest( uint depth ) {
707 _nest = depth;
708 int bits = _has_call;
709 if( _child ) bits |= _child->set_nest(depth+1);
710 if( bits ) _has_call = 1;
711 if( _next ) bits |= _next ->set_nest(depth );
712 return bits;
713 }
714
715 //------------------------------split_fall_in----------------------------------
716 // Split out multiple fall-in edges from the loop header. Move them to a
717 // private RegionNode before the loop. This becomes the loop landing pad.
718 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
719 PhaseIterGVN &igvn = phase->_igvn;
720 uint i;
721
722 // Make a new RegionNode to be the landing pad.
723 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
724 phase->set_loop(landing_pad,_parent);
725 // Gather all the fall-in control paths into the landing pad
726 uint icnt = fall_in_cnt;
727 uint oreq = _head->req();
728 for( i = oreq-1; i>0; i-- )
729 if( !phase->is_member( this, _head->in(i) ) )
730 landing_pad->set_req(icnt--,_head->in(i));
731
732 // Peel off PhiNode edges as well
733 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
734 Node *oj = _head->fast_out(j);
735 if( oj->is_Phi() ) {
736 PhiNode* old_phi = oj->as_Phi();
737 assert( old_phi->region() == _head, "" );
738 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
739 Node *p = PhiNode::make_blank(landing_pad, old_phi);
740 uint icnt = fall_in_cnt;
741 for( i = oreq-1; i>0; i-- ) {
742 if( !phase->is_member( this, _head->in(i) ) ) {
743 p->init_req(icnt--, old_phi->in(i));
744 // Go ahead and clean out old edges from old phi
745 old_phi->del_req(i);
746 }
747 }
748 // Search for CSE's here, because ZKM.jar does a lot of
749 // loop hackery and we need to be a little incremental
750 // with the CSE to avoid O(N^2) node blow-up.
751 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
752 if( p2 ) { // Found CSE
753 p->destruct(); // Recover useless new node
754 p = p2; // Use old node
755 } else {
756 igvn.register_new_node_with_optimizer(p, old_phi);
757 }
758 // Make old Phi refer to new Phi.
759 old_phi->add_req(p);
760 // Check for the special case of making the old phi useless and
761 // disappear it. In JavaGrande I have a case where this useless
762 // Phi is the loop limit and prevents recognizing a CountedLoop
763 // which in turn prevents removing an empty loop.
764 Node *id_old_phi = old_phi->Identity( &igvn );
765 if( id_old_phi != old_phi ) { // Found a simple identity?
766 // Note that I cannot call 'subsume_node' here, because
767 // that will yank the edge from old_phi to the Region and
768 // I'm mid-iteration over the Region's uses.
769 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
770 Node* use = old_phi->last_out(i);
771 igvn.hash_delete(use);
772 igvn._worklist.push(use);
773 uint uses_found = 0;
774 for (uint j = 0; j < use->len(); j++) {
775 if (use->in(j) == old_phi) {
776 if (j < use->req()) use->set_req (j, id_old_phi);
777 else use->set_prec(j, id_old_phi);
778 uses_found++;
779 }
780 }
781 i -= uses_found; // we deleted 1 or more copies of this edge
782 }
783 }
784 igvn._worklist.push(old_phi);
785 }
786 }
787 // Finally clean out the fall-in edges from the RegionNode
788 for( i = oreq-1; i>0; i-- ) {
789 if( !phase->is_member( this, _head->in(i) ) ) {
790 _head->del_req(i);
791 }
792 }
793 // Transform landing pad
794 igvn.register_new_node_with_optimizer(landing_pad, _head);
795 // Insert landing pad into the header
796 _head->add_req(landing_pad);
797 }
798
799 //------------------------------split_outer_loop-------------------------------
800 // Split out the outermost loop from this shared header.
801 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
802 PhaseIterGVN &igvn = phase->_igvn;
803
804 // Find index of outermost loop; it should also be my tail.
805 uint outer_idx = 1;
806 while( _head->in(outer_idx) != _tail ) outer_idx++;
807
808 // Make a LoopNode for the outermost loop.
809 Node *ctl = _head->in(LoopNode::EntryControl);
810 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
811 outer = igvn.register_new_node_with_optimizer(outer, _head);
812 phase->set_created_loop_node();
813 // Outermost loop falls into '_head' loop
814 _head->set_req(LoopNode::EntryControl, outer);
815 _head->del_req(outer_idx);
816 // Split all the Phis up between '_head' loop and 'outer' loop.
817 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
818 Node *out = _head->fast_out(j);
819 if( out->is_Phi() ) {
820 PhiNode *old_phi = out->as_Phi();
821 assert( old_phi->region() == _head, "" );
822 Node *phi = PhiNode::make_blank(outer, old_phi);
823 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
824 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
825 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
826 // Make old Phi point to new Phi on the fall-in path
827 igvn.hash_delete(old_phi);
828 old_phi->set_req(LoopNode::EntryControl, phi);
829 old_phi->del_req(outer_idx);
830 igvn._worklist.push(old_phi);
831 }
832 }
833
834 // Use the new loop head instead of the old shared one
835 _head = outer;
836 phase->set_loop(_head, this);
837 }
838
839 //------------------------------fix_parent-------------------------------------
840 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
841 loop->_parent = parent;
842 if( loop->_child ) fix_parent( loop->_child, loop );
843 if( loop->_next ) fix_parent( loop->_next , parent );
844 }
845
846 //------------------------------estimate_path_freq-----------------------------
847 static float estimate_path_freq( Node *n ) {
848 // Try to extract some path frequency info
849 IfNode *iff;
850 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
851 uint nop = n->Opcode();
852 if( nop == Op_SafePoint ) { // Skip any safepoint
853 n = n->in(0);
854 continue;
855 }
856 if( nop == Op_CatchProj ) { // Get count from a prior call
857 // Assume call does not always throw exceptions: means the call-site
858 // count is also the frequency of the fall-through path.
859 assert( n->is_CatchProj(), "" );
860 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
861 return 0.0f; // Assume call exception path is rare
862 Node *call = n->in(0)->in(0)->in(0);
863 assert( call->is_Call(), "expect a call here" );
864 const JVMState *jvms = ((CallNode*)call)->jvms();
865 ciMethodData* methodData = jvms->method()->method_data();
866 if (!methodData->is_mature()) return 0.0f; // No call-site data
867 ciProfileData* data = methodData->bci_to_data(jvms->bci());
868 if ((data == NULL) || !data->is_CounterData()) {
869 // no call profile available, try call's control input
870 n = n->in(0);
871 continue;
872 }
873 return data->as_CounterData()->count()/FreqCountInvocations;
874 }
875 // See if there's a gating IF test
876 Node *n_c = n->in(0);
877 if( !n_c->is_If() ) break; // No estimate available
878 iff = n_c->as_If();
879 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
880 // Compute how much count comes on this path
881 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
882 // Have no count info. Skip dull uncommon-trap like branches.
883 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
884 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
885 break;
886 // Skip through never-taken branch; look for a real loop exit.
887 n = iff->in(0);
888 }
889 return 0.0f; // No estimate available
890 }
891
892 //------------------------------merge_many_backedges---------------------------
893 // Merge all the backedges from the shared header into a private Region.
894 // Feed that region as the one backedge to this loop.
895 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
896 uint i;
897
898 // Scan for the top 2 hottest backedges
899 float hotcnt = 0.0f;
900 float warmcnt = 0.0f;
901 uint hot_idx = 0;
902 // Loop starts at 2 because slot 1 is the fall-in path
903 for( i = 2; i < _head->req(); i++ ) {
904 float cnt = estimate_path_freq(_head->in(i));
905 if( cnt > hotcnt ) { // Grab hottest path
906 warmcnt = hotcnt;
907 hotcnt = cnt;
908 hot_idx = i;
909 } else if( cnt > warmcnt ) { // And 2nd hottest path
910 warmcnt = cnt;
911 }
912 }
913
914 // See if the hottest backedge is worthy of being an inner loop
915 // by being much hotter than the next hottest backedge.
916 if( hotcnt <= 0.0001 ||
917 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
918
919 // Peel out the backedges into a private merge point; peel
920 // them all except optionally hot_idx.
921 PhaseIterGVN &igvn = phase->_igvn;
922
923 Node *hot_tail = NULL;
924 // Make a Region for the merge point
925 Node *r = new (phase->C, 1) RegionNode(1);
926 for( i = 2; i < _head->req(); i++ ) {
927 if( i != hot_idx )
928 r->add_req( _head->in(i) );
929 else hot_tail = _head->in(i);
930 }
931 igvn.register_new_node_with_optimizer(r, _head);
932 // Plug region into end of loop _head, followed by hot_tail
933 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
934 _head->set_req(2, r);
935 if( hot_idx ) _head->add_req(hot_tail);
936
937 // Split all the Phis up between '_head' loop and the Region 'r'
938 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
939 Node *out = _head->fast_out(j);
940 if( out->is_Phi() ) {
941 PhiNode* n = out->as_Phi();
942 igvn.hash_delete(n); // Delete from hash before hacking edges
943 Node *hot_phi = NULL;
944 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
945 // Check all inputs for the ones to peel out
946 uint j = 1;
947 for( uint i = 2; i < n->req(); i++ ) {
948 if( i != hot_idx )
949 phi->set_req( j++, n->in(i) );
950 else hot_phi = n->in(i);
951 }
952 // Register the phi but do not transform until whole place transforms
953 igvn.register_new_node_with_optimizer(phi, n);
954 // Add the merge phi to the old Phi
955 while( n->req() > 3 ) n->del_req( n->req()-1 );
956 n->set_req(2, phi);
957 if( hot_idx ) n->add_req(hot_phi);
958 }
959 }
960
961
962 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
963 // of self loop tree. Turn self into a loop headed by _head and with
964 // tail being the new merge point.
965 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
966 phase->set_loop(_tail,ilt); // Adjust tail
967 _tail = r; // Self's tail is new merge point
968 phase->set_loop(r,this);
969 ilt->_child = _child; // New guy has my children
970 _child = ilt; // Self has new guy as only child
971 ilt->_parent = this; // new guy has self for parent
972 ilt->_nest = _nest; // Same nesting depth (for now)
973
974 // Starting with 'ilt', look for child loop trees using the same shared
975 // header. Flatten these out; they will no longer be loops in the end.
976 IdealLoopTree **pilt = &_child;
977 while( ilt ) {
978 if( ilt->_head == _head ) {
979 uint i;
980 for( i = 2; i < _head->req(); i++ )
981 if( _head->in(i) == ilt->_tail )
982 break; // Still a loop
983 if( i == _head->req() ) { // No longer a loop
984 // Flatten ilt. Hang ilt's "_next" list from the end of
985 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
986 IdealLoopTree **cp = &ilt->_child;
987 while( *cp ) cp = &(*cp)->_next; // Find end of child list
988 *cp = ilt->_next; // Hang next list at end of child list
989 *pilt = ilt->_child; // Move child up to replace ilt
990 ilt->_head = NULL; // Flag as a loop UNIONED into parent
991 ilt = ilt->_child; // Repeat using new ilt
992 continue; // do not advance over ilt->_child
993 }
994 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
995 phase->set_loop(_head,ilt);
996 }
997 pilt = &ilt->_child; // Advance to next
998 ilt = *pilt;
999 }
1000
1001 if( _child ) fix_parent( _child, this );
1002 }
1003
1004 //------------------------------beautify_loops---------------------------------
1005 // Split shared headers and insert loop landing pads.
1006 // Insert a LoopNode to replace the RegionNode.
1007 // Return TRUE if loop tree is structurally changed.
1008 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1009 bool result = false;
1010 // Cache parts in locals for easy
1011 PhaseIterGVN &igvn = phase->_igvn;
1012
1013 phase->C->print_method("Before beautify loops", 3);
1014
1015 igvn.hash_delete(_head); // Yank from hash before hacking edges
1016
1017 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1018 int fall_in_cnt = 0;
1019 for( uint i = 1; i < _head->req(); i++ )
1020 if( !phase->is_member( this, _head->in(i) ) )
1021 fall_in_cnt++;
1022 assert( fall_in_cnt, "at least 1 fall-in path" );
1023 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1024 split_fall_in( phase, fall_in_cnt );
1025
1026 // Swap inputs to the _head and all Phis to move the fall-in edge to
1027 // the left.
1028 fall_in_cnt = 1;
1029 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1030 fall_in_cnt++;
1031 if( fall_in_cnt > 1 ) {
1032 // Since I am just swapping inputs I do not need to update def-use info
1033 Node *tmp = _head->in(1);
1034 _head->set_req( 1, _head->in(fall_in_cnt) );
1035 _head->set_req( fall_in_cnt, tmp );
1036 // Swap also all Phis
1037 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1038 Node* phi = _head->fast_out(i);
1039 if( phi->is_Phi() ) {
1040 igvn.hash_delete(phi); // Yank from hash before hacking edges
1041 tmp = phi->in(1);
1042 phi->set_req( 1, phi->in(fall_in_cnt) );
1043 phi->set_req( fall_in_cnt, tmp );
1044 }
1045 }
1046 }
1047 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1048 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1049
1050 // If I am a shared header (multiple backedges), peel off the many
1051 // backedges into a private merge point and use the merge point as
1052 // the one true backedge.
1053 if( _head->req() > 3 ) {
1054 // Merge the many backedges into a single backedge.
1055 merge_many_backedges( phase );
1056 result = true;
1057 }
1058
1059 // If I am a shared header (multiple backedges), peel off myself loop.
1060 // I better be the outermost loop.
1061 if( _head->req() > 3 ) {
1062 split_outer_loop( phase );
1063 result = true;
1064
1065 } else if( !_head->is_Loop() && !_irreducible ) {
1066 // Make a new LoopNode to replace the old loop head
1067 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1068 l = igvn.register_new_node_with_optimizer(l, _head);
1069 phase->set_created_loop_node();
1070 // Go ahead and replace _head
1071 phase->_igvn.subsume_node( _head, l );
1072 _head = l;
1073 phase->set_loop(_head, this);
1074 for (DUIterator_Fast imax, i = l->fast_outs(imax); i < imax; i++)
1075 phase->_igvn.add_users_to_worklist(l->fast_out(i));
1076 }
1077
1078 phase->C->print_method("After beautify loops", 3);
1079
1080 // Now recursively beautify nested loops
1081 if( _child ) result |= _child->beautify_loops( phase );
1082 if( _next ) result |= _next ->beautify_loops( phase );
1083 return result;
1084 }
1085
1086 //------------------------------allpaths_check_safepts----------------------------
1087 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1088 // encountered. Helper for check_safepts.
1089 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1090 assert(stack.size() == 0, "empty stack");
1091 stack.push(_tail);
1092 visited.Clear();
1093 visited.set(_tail->_idx);
1094 while (stack.size() > 0) {
1095 Node* n = stack.pop();
1096 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1097 // Terminate this path
1098 } else if (n->Opcode() == Op_SafePoint) {
1099 if (_phase->get_loop(n) != this) {
1100 if (_required_safept == NULL) _required_safept = new Node_List();
1101 _required_safept->push(n); // save the one closest to the tail
1102 }
1103 // Terminate this path
1104 } else {
1105 uint start = n->is_Region() ? 1 : 0;
1106 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1107 for (uint i = start; i < end; i++) {
1108 Node* in = n->in(i);
1109 assert(in->is_CFG(), "must be");
1110 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1111 stack.push(in);
1112 }
1113 }
1114 }
1115 }
1116 }
1117
1118 //------------------------------check_safepts----------------------------
1119 // Given dominators, try to find loops with calls that must always be
1120 // executed (call dominates loop tail). These loops do not need non-call
1121 // safepoints (ncsfpt).
1122 //
1123 // A complication is that a safepoint in a inner loop may be needed
1124 // by an outer loop. In the following, the inner loop sees it has a
1125 // call (block 3) on every path from the head (block 2) to the
1126 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1127 // in block 2, _but_ this leaves the outer loop without a safepoint.
1128 //
1129 // entry 0
1130 // |
1131 // v
1132 // outer 1,2 +->1
1133 // | |
1134 // | v
1135 // | 2<---+ ncsfpt in 2
1136 // |_/|\ |
1137 // | v |
1138 // inner 2,3 / 3 | call in 3
1139 // / | |
1140 // v +--+
1141 // exit 4
1142 //
1143 //
1144 // This method creates a list (_required_safept) of ncsfpt nodes that must
1145 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1146 // is first looked for in the lists for the outer loops of the current loop.
1147 //
1148 // The insights into the problem:
1149 // A) counted loops are okay
1150 // B) innermost loops are okay (only an inner loop can delete
1151 // a ncsfpt needed by an outer loop)
1152 // C) a loop is immune from an inner loop deleting a safepoint
1153 // if the loop has a call on the idom-path
1154 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1155 // idom-path that is not in a nested loop
1156 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1157 // loop needs to be prevented from deletion by an inner loop
1158 //
1159 // There are two analyses:
1160 // 1) The first, and cheaper one, scans the loop body from
1161 // tail to head following the idom (immediate dominator)
1162 // chain, looking for the cases (C,D,E) above.
1163 // Since inner loops are scanned before outer loops, there is summary
1164 // information about inner loops. Inner loops can be skipped over
1165 // when the tail of an inner loop is encountered.
1166 //
1167 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1168 // the idom path (which is rare), scans all predecessor control paths
1169 // from the tail to the head, terminating a path when a call or sfpt
1170 // is encountered, to find the ncsfpt's that are closest to the tail.
1171 //
1172 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1173 // Bottom up traversal
1174 IdealLoopTree* ch = _child;
1175 while (ch != NULL) {
1176 ch->check_safepts(visited, stack);
1177 ch = ch->_next;
1178 }
1179
1180 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1181 bool has_call = false; // call on dom-path
1182 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1183 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1184 // Scan the dom-path nodes from tail to head
1185 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1186 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1187 has_call = true;
1188 _has_sfpt = 1; // Then no need for a safept!
1189 break;
1190 } else if (n->Opcode() == Op_SafePoint) {
1191 if (_phase->get_loop(n) == this) {
1192 has_local_ncsfpt = true;
1193 break;
1194 }
1195 if (nonlocal_ncsfpt == NULL) {
1196 nonlocal_ncsfpt = n; // save the one closest to the tail
1197 }
1198 } else {
1199 IdealLoopTree* nlpt = _phase->get_loop(n);
1200 if (this != nlpt) {
1201 // If at an inner loop tail, see if the inner loop has already
1202 // recorded seeing a call on the dom-path (and stop.) If not,
1203 // jump to the head of the inner loop.
1204 assert(is_member(nlpt), "nested loop");
1205 Node* tail = nlpt->_tail;
1206 if (tail->in(0)->is_If()) tail = tail->in(0);
1207 if (n == tail) {
1208 // If inner loop has call on dom-path, so does outer loop
1209 if (nlpt->_has_sfpt) {
1210 has_call = true;
1211 _has_sfpt = 1;
1212 break;
1213 }
1214 // Skip to head of inner loop
1215 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1216 n = nlpt->_head;
1217 }
1218 }
1219 }
1220 }
1221 // Record safept's that this loop needs preserved when an
1222 // inner loop attempts to delete it's safepoints.
1223 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1224 if (nonlocal_ncsfpt != NULL) {
1225 if (_required_safept == NULL) _required_safept = new Node_List();
1226 _required_safept->push(nonlocal_ncsfpt);
1227 } else {
1228 // Failed to find a suitable safept on the dom-path. Now use
1229 // an all paths walk from tail to head, looking for safepoints to preserve.
1230 allpaths_check_safepts(visited, stack);
1231 }
1232 }
1233 }
1234 }
1235
1236 //---------------------------is_deleteable_safept----------------------------
1237 // Is safept not required by an outer loop?
1238 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1239 assert(sfpt->Opcode() == Op_SafePoint, "");
1240 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1241 while (lp != NULL) {
1242 Node_List* sfpts = lp->_required_safept;
1243 if (sfpts != NULL) {
1244 for (uint i = 0; i < sfpts->size(); i++) {
1245 if (sfpt == sfpts->at(i))
1246 return false;
1247 }
1248 }
1249 lp = lp->_parent;
1250 }
1251 return true;
1252 }
1253
1254 //------------------------------counted_loop-----------------------------------
1255 // Convert to counted loops where possible
1256 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1257
1258 // For grins, set the inner-loop flag here
1259 if( !_child ) {
1260 if( _head->is_Loop() ) _head->as_Loop()->set_inner_loop();
1261 }
1262
1263 if( _head->is_CountedLoop() ||
1264 phase->is_counted_loop( _head, this ) ) {
1265 _has_sfpt = 1; // Indicate we do not need a safepoint here
1266
1267 // Look for a safepoint to remove
1268 for (Node* n = tail(); n != _head; n = phase->idom(n))
1269 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1270 phase->is_deleteable_safept(n))
1271 phase->lazy_replace(n,n->in(TypeFunc::Control));
1272
1273 CountedLoopNode *cl = _head->as_CountedLoop();
1274 Node *incr = cl->incr();
1275 if( !incr ) return; // Dead loop?
1276 Node *init = cl->init_trip();
1277 Node *phi = cl->phi();
1278 // protect against stride not being a constant
1279 if( !cl->stride_is_con() ) return;
1280 int stride_con = cl->stride_con();
1281
1282 // Look for induction variables
1283
1284 // Visit all children, looking for Phis
1285 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1286 Node *out = cl->out(i);
1287 if (!out->is_Phi()) continue; // Looking for phis
1288 PhiNode* phi2 = out->as_Phi();
1289 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1290 // Look for induction variables of the form: X += constant
1291 if( phi2->region() != _head ||
1292 incr2->req() != 3 ||
1293 incr2->in(1) != phi2 ||
1294 incr2 == incr ||
1295 incr2->Opcode() != Op_AddI ||
1296 !incr2->in(2)->is_Con() )
1297 continue;
1298
1299 // Check for parallel induction variable (parallel to trip counter)
1300 // via an affine function. In particular, count-down loops with
1301 // count-up array indices are common. We only RCE references off
1302 // the trip-counter, so we need to convert all these to trip-counter
1303 // expressions.
1304 Node *init2 = phi2->in( LoopNode::EntryControl );
1305 int stride_con2 = incr2->in(2)->get_int();
1306
1307 // The general case here gets a little tricky. We want to find the
1308 // GCD of all possible parallel IV's and make a new IV using this
1309 // GCD for the loop. Then all possible IVs are simple multiples of
1310 // the GCD. In practice, this will cover very few extra loops.
1311 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1312 // where +/-1 is the common case, but other integer multiples are
1313 // also easy to handle.
1314 int ratio_con = stride_con2/stride_con;
1315
1316 if( ratio_con * stride_con == stride_con2 ) { // Check for exact
1317 // Convert to using the trip counter. The parallel induction
1318 // variable differs from the trip counter by a loop-invariant
1319 // amount, the difference between their respective initial values.
1320 // It is scaled by the 'ratio_con'.
1321 Compile* C = phase->C;
1322 Node* ratio = phase->_igvn.intcon(ratio_con);
1323 phase->set_ctrl(ratio, C->root());
1324 Node* ratio_init = new (C, 3) MulINode(init, ratio);
1325 phase->_igvn.register_new_node_with_optimizer(ratio_init, init);
1326 phase->set_early_ctrl(ratio_init);
1327 Node* diff = new (C, 3) SubINode(init2, ratio_init);
1328 phase->_igvn.register_new_node_with_optimizer(diff, init2);
1329 phase->set_early_ctrl(diff);
1330 Node* ratio_idx = new (C, 3) MulINode(phi, ratio);
1331 phase->_igvn.register_new_node_with_optimizer(ratio_idx, phi);
1332 phase->set_ctrl(ratio_idx, cl);
1333 Node* add = new (C, 3) AddINode(ratio_idx, diff);
1334 phase->_igvn.register_new_node_with_optimizer(add);
1335 phase->set_ctrl(add, cl);
1336 phase->_igvn.hash_delete( phi2 );
1337 phase->_igvn.subsume_node( phi2, add );
1338 // Sometimes an induction variable is unused
1339 if (add->outcnt() == 0) {
1340 phase->_igvn.remove_dead_node(add);
1341 }
1342 --i; // deleted this phi; rescan starting with next position
1343 continue;
1344 }
1345 }
1346 } else if (_parent != NULL && !_irreducible) {
1347 // Not a counted loop.
1348 // Look for a safepoint on the idom-path to remove, preserving the first one
1349 bool found = false;
1350 Node* n = tail();
1351 for (; n != _head && !found; n = phase->idom(n)) {
1352 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1353 found = true; // Found one
1354 }
1355 // Skip past it and delete the others
1356 for (; n != _head; n = phase->idom(n)) {
1357 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1358 phase->is_deleteable_safept(n))
1359 phase->lazy_replace(n,n->in(TypeFunc::Control));
1360 }
1361 }
1362
1363 // Recursively
1364 if( _child ) _child->counted_loop( phase );
1365 if( _next ) _next ->counted_loop( phase );
1366 }
1367
1368 #ifndef PRODUCT
1369 //------------------------------dump_head--------------------------------------
1370 // Dump 1 liner for loop header info
1371 void IdealLoopTree::dump_head( ) const {
1372 for( uint i=0; i<_nest; i++ )
1373 tty->print(" ");
1374 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1375 if( _irreducible ) tty->print(" IRREDUCIBLE");
1376 if( _head->is_CountedLoop() ) {
1377 CountedLoopNode *cl = _head->as_CountedLoop();
1378 tty->print(" counted");
1379 if( cl->is_pre_loop () ) tty->print(" pre" );
1380 if( cl->is_main_loop() ) tty->print(" main");
1381 if( cl->is_post_loop() ) tty->print(" post");
1382 }
1383 tty->cr();
1384 }
1385
1386 //------------------------------dump-------------------------------------------
1387 // Dump loops by loop tree
1388 void IdealLoopTree::dump( ) const {
1389 dump_head();
1390 if( _child ) _child->dump();
1391 if( _next ) _next ->dump();
1392 }
1393
1394 #endif
1395
1396 //=============================================================================
1397 //------------------------------PhaseIdealLoop---------------------------------
1398 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1399 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1400 PhaseIdealLoop::PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs )
1401 : PhaseTransform(Ideal_Loop),
1402 _igvn(igvn),
1403 _dom_lca_tags(C->comp_arena()) {
1404 // Reset major-progress flag for the driver's heuristics
1405 C->clear_major_progress();
1406
1407 #ifndef PRODUCT
1408 // Capture for later assert
1409 uint unique = C->unique();
1410 _loop_invokes++;
1411 _loop_work += unique;
1412 #endif
1413
1414 // True if the method has at least 1 irreducible loop
1415 _has_irreducible_loops = false;
1416
1417 _created_loop_node = false;
1418
1419 Arena *a = Thread::current()->resource_area();
1420 VectorSet visited(a);
1421 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1422 _nodes.map(C->unique(), NULL);
1423 memset(_nodes.adr(), 0, wordSize * C->unique());
1424
1425 // Pre-build the top-level outermost loop tree entry
1426 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1427 // Do not need a safepoint at the top level
1428 _ltree_root->_has_sfpt = 1;
1429
1430 // Empty pre-order array
1431 allocate_preorders();
1432
1433 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1434 // IdealLoopTree entries. Data nodes are NOT walked.
1435 build_loop_tree();
1436 // Check for bailout, and return
1437 if (C->failing()) {
1438 return;
1439 }
1440
1441 // No loops after all
1442 if( !_ltree_root->_child ) C->set_has_loops(false);
1443
1444 // There should always be an outer loop containing the Root and Return nodes.
1445 // If not, we have a degenerate empty program. Bail out in this case.
1446 if (!has_node(C->root())) {
1447 C->clear_major_progress();
1448 C->record_method_not_compilable("empty program detected during loop optimization");
1449 return;
1450 }
1451
1452 // Nothing to do, so get out
1453 if( !C->has_loops() && !do_split_ifs && !verify_me) {
1454 _igvn.optimize(); // Cleanup NeverBranches
1455 return;
1456 }
1457
1458 // Set loop nesting depth
1459 _ltree_root->set_nest( 0 );
1460
1461 // Split shared headers and insert loop landing pads.
1462 // Do not bother doing this on the Root loop of course.
1463 if( !verify_me && _ltree_root->_child ) {
1464 if( _ltree_root->_child->beautify_loops( this ) ) {
1465 // Re-build loop tree!
1466 _ltree_root->_child = NULL;
1467 _nodes.clear();
1468 reallocate_preorders();
1469 build_loop_tree();
1470 // Check for bailout, and return
1471 if (C->failing()) {
1472 return;
1473 }
1474 // Reset loop nesting depth
1475 _ltree_root->set_nest( 0 );
1476 }
1477 }
1478
1479 // Build Dominators for elision of NULL checks & loop finding.
1480 // Since nodes do not have a slot for immediate dominator, make
1481 // a persistant side array for that info indexed on node->_idx.
1482 _idom_size = C->unique();
1483 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1484 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1485 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1486 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1487
1488 Dominators();
1489
1490 // As a side effect, Dominators removed any unreachable CFG paths
1491 // into RegionNodes. It doesn't do this test against Root, so
1492 // we do it here.
1493 for( uint i = 1; i < C->root()->req(); i++ ) {
1494 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1495 _igvn.hash_delete(C->root());
1496 C->root()->del_req(i);
1497 _igvn._worklist.push(C->root());
1498 i--; // Rerun same iteration on compressed edges
1499 }
1500 }
1501
1502 // Given dominators, try to find inner loops with calls that must
1503 // always be executed (call dominates loop tail). These loops do
1504 // not need a seperate safepoint.
1505 Node_List cisstack(a);
1506 _ltree_root->check_safepts(visited, cisstack);
1507
1508 // Walk the DATA nodes and place into loops. Find earliest control
1509 // node. For CFG nodes, the _nodes array starts out and remains
1510 // holding the associated IdealLoopTree pointer. For DATA nodes, the
1511 // _nodes array holds the earliest legal controlling CFG node.
1512
1513 // Allocate stack with enough space to avoid frequent realloc
1514 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
1515 Node_Stack nstack( a, stack_size );
1516
1517 visited.Clear();
1518 Node_List worklist(a);
1519 // Don't need C->root() on worklist since
1520 // it will be processed among C->top() inputs
1521 worklist.push( C->top() );
1522 visited.set( C->top()->_idx ); // Set C->top() as visited now
1523 build_loop_early( visited, worklist, nstack, verify_me );
1524
1525 // Given early legal placement, try finding counted loops. This placement
1526 // is good enough to discover most loop invariants.
1527 if( !verify_me )
1528 _ltree_root->counted_loop( this );
1529
1530 // Find latest loop placement. Find ideal loop placement.
1531 visited.Clear();
1532 init_dom_lca_tags();
1533 // Need C->root() on worklist when processing outs
1534 worklist.push( C->root() );
1535 NOT_PRODUCT( C->verify_graph_edges(); )
1536 worklist.push( C->top() );
1537 build_loop_late( visited, worklist, nstack, verify_me );
1538
1539 // clear out the dead code
1540 while(_deadlist.size()) {
1541 igvn.remove_globally_dead_node(_deadlist.pop());
1542 }
1543
1544 #ifndef PRODUCT
1545 C->verify_graph_edges();
1546 if( verify_me ) { // Nested verify pass?
1547 // Check to see if the verify mode is broken
1548 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
1549 return;
1550 }
1551 if( VerifyLoopOptimizations ) verify();
1552 #endif
1553
1554 if (ReassociateInvariants) {
1555 // Reassociate invariants and prep for split_thru_phi
1556 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1557 IdealLoopTree* lpt = iter.current();
1558 if (!lpt->is_counted() || !lpt->is_inner()) continue;
1559
1560 lpt->reassociate_invariants(this);
1561
1562 // Because RCE opportunities can be masked by split_thru_phi,
1563 // look for RCE candidates and inhibit split_thru_phi
1564 // on just their loop-phi's for this pass of loop opts
1565 if( SplitIfBlocks && do_split_ifs ) {
1566 if (lpt->policy_range_check(this)) {
1567 lpt->_rce_candidate = true;
1568 }
1569 }
1570 }
1571 }
1572
1573 // Check for aggressive application of split-if and other transforms
1574 // that require basic-block info (like cloning through Phi's)
1575 if( SplitIfBlocks && do_split_ifs ) {
1576 visited.Clear();
1577 split_if_with_blocks( visited, nstack );
1578 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
1579 }
1580
1581 // Perform iteration-splitting on inner loops. Split iterations to avoid
1582 // range checks or one-shot null checks.
1583
1584 // If split-if's didn't hack the graph too bad (no CFG changes)
1585 // then do loop opts.
1586 if( C->has_loops() && !C->major_progress() ) {
1587 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
1588 _ltree_root->_child->iteration_split( this, worklist );
1589 // No verify after peeling! GCM has hoisted code out of the loop.
1590 // After peeling, the hoisted code could sink inside the peeled area.
1591 // The peeling code does not try to recompute the best location for
1592 // all the code before the peeled area, so the verify pass will always
1593 // complain about it.
1594 }
1595 // Do verify graph edges in any case
1596 NOT_PRODUCT( C->verify_graph_edges(); );
1597
1598 if( !do_split_ifs ) {
1599 // We saw major progress in Split-If to get here. We forced a
1600 // pass with unrolling and not split-if, however more split-if's
1601 // might make progress. If the unrolling didn't make progress
1602 // then the major-progress flag got cleared and we won't try
1603 // another round of Split-If. In particular the ever-common
1604 // instance-of/check-cast pattern requires at least 2 rounds of
1605 // Split-If to clear out.
1606 C->set_major_progress();
1607 }
1608
1609 // Repeat loop optimizations if new loops were seen
1610 if (created_loop_node()) {
1611 C->set_major_progress();
1612 }
1613
1614 // Convert scalar to superword operations
1615
1616 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
1617 // SuperWord transform
1618 SuperWord sw(this);
1619 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1620 IdealLoopTree* lpt = iter.current();
1621 if (lpt->is_counted()) {
1622 sw.transform_loop(lpt);
1623 }
1624 }
1625 }
1626
1627 // Cleanup any modified bits
1628 _igvn.optimize();
1629
1630 // Do not repeat loop optimizations if irreducible loops are present
1631 // by claiming no-progress.
1632 if( _has_irreducible_loops )
1633 C->clear_major_progress();
1634 }
1635
1636 #ifndef PRODUCT
1637 //------------------------------print_statistics-------------------------------
1638 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
1639 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
1640 void PhaseIdealLoop::print_statistics() {
1641 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
1642 }
1643
1644 //------------------------------verify-----------------------------------------
1645 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
1646 static int fail; // debug only, so its multi-thread dont care
1647 void PhaseIdealLoop::verify() const {
1648 int old_progress = C->major_progress();
1649 ResourceMark rm;
1650 PhaseIdealLoop loop_verify( _igvn, this, false );
1651 VectorSet visited(Thread::current()->resource_area());
1652
1653 fail = 0;
1654 verify_compare( C->root(), &loop_verify, visited );
1655 assert( fail == 0, "verify loops failed" );
1656 // Verify loop structure is the same
1657 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
1658 // Reset major-progress. It was cleared by creating a verify version of
1659 // PhaseIdealLoop.
1660 for( int i=0; i<old_progress; i++ )
1661 C->set_major_progress();
1662 }
1663
1664 //------------------------------verify_compare---------------------------------
1665 // Make sure me and the given PhaseIdealLoop agree on key data structures
1666 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
1667 if( !n ) return;
1668 if( visited.test_set( n->_idx ) ) return;
1669 if( !_nodes[n->_idx] ) { // Unreachable
1670 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
1671 return;
1672 }
1673
1674 uint i;
1675 for( i = 0; i < n->req(); i++ )
1676 verify_compare( n->in(i), loop_verify, visited );
1677
1678 // Check the '_nodes' block/loop structure
1679 i = n->_idx;
1680 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
1681 if( _nodes[i] != loop_verify->_nodes[i] &&
1682 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
1683 tty->print("Mismatched control setting for: ");
1684 n->dump();
1685 if( fail++ > 10 ) return;
1686 Node *c = get_ctrl_no_update(n);
1687 tty->print("We have it as: ");
1688 if( c->in(0) ) c->dump();
1689 else tty->print_cr("N%d",c->_idx);
1690 tty->print("Verify thinks: ");
1691 if( loop_verify->has_ctrl(n) )
1692 loop_verify->get_ctrl_no_update(n)->dump();
1693 else
1694 loop_verify->get_loop_idx(n)->dump();
1695 tty->cr();
1696 }
1697 } else { // We have a loop
1698 IdealLoopTree *us = get_loop_idx(n);
1699 if( loop_verify->has_ctrl(n) ) {
1700 tty->print("Mismatched loop setting for: ");
1701 n->dump();
1702 if( fail++ > 10 ) return;
1703 tty->print("We have it as: ");
1704 us->dump();
1705 tty->print("Verify thinks: ");
1706 loop_verify->get_ctrl_no_update(n)->dump();
1707 tty->cr();
1708 } else if (!C->major_progress()) {
1709 // Loop selection can be messed up if we did a major progress
1710 // operation, like split-if. Do not verify in that case.
1711 IdealLoopTree *them = loop_verify->get_loop_idx(n);
1712 if( us->_head != them->_head || us->_tail != them->_tail ) {
1713 tty->print("Unequals loops for: ");
1714 n->dump();
1715 if( fail++ > 10 ) return;
1716 tty->print("We have it as: ");
1717 us->dump();
1718 tty->print("Verify thinks: ");
1719 them->dump();
1720 tty->cr();
1721 }
1722 }
1723 }
1724
1725 // Check for immediate dominators being equal
1726 if( i >= _idom_size ) {
1727 if( !n->is_CFG() ) return;
1728 tty->print("CFG Node with no idom: ");
1729 n->dump();
1730 return;
1731 }
1732 if( !n->is_CFG() ) return;
1733 if( n == C->root() ) return; // No IDOM here
1734
1735 assert(n->_idx == i, "sanity");
1736 Node *id = idom_no_update(n);
1737 if( id != loop_verify->idom_no_update(n) ) {
1738 tty->print("Unequals idoms for: ");
1739 n->dump();
1740 if( fail++ > 10 ) return;
1741 tty->print("We have it as: ");
1742 id->dump();
1743 tty->print("Verify thinks: ");
1744 loop_verify->idom_no_update(n)->dump();
1745 tty->cr();
1746 }
1747
1748 }
1749
1750 //------------------------------verify_tree------------------------------------
1751 // Verify that tree structures match. Because the CFG can change, siblings
1752 // within the loop tree can be reordered. We attempt to deal with that by
1753 // reordering the verify's loop tree if possible.
1754 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
1755 assert( _parent == parent, "Badly formed loop tree" );
1756
1757 // Siblings not in same order? Attempt to re-order.
1758 if( _head != loop->_head ) {
1759 // Find _next pointer to update
1760 IdealLoopTree **pp = &loop->_parent->_child;
1761 while( *pp != loop )
1762 pp = &((*pp)->_next);
1763 // Find proper sibling to be next
1764 IdealLoopTree **nn = &loop->_next;
1765 while( (*nn) && (*nn)->_head != _head )
1766 nn = &((*nn)->_next);
1767
1768 // Check for no match.
1769 if( !(*nn) ) {
1770 // Annoyingly, irreducible loops can pick different headers
1771 // after a major_progress operation, so the rest of the loop
1772 // tree cannot be matched.
1773 if (_irreducible && Compile::current()->major_progress()) return;
1774 assert( 0, "failed to match loop tree" );
1775 }
1776
1777 // Move (*nn) to (*pp)
1778 IdealLoopTree *hit = *nn;
1779 *nn = hit->_next;
1780 hit->_next = loop;
1781 *pp = loop;
1782 loop = hit;
1783 // Now try again to verify
1784 }
1785
1786 assert( _head == loop->_head , "mismatched loop head" );
1787 Node *tail = _tail; // Inline a non-updating version of
1788 while( !tail->in(0) ) // the 'tail()' call.
1789 tail = tail->in(1);
1790 assert( tail == loop->_tail, "mismatched loop tail" );
1791
1792 // Counted loops that are guarded should be able to find their guards
1793 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
1794 CountedLoopNode *cl = _head->as_CountedLoop();
1795 Node *init = cl->init_trip();
1796 Node *ctrl = cl->in(LoopNode::EntryControl);
1797 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1798 Node *iff = ctrl->in(0);
1799 assert( iff->Opcode() == Op_If, "" );
1800 Node *bol = iff->in(1);
1801 assert( bol->Opcode() == Op_Bool, "" );
1802 Node *cmp = bol->in(1);
1803 assert( cmp->Opcode() == Op_CmpI, "" );
1804 Node *add = cmp->in(1);
1805 Node *opaq;
1806 if( add->Opcode() == Op_Opaque1 ) {
1807 opaq = add;
1808 } else {
1809 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
1810 assert( add == init, "" );
1811 opaq = cmp->in(2);
1812 }
1813 assert( opaq->Opcode() == Op_Opaque1, "" );
1814
1815 }
1816
1817 if (_child != NULL) _child->verify_tree(loop->_child, this);
1818 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
1819 // Innermost loops need to verify loop bodies,
1820 // but only if no 'major_progress'
1821 int fail = 0;
1822 if (!Compile::current()->major_progress() && _child == NULL) {
1823 for( uint i = 0; i < _body.size(); i++ ) {
1824 Node *n = _body.at(i);
1825 if (n->outcnt() == 0) continue; // Ignore dead
1826 uint j;
1827 for( j = 0; j < loop->_body.size(); j++ )
1828 if( loop->_body.at(j) == n )
1829 break;
1830 if( j == loop->_body.size() ) { // Not found in loop body
1831 // Last ditch effort to avoid assertion: Its possible that we
1832 // have some users (so outcnt not zero) but are still dead.
1833 // Try to find from root.
1834 if (Compile::current()->root()->find(n->_idx)) {
1835 fail++;
1836 tty->print("We have that verify does not: ");
1837 n->dump();
1838 }
1839 }
1840 }
1841 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
1842 Node *n = loop->_body.at(i2);
1843 if (n->outcnt() == 0) continue; // Ignore dead
1844 uint j;
1845 for( j = 0; j < _body.size(); j++ )
1846 if( _body.at(j) == n )
1847 break;
1848 if( j == _body.size() ) { // Not found in loop body
1849 // Last ditch effort to avoid assertion: Its possible that we
1850 // have some users (so outcnt not zero) but are still dead.
1851 // Try to find from root.
1852 if (Compile::current()->root()->find(n->_idx)) {
1853 fail++;
1854 tty->print("Verify has that we do not: ");
1855 n->dump();
1856 }
1857 }
1858 }
1859 assert( !fail, "loop body mismatch" );
1860 }
1861 }
1862
1863 #endif
1864
1865 //------------------------------set_idom---------------------------------------
1866 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
1867 uint idx = d->_idx;
1868 if (idx >= _idom_size) {
1869 uint newsize = _idom_size<<1;
1870 while( idx >= newsize ) {
1871 newsize <<= 1;
1872 }
1873 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
1874 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
1875 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
1876 _idom_size = newsize;
1877 }
1878 _idom[idx] = n;
1879 _dom_depth[idx] = dom_depth;
1880 }
1881
1882 //------------------------------recompute_dom_depth---------------------------------------
1883 // The dominator tree is constructed with only parent pointers.
1884 // This recomputes the depth in the tree by first tagging all
1885 // nodes as "no depth yet" marker. The next pass then runs up
1886 // the dom tree from each node marked "no depth yet", and computes
1887 // the depth on the way back down.
1888 void PhaseIdealLoop::recompute_dom_depth() {
1889 uint no_depth_marker = C->unique();
1890 uint i;
1891 // Initialize depth to "no depth yet"
1892 for (i = 0; i < _idom_size; i++) {
1893 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
1894 _dom_depth[i] = no_depth_marker;
1895 }
1896 }
1897 if (_dom_stk == NULL) {
1898 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
1899 if (init_size < 10) init_size = 10;
1900 _dom_stk = new (C->node_arena()) GrowableArray<uint>(C->node_arena(), init_size, 0, 0);
1901 }
1902 // Compute new depth for each node.
1903 for (i = 0; i < _idom_size; i++) {
1904 uint j = i;
1905 // Run up the dom tree to find a node with a depth
1906 while (_dom_depth[j] == no_depth_marker) {
1907 _dom_stk->push(j);
1908 j = _idom[j]->_idx;
1909 }
1910 // Compute the depth on the way back down this tree branch
1911 uint dd = _dom_depth[j] + 1;
1912 while (_dom_stk->length() > 0) {
1913 uint j = _dom_stk->pop();
1914 _dom_depth[j] = dd;
1915 dd++;
1916 }
1917 }
1918 }
1919
1920 //------------------------------sort-------------------------------------------
1921 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
1922 // loop tree, not the root.
1923 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
1924 if( !innermost ) return loop; // New innermost loop
1925
1926 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
1927 assert( loop_preorder, "not yet post-walked loop" );
1928 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
1929 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
1930
1931 // Insert at start of list
1932 while( l ) { // Insertion sort based on pre-order
1933 if( l == loop ) return innermost; // Already on list!
1934 int l_preorder = get_preorder(l->_head); // Cache pre-order number
1935 assert( l_preorder, "not yet post-walked l" );
1936 // Check header pre-order number to figure proper nesting
1937 if( loop_preorder > l_preorder )
1938 break; // End of insertion
1939 // If headers tie (e.g., shared headers) check tail pre-order numbers.
1940 // Since I split shared headers, you'd think this could not happen.
1941 // BUT: I must first do the preorder numbering before I can discover I
1942 // have shared headers, so the split headers all get the same preorder
1943 // number as the RegionNode they split from.
1944 if( loop_preorder == l_preorder &&
1945 get_preorder(loop->_tail) < get_preorder(l->_tail) )
1946 break; // Also check for shared headers (same pre#)
1947 pp = &l->_parent; // Chain up list
1948 l = *pp;
1949 }
1950 // Link into list
1951 // Point predecessor to me
1952 *pp = loop;
1953 // Point me to successor
1954 IdealLoopTree *p = loop->_parent;
1955 loop->_parent = l; // Point me to successor
1956 if( p ) sort( p, innermost ); // Insert my parents into list as well
1957 return innermost;
1958 }
1959
1960 //------------------------------build_loop_tree--------------------------------
1961 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
1962 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
1963 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
1964 // tightest enclosing IdealLoopTree for post-walked.
1965 //
1966 // During my forward walk I do a short 1-layer lookahead to see if I can find
1967 // a loop backedge with that doesn't have any work on the backedge. This
1968 // helps me construct nested loops with shared headers better.
1969 //
1970 // Once I've done the forward recursion, I do the post-work. For each child
1971 // I check to see if there is a backedge. Backedges define a loop! I
1972 // insert an IdealLoopTree at the target of the backedge.
1973 //
1974 // During the post-work I also check to see if I have several children
1975 // belonging to different loops. If so, then this Node is a decision point
1976 // where control flow can choose to change loop nests. It is at this
1977 // decision point where I can figure out how loops are nested. At this
1978 // time I can properly order the different loop nests from my children.
1979 // Note that there may not be any backedges at the decision point!
1980 //
1981 // Since the decision point can be far removed from the backedges, I can't
1982 // order my loops at the time I discover them. Thus at the decision point
1983 // I need to inspect loop header pre-order numbers to properly nest my
1984 // loops. This means I need to sort my childrens' loops by pre-order.
1985 // The sort is of size number-of-control-children, which generally limits
1986 // it to size 2 (i.e., I just choose between my 2 target loops).
1987 void PhaseIdealLoop::build_loop_tree() {
1988 // Allocate stack of size C->unique()/2 to avoid frequent realloc
1989 GrowableArray <Node *> bltstack(C->unique() >> 1);
1990 Node *n = C->root();
1991 bltstack.push(n);
1992 int pre_order = 1;
1993 int stack_size;
1994
1995 while ( ( stack_size = bltstack.length() ) != 0 ) {
1996 n = bltstack.top(); // Leave node on stack
1997 if ( !is_visited(n) ) {
1998 // ---- Pre-pass Work ----
1999 // Pre-walked but not post-walked nodes need a pre_order number.
2000
2001 set_preorder_visited( n, pre_order ); // set as visited
2002
2003 // ---- Scan over children ----
2004 // Scan first over control projections that lead to loop headers.
2005 // This helps us find inner-to-outer loops with shared headers better.
2006
2007 // Scan children's children for loop headers.
2008 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2009 Node* m = n->raw_out(i); // Child
2010 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2011 // Scan over children's children to find loop
2012 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2013 Node* l = m->fast_out(j);
2014 if( is_visited(l) && // Been visited?
2015 !is_postvisited(l) && // But not post-visited
2016 get_preorder(l) < pre_order ) { // And smaller pre-order
2017 // Found! Scan the DFS down this path before doing other paths
2018 bltstack.push(m);
2019 break;
2020 }
2021 }
2022 }
2023 }
2024 pre_order++;
2025 }
2026 else if ( !is_postvisited(n) ) {
2027 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2028 // such as com.sun.rsasign.am::a.
2029 // For non-recursive version, first, process current children.
2030 // On next iteration, check if additional children were added.
2031 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2032 Node* u = n->raw_out(k);
2033 if ( u->is_CFG() && !is_visited(u) ) {
2034 bltstack.push(u);
2035 }
2036 }
2037 if ( bltstack.length() == stack_size ) {
2038 // There were no additional children, post visit node now
2039 (void)bltstack.pop(); // Remove node from stack
2040 pre_order = build_loop_tree_impl( n, pre_order );
2041 // Check for bailout
2042 if (C->failing()) {
2043 return;
2044 }
2045 // Check to grow _preorders[] array for the case when
2046 // build_loop_tree_impl() adds new nodes.
2047 check_grow_preorders();
2048 }
2049 }
2050 else {
2051 (void)bltstack.pop(); // Remove post-visited node from stack
2052 }
2053 }
2054 }
2055
2056 //------------------------------build_loop_tree_impl---------------------------
2057 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2058 // ---- Post-pass Work ----
2059 // Pre-walked but not post-walked nodes need a pre_order number.
2060
2061 // Tightest enclosing loop for this Node
2062 IdealLoopTree *innermost = NULL;
2063
2064 // For all children, see if any edge is a backedge. If so, make a loop
2065 // for it. Then find the tightest enclosing loop for the self Node.
2066 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2067 Node* m = n->fast_out(i); // Child
2068 if( n == m ) continue; // Ignore control self-cycles
2069 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2070
2071 IdealLoopTree *l; // Child's loop
2072 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2073 // Found a backedge
2074 assert( get_preorder(m) < pre_order, "should be backedge" );
2075 // Check for the RootNode, which is already a LoopNode and is allowed
2076 // to have multiple "backedges".
2077 if( m == C->root()) { // Found the root?
2078 l = _ltree_root; // Root is the outermost LoopNode
2079 } else { // Else found a nested loop
2080 // Insert a LoopNode to mark this loop.
2081 l = new IdealLoopTree(this, m, n);
2082 } // End of Else found a nested loop
2083 if( !has_loop(m) ) // If 'm' does not already have a loop set
2084 set_loop(m, l); // Set loop header to loop now
2085
2086 } else { // Else not a nested loop
2087 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2088 l = get_loop(m); // Get previously determined loop
2089 // If successor is header of a loop (nest), move up-loop till it
2090 // is a member of some outer enclosing loop. Since there are no
2091 // shared headers (I've split them already) I only need to go up
2092 // at most 1 level.
2093 while( l && l->_head == m ) // Successor heads loop?
2094 l = l->_parent; // Move up 1 for me
2095 // If this loop is not properly parented, then this loop
2096 // has no exit path out, i.e. its an infinite loop.
2097 if( !l ) {
2098 // Make loop "reachable" from root so the CFG is reachable. Basically
2099 // insert a bogus loop exit that is never taken. 'm', the loop head,
2100 // points to 'n', one (of possibly many) fall-in paths. There may be
2101 // many backedges as well.
2102
2103 // Here I set the loop to be the root loop. I could have, after
2104 // inserting a bogus loop exit, restarted the recursion and found my
2105 // new loop exit. This would make the infinite loop a first-class
2106 // loop and it would then get properly optimized. What's the use of
2107 // optimizing an infinite loop?
2108 l = _ltree_root; // Oops, found infinite loop
2109
2110 // Insert the NeverBranch between 'm' and it's control user.
2111 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2112 _igvn.register_new_node_with_optimizer(iff);
2113 set_loop(iff, l);
2114 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2115 _igvn.register_new_node_with_optimizer(if_t);
2116 set_loop(if_t, l);
2117
2118 Node* cfg = NULL; // Find the One True Control User of m
2119 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2120 Node* x = m->fast_out(j);
2121 if (x->is_CFG() && x != m && x != iff)
2122 { cfg = x; break; }
2123 }
2124 assert(cfg != NULL, "must find the control user of m");
2125 uint k = 0; // Probably cfg->in(0)
2126 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2127 cfg->set_req( k, if_t ); // Now point to NeverBranch
2128
2129 // Now create the never-taken loop exit
2130 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2131 _igvn.register_new_node_with_optimizer(if_f);
2132 set_loop(if_f, l);
2133 // Find frame ptr for Halt. Relies on the optimizer
2134 // V-N'ing. Easier and quicker than searching through
2135 // the program structure.
2136 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2137 _igvn.register_new_node_with_optimizer(frame);
2138 // Halt & Catch Fire
2139 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2140 _igvn.register_new_node_with_optimizer(halt);
2141 set_loop(halt, l);
2142 C->root()->add_req(halt);
2143 set_loop(C->root(), _ltree_root);
2144 }
2145 }
2146 // Weeny check for irreducible. This child was already visited (this
2147 // IS the post-work phase). Is this child's loop header post-visited
2148 // as well? If so, then I found another entry into the loop.
2149 while( is_postvisited(l->_head) ) {
2150 // found irreducible
2151 l->_irreducible = true;
2152 l = l->_parent;
2153 _has_irreducible_loops = true;
2154 // Check for bad CFG here to prevent crash, and bailout of compile
2155 if (l == NULL) {
2156 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2157 return pre_order;
2158 }
2159 }
2160
2161 // This Node might be a decision point for loops. It is only if
2162 // it's children belong to several different loops. The sort call
2163 // does a trivial amount of work if there is only 1 child or all
2164 // children belong to the same loop. If however, the children
2165 // belong to different loops, the sort call will properly set the
2166 // _parent pointers to show how the loops nest.
2167 //
2168 // In any case, it returns the tightest enclosing loop.
2169 innermost = sort( l, innermost );
2170 }
2171
2172 // Def-use info will have some dead stuff; dead stuff will have no
2173 // loop decided on.
2174
2175 // Am I a loop header? If so fix up my parent's child and next ptrs.
2176 if( innermost && innermost->_head == n ) {
2177 assert( get_loop(n) == innermost, "" );
2178 IdealLoopTree *p = innermost->_parent;
2179 IdealLoopTree *l = innermost;
2180 while( p && l->_head == n ) {
2181 l->_next = p->_child; // Put self on parents 'next child'
2182 p->_child = l; // Make self as first child of parent
2183 l = p; // Now walk up the parent chain
2184 p = l->_parent;
2185 }
2186 } else {
2187 // Note that it is possible for a LoopNode to reach here, if the
2188 // backedge has been made unreachable (hence the LoopNode no longer
2189 // denotes a Loop, and will eventually be removed).
2190
2191 // Record tightest enclosing loop for self. Mark as post-visited.
2192 set_loop(n, innermost);
2193 // Also record has_call flag early on
2194 if( innermost ) {
2195 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2196 // Do not count uncommon calls
2197 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2198 Node *iff = n->in(0)->in(0);
2199 if( !iff->is_If() ||
2200 (n->in(0)->Opcode() == Op_IfFalse &&
2201 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2202 (iff->as_If()->_prob >= 0.01) )
2203 innermost->_has_call = 1;
2204 }
2205 }
2206 }
2207 }
2208
2209 // Flag as post-visited now
2210 set_postvisited(n);
2211 return pre_order;
2212 }
2213
2214
2215 //------------------------------build_loop_early-------------------------------
2216 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2217 // First pass computes the earliest controlling node possible. This is the
2218 // controlling input with the deepest dominating depth.
2219 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) {
2220 while (worklist.size() != 0) {
2221 // Use local variables nstack_top_n & nstack_top_i to cache values
2222 // on nstack's top.
2223 Node *nstack_top_n = worklist.pop();
2224 uint nstack_top_i = 0;
2225 //while_nstack_nonempty:
2226 while (true) {
2227 // Get parent node and next input's index from stack's top.
2228 Node *n = nstack_top_n;
2229 uint i = nstack_top_i;
2230 uint cnt = n->req(); // Count of inputs
2231 if (i == 0) { // Pre-process the node.
2232 if( has_node(n) && // Have either loop or control already?
2233 !has_ctrl(n) ) { // Have loop picked out already?
2234 // During "merge_many_backedges" we fold up several nested loops
2235 // into a single loop. This makes the members of the original
2236 // loop bodies pointing to dead loops; they need to move up
2237 // to the new UNION'd larger loop. I set the _head field of these
2238 // dead loops to NULL and the _parent field points to the owning
2239 // loop. Shades of UNION-FIND algorithm.
2240 IdealLoopTree *ilt;
2241 while( !(ilt = get_loop(n))->_head ) {
2242 // Normally I would use a set_loop here. But in this one special
2243 // case, it is legal (and expected) to change what loop a Node
2244 // belongs to.
2245 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2246 }
2247 // Remove safepoints ONLY if I've already seen I don't need one.
2248 // (the old code here would yank a 2nd safepoint after seeing a
2249 // first one, even though the 1st did not dominate in the loop body
2250 // and thus could be avoided indefinitely)
2251 if( !verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2252 is_deleteable_safept(n)) {
2253 Node *in = n->in(TypeFunc::Control);
2254 lazy_replace(n,in); // Pull safepoint now
2255 // Carry on with the recursion "as if" we are walking
2256 // only the control input
2257 if( !visited.test_set( in->_idx ) ) {
2258 worklist.push(in); // Visit this guy later, using worklist
2259 }
2260 // Get next node from nstack:
2261 // - skip n's inputs processing by setting i > cnt;
2262 // - we also will not call set_early_ctrl(n) since
2263 // has_node(n) == true (see the condition above).
2264 i = cnt + 1;
2265 }
2266 }
2267 } // if (i == 0)
2268
2269 // Visit all inputs
2270 bool done = true; // Assume all n's inputs will be processed
2271 while (i < cnt) {
2272 Node *in = n->in(i);
2273 ++i;
2274 if (in == NULL) continue;
2275 if (in->pinned() && !in->is_CFG())
2276 set_ctrl(in, in->in(0));
2277 int is_visited = visited.test_set( in->_idx );
2278 if (!has_node(in)) { // No controlling input yet?
2279 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2280 assert( !is_visited, "visit only once" );
2281 nstack.push(n, i); // Save parent node and next input's index.
2282 nstack_top_n = in; // Process current input now.
2283 nstack_top_i = 0;
2284 done = false; // Not all n's inputs processed.
2285 break; // continue while_nstack_nonempty;
2286 } else if (!is_visited) {
2287 // This guy has a location picked out for him, but has not yet
2288 // been visited. Happens to all CFG nodes, for instance.
2289 // Visit him using the worklist instead of recursion, to break
2290 // cycles. Since he has a location already we do not need to
2291 // find his location before proceeding with the current Node.
2292 worklist.push(in); // Visit this guy later, using worklist
2293 }
2294 }
2295 if (done) {
2296 // All of n's inputs have been processed, complete post-processing.
2297
2298 // Compute earilest point this Node can go.
2299 // CFG, Phi, pinned nodes already know their controlling input.
2300 if (!has_node(n)) {
2301 // Record earliest legal location
2302 set_early_ctrl( n );
2303 }
2304 if (nstack.is_empty()) {
2305 // Finished all nodes on stack.
2306 // Process next node on the worklist.
2307 break;
2308 }
2309 // Get saved parent node and next input's index.
2310 nstack_top_n = nstack.node();
2311 nstack_top_i = nstack.index();
2312 nstack.pop();
2313 }
2314 } // while (true)
2315 }
2316 }
2317
2318 //------------------------------dom_lca_internal--------------------------------
2319 // Pair-wise LCA
2320 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2321 if( !n1 ) return n2; // Handle NULL original LCA
2322 assert( n1->is_CFG(), "" );
2323 assert( n2->is_CFG(), "" );
2324 // find LCA of all uses
2325 uint d1 = dom_depth(n1);
2326 uint d2 = dom_depth(n2);
2327 while (n1 != n2) {
2328 if (d1 > d2) {
2329 n1 = idom(n1);
2330 d1 = dom_depth(n1);
2331 } else if (d1 < d2) {
2332 n2 = idom(n2);
2333 d2 = dom_depth(n2);
2334 } else {
2335 // Here d1 == d2. Due to edits of the dominator-tree, sections
2336 // of the tree might have the same depth. These sections have
2337 // to be searched more carefully.
2338
2339 // Scan up all the n1's with equal depth, looking for n2.
2340 Node *t1 = idom(n1);
2341 while (dom_depth(t1) == d1) {
2342 if (t1 == n2) return n2;
2343 t1 = idom(t1);
2344 }
2345 // Scan up all the n2's with equal depth, looking for n1.
2346 Node *t2 = idom(n2);
2347 while (dom_depth(t2) == d2) {
2348 if (t2 == n1) return n1;
2349 t2 = idom(t2);
2350 }
2351 // Move up to a new dominator-depth value as well as up the dom-tree.
2352 n1 = t1;
2353 n2 = t2;
2354 d1 = dom_depth(n1);
2355 d2 = dom_depth(n2);
2356 }
2357 }
2358 return n1;
2359 }
2360
2361 //------------------------------compute_idom-----------------------------------
2362 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2363 // IDOMs are correct.
2364 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2365 assert( region->is_Region(), "" );
2366 Node *LCA = NULL;
2367 for( uint i = 1; i < region->req(); i++ ) {
2368 if( region->in(i) != C->top() )
2369 LCA = dom_lca( LCA, region->in(i) );
2370 }
2371 return LCA;
2372 }
2373
2374 //------------------------------get_late_ctrl----------------------------------
2375 // Compute latest legal control.
2376 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2377 assert(early != NULL, "early control should not be NULL");
2378
2379 // Compute LCA over list of uses
2380 Node *LCA = NULL;
2381 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2382 Node* c = n->fast_out(i);
2383 if (_nodes[c->_idx] == NULL)
2384 continue; // Skip the occasional dead node
2385 if( c->is_Phi() ) { // For Phis, we must land above on the path
2386 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2387 if( c->in(j) == n ) { // Found matching input?
2388 Node *use = c->in(0)->in(j);
2389 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2390 }
2391 }
2392 } else {
2393 // For CFG data-users, use is in the block just prior
2394 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2395 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2396 }
2397 }
2398
2399 // if this is a load, check for anti-dependent stores
2400 // We use a conservative algorithm to identify potential interfering
2401 // instructions and for rescheduling the load. The users of the memory
2402 // input of this load are examined. Any use which is not a load and is
2403 // dominated by early is considered a potentially interfering store.
2404 // This can produce false positives.
2405 if (n->is_Load() && LCA != early) {
2406 Node_List worklist;
2407
2408 Node *mem = n->in(MemNode::Memory);
2409 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2410 Node* s = mem->fast_out(i);
2411 worklist.push(s);
2412 }
2413 while(worklist.size() != 0 && LCA != early) {
2414 Node* s = worklist.pop();
2415 if (s->is_Load()) {
2416 continue;
2417 } else if (s->is_MergeMem()) {
2418 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
2419 Node* s1 = s->fast_out(i);
2420 worklist.push(s1);
2421 }
2422 } else {
2423 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
2424 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
2425 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
2426 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
2427 }
2428 }
2429 }
2430 }
2431
2432 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
2433 return LCA;
2434 }
2435
2436 // true if CFG node d dominates CFG node n
2437 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
2438 if (d == n)
2439 return true;
2440 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
2441 uint dd = dom_depth(d);
2442 while (dom_depth(n) >= dd) {
2443 if (n == d)
2444 return true;
2445 n = idom(n);
2446 }
2447 return false;
2448 }
2449
2450 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
2451 // Pair-wise LCA with tags.
2452 // Tag each index with the node 'tag' currently being processed
2453 // before advancing up the dominator chain using idom().
2454 // Later calls that find a match to 'tag' know that this path has already
2455 // been considered in the current LCA (which is input 'n1' by convention).
2456 // Since get_late_ctrl() is only called once for each node, the tag array
2457 // does not need to be cleared between calls to get_late_ctrl().
2458 // Algorithm trades a larger constant factor for better asymptotic behavior
2459 //
2460 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
2461 uint d1 = dom_depth(n1);
2462 uint d2 = dom_depth(n2);
2463
2464 do {
2465 if (d1 > d2) {
2466 // current lca is deeper than n2
2467 _dom_lca_tags.map(n1->_idx, tag);
2468 n1 = idom(n1);
2469 d1 = dom_depth(n1);
2470 } else if (d1 < d2) {
2471 // n2 is deeper than current lca
2472 Node *memo = _dom_lca_tags[n2->_idx];
2473 if( memo == tag ) {
2474 return n1; // Return the current LCA
2475 }
2476 _dom_lca_tags.map(n2->_idx, tag);
2477 n2 = idom(n2);
2478 d2 = dom_depth(n2);
2479 } else {
2480 // Here d1 == d2. Due to edits of the dominator-tree, sections
2481 // of the tree might have the same depth. These sections have
2482 // to be searched more carefully.
2483
2484 // Scan up all the n1's with equal depth, looking for n2.
2485 _dom_lca_tags.map(n1->_idx, tag);
2486 Node *t1 = idom(n1);
2487 while (dom_depth(t1) == d1) {
2488 if (t1 == n2) return n2;
2489 _dom_lca_tags.map(t1->_idx, tag);
2490 t1 = idom(t1);
2491 }
2492 // Scan up all the n2's with equal depth, looking for n1.
2493 _dom_lca_tags.map(n2->_idx, tag);
2494 Node *t2 = idom(n2);
2495 while (dom_depth(t2) == d2) {
2496 if (t2 == n1) return n1;
2497 _dom_lca_tags.map(t2->_idx, tag);
2498 t2 = idom(t2);
2499 }
2500 // Move up to a new dominator-depth value as well as up the dom-tree.
2501 n1 = t1;
2502 n2 = t2;
2503 d1 = dom_depth(n1);
2504 d2 = dom_depth(n2);
2505 }
2506 } while (n1 != n2);
2507 return n1;
2508 }
2509
2510 //------------------------------init_dom_lca_tags------------------------------
2511 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2512 // Intended use does not involve any growth for the array, so it could
2513 // be of fixed size.
2514 void PhaseIdealLoop::init_dom_lca_tags() {
2515 uint limit = C->unique() + 1;
2516 _dom_lca_tags.map( limit, NULL );
2517 #ifdef ASSERT
2518 for( uint i = 0; i < limit; ++i ) {
2519 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2520 }
2521 #endif // ASSERT
2522 }
2523
2524 //------------------------------clear_dom_lca_tags------------------------------
2525 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2526 // Intended use does not involve any growth for the array, so it could
2527 // be of fixed size.
2528 void PhaseIdealLoop::clear_dom_lca_tags() {
2529 uint limit = C->unique() + 1;
2530 _dom_lca_tags.map( limit, NULL );
2531 _dom_lca_tags.clear();
2532 #ifdef ASSERT
2533 for( uint i = 0; i < limit; ++i ) {
2534 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2535 }
2536 #endif // ASSERT
2537 }
2538
2539 //------------------------------build_loop_late--------------------------------
2540 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2541 // Second pass finds latest legal placement, and ideal loop placement.
2542 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) {
2543 while (worklist.size() != 0) {
2544 Node *n = worklist.pop();
2545 // Only visit once
2546 if (visited.test_set(n->_idx)) continue;
2547 uint cnt = n->outcnt();
2548 uint i = 0;
2549 while (true) {
2550 assert( _nodes[n->_idx], "no dead nodes" );
2551 // Visit all children
2552 if (i < cnt) {
2553 Node* use = n->raw_out(i);
2554 ++i;
2555 // Check for dead uses. Aggressively prune such junk. It might be
2556 // dead in the global sense, but still have local uses so I cannot
2557 // easily call 'remove_dead_node'.
2558 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
2559 // Due to cycles, we might not hit the same fixed point in the verify
2560 // pass as we do in the regular pass. Instead, visit such phis as
2561 // simple uses of the loop head.
2562 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
2563 if( !visited.test(use->_idx) )
2564 worklist.push(use);
2565 } else if( !visited.test_set(use->_idx) ) {
2566 nstack.push(n, i); // Save parent and next use's index.
2567 n = use; // Process all children of current use.
2568 cnt = use->outcnt();
2569 i = 0;
2570 }
2571 } else {
2572 // Do not visit around the backedge of loops via data edges.
2573 // push dead code onto a worklist
2574 _deadlist.push(use);
2575 }
2576 } else {
2577 // All of n's children have been processed, complete post-processing.
2578 build_loop_late_post(n, verify_me);
2579 if (nstack.is_empty()) {
2580 // Finished all nodes on stack.
2581 // Process next node on the worklist.
2582 break;
2583 }
2584 // Get saved parent node and next use's index. Visit the rest of uses.
2585 n = nstack.node();
2586 cnt = n->outcnt();
2587 i = nstack.index();
2588 nstack.pop();
2589 }
2590 }
2591 }
2592 }
2593
2594 //------------------------------build_loop_late_post---------------------------
2595 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2596 // Second pass finds latest legal placement, and ideal loop placement.
2597 void PhaseIdealLoop::build_loop_late_post( Node *n, const PhaseIdealLoop *verify_me ) {
2598
2599 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress()) {
2600 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
2601 }
2602
2603 // CFG and pinned nodes already handled
2604 if( n->in(0) ) {
2605 if( n->in(0)->is_top() ) return; // Dead?
2606
2607 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
2608 // _must_ be pinned (they have to observe their control edge of course).
2609 // Unlike Stores (which modify an unallocable resource, the memory
2610 // state), Mods/Loads can float around. So free them up.
2611 bool pinned = true;
2612 switch( n->Opcode() ) {
2613 case Op_DivI:
2614 case Op_DivF:
2615 case Op_DivD:
2616 case Op_ModI:
2617 case Op_ModF:
2618 case Op_ModD:
2619 case Op_LoadB: // Same with Loads; they can sink
2620 case Op_LoadC: // during loop optimizations.
2621 case Op_LoadD:
2622 case Op_LoadF:
2623 case Op_LoadI:
2624 case Op_LoadKlass:
2625 case Op_LoadL:
2626 case Op_LoadS:
2627 case Op_LoadP:
2628 case Op_LoadRange:
2629 case Op_LoadD_unaligned:
2630 case Op_LoadL_unaligned:
2631 case Op_StrComp: // Does a bunch of load-like effects
2632 pinned = false;
2633 }
2634 if( pinned ) {
2635 IdealLoopTree *choosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
2636 if( !choosen_loop->_child ) // Inner loop?
2637 choosen_loop->_body.push(n); // Collect inner loops
2638 return;
2639 }
2640 } else { // No slot zero
2641 if( n->is_CFG() ) { // CFG with no slot 0 is dead
2642 _nodes.map(n->_idx,0); // No block setting, it's globally dead
2643 return;
2644 }
2645 assert(!n->is_CFG() || n->outcnt() == 0, "");
2646 }
2647
2648 // Do I have a "safe range" I can select over?
2649 Node *early = get_ctrl(n);// Early location already computed
2650
2651 // Compute latest point this Node can go
2652 Node *LCA = get_late_ctrl( n, early );
2653 // LCA is NULL due to uses being dead
2654 if( LCA == NULL ) {
2655 #ifdef ASSERT
2656 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
2657 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
2658 }
2659 #endif
2660 _nodes.map(n->_idx, 0); // This node is useless
2661 _deadlist.push(n);
2662 return;
2663 }
2664 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
2665
2666 Node *legal = LCA; // Walk 'legal' up the IDOM chain
2667 Node *least = legal; // Best legal position so far
2668 while( early != legal ) { // While not at earliest legal
2669 // Find least loop nesting depth
2670 legal = idom(legal); // Bump up the IDOM tree
2671 // Check for lower nesting depth
2672 if( get_loop(legal)->_nest < get_loop(least)->_nest )
2673 least = legal;
2674 }
2675
2676 // Try not to place code on a loop entry projection
2677 // which can inhibit range check elimination.
2678 if (least != early) {
2679 Node* ctrl_out = least->unique_ctrl_out();
2680 if (ctrl_out && ctrl_out->is_CountedLoop() &&
2681 least == ctrl_out->in(LoopNode::EntryControl)) {
2682 Node* least_dom = idom(least);
2683 if (get_loop(least_dom)->is_member(get_loop(least))) {
2684 least = least_dom;
2685 }
2686 }
2687 }
2688
2689 #ifdef ASSERT
2690 // If verifying, verify that 'verify_me' has a legal location
2691 // and choose it as our location.
2692 if( verify_me ) {
2693 Node *v_ctrl = verify_me->get_ctrl_no_update(n);
2694 Node *legal = LCA;
2695 while( early != legal ) { // While not at earliest legal
2696 if( legal == v_ctrl ) break; // Check for prior good location
2697 legal = idom(legal) ;// Bump up the IDOM tree
2698 }
2699 // Check for prior good location
2700 if( legal == v_ctrl ) least = legal; // Keep prior if found
2701 }
2702 #endif
2703
2704 // Assign discovered "here or above" point
2705 least = find_non_split_ctrl(least);
2706 set_ctrl(n, least);
2707
2708 // Collect inner loop bodies
2709 IdealLoopTree *choosen_loop = get_loop(least);
2710 if( !choosen_loop->_child ) // Inner loop?
2711 choosen_loop->_body.push(n);// Collect inner loops
2712 }
2713
2714 #ifndef PRODUCT
2715 //------------------------------dump-------------------------------------------
2716 void PhaseIdealLoop::dump( ) const {
2717 ResourceMark rm;
2718 Arena* arena = Thread::current()->resource_area();
2719 Node_Stack stack(arena, C->unique() >> 2);
2720 Node_List rpo_list;
2721 VectorSet visited(arena);
2722 visited.set(C->top()->_idx);
2723 rpo( C->root(), stack, visited, rpo_list );
2724 // Dump root loop indexed by last element in PO order
2725 dump( _ltree_root, rpo_list.size(), rpo_list );
2726 }
2727
2728 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
2729
2730 // Indent by loop nesting depth
2731 for( uint x = 0; x < loop->_nest; x++ )
2732 tty->print(" ");
2733 tty->print_cr("---- Loop N%d-N%d ----", loop->_head->_idx,loop->_tail->_idx);
2734
2735 // Now scan for CFG nodes in the same loop
2736 for( uint j=idx; j > 0; j-- ) {
2737 Node *n = rpo_list[j-1];
2738 if( !_nodes[n->_idx] ) // Skip dead nodes
2739 continue;
2740 if( get_loop(n) != loop ) { // Wrong loop nest
2741 if( get_loop(n)->_head == n && // Found nested loop?
2742 get_loop(n)->_parent == loop )
2743 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
2744 continue;
2745 }
2746
2747 // Dump controlling node
2748 for( uint x = 0; x < loop->_nest; x++ )
2749 tty->print(" ");
2750 tty->print("C");
2751 if( n == C->root() ) {
2752 n->dump();
2753 } else {
2754 Node* cached_idom = idom_no_update(n);
2755 Node *computed_idom = n->in(0);
2756 if( n->is_Region() ) {
2757 computed_idom = compute_idom(n);
2758 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
2759 // any MultiBranch ctrl node), so apply a similar transform to
2760 // the cached idom returned from idom_no_update.
2761 cached_idom = find_non_split_ctrl(cached_idom);
2762 }
2763 tty->print(" ID:%d",computed_idom->_idx);
2764 n->dump();
2765 if( cached_idom != computed_idom ) {
2766 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
2767 computed_idom->_idx, cached_idom->_idx);
2768 }
2769 }
2770 // Dump nodes it controls
2771 for( uint k = 0; k < _nodes.Size(); k++ ) {
2772 // (k < C->unique() && get_ctrl(find(k)) == n)
2773 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
2774 Node *m = C->root()->find(k);
2775 if( m && m->outcnt() > 0 ) {
2776 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
2777 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
2778 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
2779 }
2780 for( uint j = 0; j < loop->_nest; j++ )
2781 tty->print(" ");
2782 tty->print(" ");
2783 m->dump();
2784 }
2785 }
2786 }
2787 }
2788 }
2789
2790 // Collect a R-P-O for the whole CFG.
2791 // Result list is in post-order (scan backwards for RPO)
2792 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
2793 stk.push(start, 0);
2794 visited.set(start->_idx);
2795
2796 while (stk.is_nonempty()) {
2797 Node* m = stk.node();
2798 uint idx = stk.index();
2799 if (idx < m->outcnt()) {
2800 stk.set_index(idx + 1);
2801 Node* n = m->raw_out(idx);
2802 if (n->is_CFG() && !visited.test_set(n->_idx)) {
2803 stk.push(n, 0);
2804 }
2805 } else {
2806 rpo_list.push(m);
2807 stk.pop();
2808 }
2809 }
2810 }
2811 #endif
2812
2813
2814 //=============================================================================
2815 //------------------------------LoopTreeIterator-----------------------------------
2816
2817 // Advance to next loop tree using a preorder, left-to-right traversal.
2818 void LoopTreeIterator::next() {
2819 assert(!done(), "must not be done.");
2820 if (_curnt->_child != NULL) {
2821 _curnt = _curnt->_child;
2822 } else if (_curnt->_next != NULL) {
2823 _curnt = _curnt->_next;
2824 } else {
2825 while (_curnt != _root && _curnt->_next == NULL) {
2826 _curnt = _curnt->_parent;
2827 }
2828 if (_curnt == _root) {
2829 _curnt = NULL;
2830 assert(done(), "must be done.");
2831 } else {
2832 assert(_curnt->_next != NULL, "must be more to do");
2833 _curnt = _curnt->_next;
2834 }
2835 }
2836 }
2837