Mon, 20 Aug 2012 09:07:21 -0700
6340864: Implement vectorization optimizations in hotspot-server
Summary: Added asm encoding and mach nodes for vector arithmetic instructions on x86.
Reviewed-by: roland
1 /*
2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/superword.hpp"
40 //=============================================================================
41 //------------------------------is_loop_iv-------------------------------------
42 // Determine if a node is Counted loop induction variable.
43 // The method is declared in node.hpp.
44 const Node* Node::is_loop_iv() const {
45 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
46 this->as_Phi()->region()->is_CountedLoop() &&
47 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
48 return this;
49 } else {
50 return NULL;
51 }
52 }
54 //=============================================================================
55 //------------------------------dump_spec--------------------------------------
56 // Dump special per-node info
57 #ifndef PRODUCT
58 void LoopNode::dump_spec(outputStream *st) const {
59 if (is_inner_loop()) st->print( "inner " );
60 if (is_partial_peel_loop()) st->print( "partial_peel " );
61 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
62 }
63 #endif
65 //------------------------------is_valid_counted_loop-------------------------
66 bool LoopNode::is_valid_counted_loop() const {
67 if (is_CountedLoop()) {
68 CountedLoopNode* l = as_CountedLoop();
69 CountedLoopEndNode* le = l->loopexit();
70 if (le != NULL &&
71 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
72 Node* phi = l->phi();
73 Node* exit = le->proj_out(0 /* false */);
74 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
75 phi != NULL && phi->is_Phi() &&
76 phi->in(LoopNode::LoopBackControl) == l->incr() &&
77 le->loopnode() == l && le->stride_is_con()) {
78 return true;
79 }
80 }
81 }
82 return false;
83 }
85 //------------------------------get_early_ctrl---------------------------------
86 // Compute earliest legal control
87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
88 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
89 uint i;
90 Node *early;
91 if( n->in(0) ) {
92 early = n->in(0);
93 if( !early->is_CFG() ) // Might be a non-CFG multi-def
94 early = get_ctrl(early); // So treat input as a straight data input
95 i = 1;
96 } else {
97 early = get_ctrl(n->in(1));
98 i = 2;
99 }
100 uint e_d = dom_depth(early);
101 assert( early, "" );
102 for( ; i < n->req(); i++ ) {
103 Node *cin = get_ctrl(n->in(i));
104 assert( cin, "" );
105 // Keep deepest dominator depth
106 uint c_d = dom_depth(cin);
107 if( c_d > e_d ) { // Deeper guy?
108 early = cin; // Keep deepest found so far
109 e_d = c_d;
110 } else if( c_d == e_d && // Same depth?
111 early != cin ) { // If not equal, must use slower algorithm
112 // If same depth but not equal, one _must_ dominate the other
113 // and we want the deeper (i.e., dominated) guy.
114 Node *n1 = early;
115 Node *n2 = cin;
116 while( 1 ) {
117 n1 = idom(n1); // Walk up until break cycle
118 n2 = idom(n2);
119 if( n1 == cin || // Walked early up to cin
120 dom_depth(n2) < c_d )
121 break; // early is deeper; keep him
122 if( n2 == early || // Walked cin up to early
123 dom_depth(n1) < c_d ) {
124 early = cin; // cin is deeper; keep him
125 break;
126 }
127 }
128 e_d = dom_depth(early); // Reset depth register cache
129 }
130 }
132 // Return earliest legal location
133 assert(early == find_non_split_ctrl(early), "unexpected early control");
135 return early;
136 }
138 //------------------------------set_early_ctrl---------------------------------
139 // Set earliest legal control
140 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
141 Node *early = get_early_ctrl(n);
143 // Record earliest legal location
144 set_ctrl(n, early);
145 }
147 //------------------------------set_subtree_ctrl-------------------------------
148 // set missing _ctrl entries on new nodes
149 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
150 // Already set? Get out.
151 if( _nodes[n->_idx] ) return;
152 // Recursively set _nodes array to indicate where the Node goes
153 uint i;
154 for( i = 0; i < n->req(); ++i ) {
155 Node *m = n->in(i);
156 if( m && m != C->root() )
157 set_subtree_ctrl( m );
158 }
160 // Fixup self
161 set_early_ctrl( n );
162 }
164 //------------------------------is_counted_loop--------------------------------
165 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
166 PhaseGVN *gvn = &_igvn;
168 // Counted loop head must be a good RegionNode with only 3 not NULL
169 // control input edges: Self, Entry, LoopBack.
170 if (x->in(LoopNode::Self) == NULL || x->req() != 3)
171 return false;
173 Node *init_control = x->in(LoopNode::EntryControl);
174 Node *back_control = x->in(LoopNode::LoopBackControl);
175 if (init_control == NULL || back_control == NULL) // Partially dead
176 return false;
177 // Must also check for TOP when looking for a dead loop
178 if (init_control->is_top() || back_control->is_top())
179 return false;
181 // Allow funny placement of Safepoint
182 if (back_control->Opcode() == Op_SafePoint)
183 back_control = back_control->in(TypeFunc::Control);
185 // Controlling test for loop
186 Node *iftrue = back_control;
187 uint iftrue_op = iftrue->Opcode();
188 if (iftrue_op != Op_IfTrue &&
189 iftrue_op != Op_IfFalse)
190 // I have a weird back-control. Probably the loop-exit test is in
191 // the middle of the loop and I am looking at some trailing control-flow
192 // merge point. To fix this I would have to partially peel the loop.
193 return false; // Obscure back-control
195 // Get boolean guarding loop-back test
196 Node *iff = iftrue->in(0);
197 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
198 return false;
199 BoolNode *test = iff->in(1)->as_Bool();
200 BoolTest::mask bt = test->_test._test;
201 float cl_prob = iff->as_If()->_prob;
202 if (iftrue_op == Op_IfFalse) {
203 bt = BoolTest(bt).negate();
204 cl_prob = 1.0 - cl_prob;
205 }
206 // Get backedge compare
207 Node *cmp = test->in(1);
208 int cmp_op = cmp->Opcode();
209 if (cmp_op != Op_CmpI)
210 return false; // Avoid pointer & float compares
212 // Find the trip-counter increment & limit. Limit must be loop invariant.
213 Node *incr = cmp->in(1);
214 Node *limit = cmp->in(2);
216 // ---------
217 // need 'loop()' test to tell if limit is loop invariant
218 // ---------
220 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
221 Node *tmp = incr; // Then reverse order into the CmpI
222 incr = limit;
223 limit = tmp;
224 bt = BoolTest(bt).commute(); // And commute the exit test
225 }
226 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
227 return false;
228 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
229 return false;
231 Node* phi_incr = NULL;
232 // Trip-counter increment must be commutative & associative.
233 if (incr->is_Phi()) {
234 if (incr->as_Phi()->region() != x || incr->req() != 3)
235 return false; // Not simple trip counter expression
236 phi_incr = incr;
237 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
238 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
239 return false;
240 }
242 Node* trunc1 = NULL;
243 Node* trunc2 = NULL;
244 const TypeInt* iv_trunc_t = NULL;
245 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
246 return false; // Funny increment opcode
247 }
248 assert(incr->Opcode() == Op_AddI, "wrong increment code");
250 // Get merge point
251 Node *xphi = incr->in(1);
252 Node *stride = incr->in(2);
253 if (!stride->is_Con()) { // Oops, swap these
254 if (!xphi->is_Con()) // Is the other guy a constant?
255 return false; // Nope, unknown stride, bail out
256 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
257 xphi = stride;
258 stride = tmp;
259 }
260 // Stride must be constant
261 int stride_con = stride->get_int();
262 if (stride_con == 0)
263 return false; // missed some peephole opt
265 if (!xphi->is_Phi())
266 return false; // Too much math on the trip counter
267 if (phi_incr != NULL && phi_incr != xphi)
268 return false;
269 PhiNode *phi = xphi->as_Phi();
271 // Phi must be of loop header; backedge must wrap to increment
272 if (phi->region() != x)
273 return false;
274 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
275 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
276 return false;
277 }
278 Node *init_trip = phi->in(LoopNode::EntryControl);
280 // If iv trunc type is smaller than int, check for possible wrap.
281 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
282 assert(trunc1 != NULL, "must have found some truncation");
284 // Get a better type for the phi (filtered thru if's)
285 const TypeInt* phi_ft = filtered_type(phi);
287 // Can iv take on a value that will wrap?
288 //
289 // Ensure iv's limit is not within "stride" of the wrap value.
290 //
291 // Example for "short" type
292 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
293 // If the stride is +10, then the last value of the induction
294 // variable before the increment (phi_ft->_hi) must be
295 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
296 // ensure no truncation occurs after the increment.
298 if (stride_con > 0) {
299 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
300 iv_trunc_t->_lo > phi_ft->_lo) {
301 return false; // truncation may occur
302 }
303 } else if (stride_con < 0) {
304 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
305 iv_trunc_t->_hi < phi_ft->_hi) {
306 return false; // truncation may occur
307 }
308 }
309 // No possibility of wrap so truncation can be discarded
310 // Promote iv type to Int
311 } else {
312 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
313 }
315 // If the condition is inverted and we will be rolling
316 // through MININT to MAXINT, then bail out.
317 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
318 // Odd stride
319 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
320 // Count down loop rolls through MAXINT
321 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
322 // Count up loop rolls through MININT
323 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
324 return false; // Bail out
325 }
327 const TypeInt* init_t = gvn->type(init_trip)->is_int();
328 const TypeInt* limit_t = gvn->type(limit)->is_int();
330 if (stride_con > 0) {
331 long init_p = (long)init_t->_lo + stride_con;
332 if (init_p > (long)max_jint || init_p > (long)limit_t->_hi)
333 return false; // cyclic loop or this loop trips only once
334 } else {
335 long init_p = (long)init_t->_hi + stride_con;
336 if (init_p < (long)min_jint || init_p < (long)limit_t->_lo)
337 return false; // cyclic loop or this loop trips only once
338 }
340 // =================================================
341 // ---- SUCCESS! Found A Trip-Counted Loop! -----
342 //
343 assert(x->Opcode() == Op_Loop, "regular loops only");
344 C->print_method("Before CountedLoop", 3);
346 Node *hook = new (C, 6) Node(6);
348 if (LoopLimitCheck) {
350 // ===================================================
351 // Generate loop limit check to avoid integer overflow
352 // in cases like next (cyclic loops):
353 //
354 // for (i=0; i <= max_jint; i++) {}
355 // for (i=0; i < max_jint; i+=2) {}
356 //
357 //
358 // Limit check predicate depends on the loop test:
359 //
360 // for(;i != limit; i++) --> limit <= (max_jint)
361 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
362 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
363 //
365 // Check if limit is excluded to do more precise int overflow check.
366 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
367 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
369 // If compare points directly to the phi we need to adjust
370 // the compare so that it points to the incr. Limit have
371 // to be adjusted to keep trip count the same and the
372 // adjusted limit should be checked for int overflow.
373 if (phi_incr != NULL) {
374 stride_m += stride_con;
375 }
377 if (limit->is_Con()) {
378 int limit_con = limit->get_int();
379 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
380 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
381 // Bailout: it could be integer overflow.
382 return false;
383 }
384 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
385 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
386 // Limit's type may satisfy the condition, for example,
387 // when it is an array length.
388 } else {
389 // Generate loop's limit check.
390 // Loop limit check predicate should be near the loop.
391 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
392 if (!limit_check_proj) {
393 // The limit check predicate is not generated if this method trapped here before.
394 #ifdef ASSERT
395 if (TraceLoopLimitCheck) {
396 tty->print("missing loop limit check:");
397 loop->dump_head();
398 x->dump(1);
399 }
400 #endif
401 return false;
402 }
404 IfNode* check_iff = limit_check_proj->in(0)->as_If();
405 Node* cmp_limit;
406 Node* bol;
408 if (stride_con > 0) {
409 cmp_limit = new (C, 3) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
410 bol = new (C, 2) BoolNode(cmp_limit, BoolTest::le);
411 } else {
412 cmp_limit = new (C, 3) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
413 bol = new (C, 2) BoolNode(cmp_limit, BoolTest::ge);
414 }
415 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
416 bol = _igvn.register_new_node_with_optimizer(bol);
417 set_subtree_ctrl(bol);
419 // Replace condition in original predicate but preserve Opaque node
420 // so that previous predicates could be found.
421 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
422 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
423 Node* opq = check_iff->in(1)->in(1);
424 _igvn.hash_delete(opq);
425 opq->set_req(1, bol);
426 // Update ctrl.
427 set_ctrl(opq, check_iff->in(0));
428 set_ctrl(check_iff->in(1), check_iff->in(0));
430 #ifndef PRODUCT
431 // report that the loop predication has been actually performed
432 // for this loop
433 if (TraceLoopLimitCheck) {
434 tty->print_cr("Counted Loop Limit Check generated:");
435 debug_only( bol->dump(2); )
436 }
437 #endif
438 }
440 if (phi_incr != NULL) {
441 // If compare points directly to the phi we need to adjust
442 // the compare so that it points to the incr. Limit have
443 // to be adjusted to keep trip count the same and we
444 // should avoid int overflow.
445 //
446 // i = init; do {} while(i++ < limit);
447 // is converted to
448 // i = init; do {} while(++i < limit+1);
449 //
450 limit = gvn->transform(new (C, 3) AddINode(limit, stride));
451 }
453 // Now we need to canonicalize loop condition.
454 if (bt == BoolTest::ne) {
455 assert(stride_con == 1 || stride_con == -1, "simple increment only");
456 // 'ne' can be replaced with 'lt' only when init < limit.
457 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
458 bt = BoolTest::lt;
459 // 'ne' can be replaced with 'gt' only when init > limit.
460 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
461 bt = BoolTest::gt;
462 }
464 if (incl_limit) {
465 // The limit check guaranties that 'limit <= (max_jint - stride)' so
466 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
467 //
468 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
469 limit = gvn->transform(new (C, 3) AddINode(limit, one));
470 if (bt == BoolTest::le)
471 bt = BoolTest::lt;
472 else if (bt == BoolTest::ge)
473 bt = BoolTest::gt;
474 else
475 ShouldNotReachHere();
476 }
477 set_subtree_ctrl( limit );
479 } else { // LoopLimitCheck
481 // If compare points to incr, we are ok. Otherwise the compare
482 // can directly point to the phi; in this case adjust the compare so that
483 // it points to the incr by adjusting the limit.
484 if (cmp->in(1) == phi || cmp->in(2) == phi)
485 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
487 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
488 // Final value for iterator should be: trip_count * stride + init_trip.
489 Node *one_p = gvn->intcon( 1);
490 Node *one_m = gvn->intcon(-1);
492 Node *trip_count = NULL;
493 switch( bt ) {
494 case BoolTest::eq:
495 ShouldNotReachHere();
496 case BoolTest::ne: // Ahh, the case we desire
497 if (stride_con == 1)
498 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
499 else if (stride_con == -1)
500 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
501 else
502 ShouldNotReachHere();
503 set_subtree_ctrl(trip_count);
504 //_loop.map(trip_count->_idx,loop(limit));
505 break;
506 case BoolTest::le: // Maybe convert to '<' case
507 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
508 set_subtree_ctrl( limit );
509 hook->init_req(4, limit);
511 bt = BoolTest::lt;
512 // Make the new limit be in the same loop nest as the old limit
513 //_loop.map(limit->_idx,limit_loop);
514 // Fall into next case
515 case BoolTest::lt: { // Maybe convert to '!=' case
516 if (stride_con < 0) // Count down loop rolls through MAXINT
517 ShouldNotReachHere();
518 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
519 set_subtree_ctrl( range );
520 hook->init_req(0, range);
522 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
523 set_subtree_ctrl( bias );
524 hook->init_req(1, bias);
526 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
527 set_subtree_ctrl( bias1 );
528 hook->init_req(2, bias1);
530 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
531 set_subtree_ctrl( trip_count );
532 hook->init_req(3, trip_count);
533 break;
534 }
536 case BoolTest::ge: // Maybe convert to '>' case
537 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
538 set_subtree_ctrl( limit );
539 hook->init_req(4 ,limit);
541 bt = BoolTest::gt;
542 // Make the new limit be in the same loop nest as the old limit
543 //_loop.map(limit->_idx,limit_loop);
544 // Fall into next case
545 case BoolTest::gt: { // Maybe convert to '!=' case
546 if (stride_con > 0) // count up loop rolls through MININT
547 ShouldNotReachHere();
548 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
549 set_subtree_ctrl( range );
550 hook->init_req(0, range);
552 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
553 set_subtree_ctrl( bias );
554 hook->init_req(1, bias);
556 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
557 set_subtree_ctrl( bias1 );
558 hook->init_req(2, bias1);
560 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
561 set_subtree_ctrl( trip_count );
562 hook->init_req(3, trip_count);
563 break;
564 }
565 } // switch( bt )
567 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
568 set_subtree_ctrl( span );
569 hook->init_req(5, span);
571 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
572 set_subtree_ctrl( limit );
574 } // LoopLimitCheck
576 // Check for SafePoint on backedge and remove
577 Node *sfpt = x->in(LoopNode::LoopBackControl);
578 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
579 lazy_replace( sfpt, iftrue );
580 loop->_tail = iftrue;
581 }
583 // Build a canonical trip test.
584 // Clone code, as old values may be in use.
585 incr = incr->clone();
586 incr->set_req(1,phi);
587 incr->set_req(2,stride);
588 incr = _igvn.register_new_node_with_optimizer(incr);
589 set_early_ctrl( incr );
590 _igvn.hash_delete(phi);
591 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
593 // If phi type is more restrictive than Int, raise to
594 // Int to prevent (almost) infinite recursion in igvn
595 // which can only handle integer types for constants or minint..maxint.
596 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
597 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
598 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
599 nphi = _igvn.register_new_node_with_optimizer(nphi);
600 set_ctrl(nphi, get_ctrl(phi));
601 _igvn.replace_node(phi, nphi);
602 phi = nphi->as_Phi();
603 }
604 cmp = cmp->clone();
605 cmp->set_req(1,incr);
606 cmp->set_req(2,limit);
607 cmp = _igvn.register_new_node_with_optimizer(cmp);
608 set_ctrl(cmp, iff->in(0));
610 test = test->clone()->as_Bool();
611 (*(BoolTest*)&test->_test)._test = bt;
612 test->set_req(1,cmp);
613 _igvn.register_new_node_with_optimizer(test);
614 set_ctrl(test, iff->in(0));
616 // Replace the old IfNode with a new LoopEndNode
617 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
618 IfNode *le = lex->as_If();
619 uint dd = dom_depth(iff);
620 set_idom(le, le->in(0), dd); // Update dominance for loop exit
621 set_loop(le, loop);
623 // Get the loop-exit control
624 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
626 // Need to swap loop-exit and loop-back control?
627 if (iftrue_op == Op_IfFalse) {
628 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
629 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
631 loop->_tail = back_control = ift2;
632 set_loop(ift2, loop);
633 set_loop(iff2, get_loop(iffalse));
635 // Lazy update of 'get_ctrl' mechanism.
636 lazy_replace_proj( iffalse, iff2 );
637 lazy_replace_proj( iftrue, ift2 );
639 // Swap names
640 iffalse = iff2;
641 iftrue = ift2;
642 } else {
643 _igvn.hash_delete(iffalse);
644 _igvn.hash_delete(iftrue);
645 iffalse->set_req_X( 0, le, &_igvn );
646 iftrue ->set_req_X( 0, le, &_igvn );
647 }
649 set_idom(iftrue, le, dd+1);
650 set_idom(iffalse, le, dd+1);
651 assert(iff->outcnt() == 0, "should be dead now");
652 lazy_replace( iff, le ); // fix 'get_ctrl'
654 // Now setup a new CountedLoopNode to replace the existing LoopNode
655 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
656 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
657 // The following assert is approximately true, and defines the intention
658 // of can_be_counted_loop. It fails, however, because phase->type
659 // is not yet initialized for this loop and its parts.
660 //assert(l->can_be_counted_loop(this), "sanity");
661 _igvn.register_new_node_with_optimizer(l);
662 set_loop(l, loop);
663 loop->_head = l;
664 // Fix all data nodes placed at the old loop head.
665 // Uses the lazy-update mechanism of 'get_ctrl'.
666 lazy_replace( x, l );
667 set_idom(l, init_control, dom_depth(x));
669 // Check for immediately preceding SafePoint and remove
670 Node *sfpt2 = le->in(0);
671 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
672 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
674 // Free up intermediate goo
675 _igvn.remove_dead_node(hook);
677 #ifdef ASSERT
678 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
679 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
680 #endif
681 #ifndef PRODUCT
682 if (TraceLoopOpts) {
683 tty->print("Counted ");
684 loop->dump_head();
685 }
686 #endif
688 C->print_method("After CountedLoop", 3);
690 return true;
691 }
693 //----------------------exact_limit-------------------------------------------
694 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
695 assert(loop->_head->is_CountedLoop(), "");
696 CountedLoopNode *cl = loop->_head->as_CountedLoop();
697 assert(cl->is_valid_counted_loop(), "");
699 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
700 cl->limit()->Opcode() == Op_LoopLimit) {
701 // Old code has exact limit (it could be incorrect in case of int overflow).
702 // Loop limit is exact with stride == 1. And loop may already have exact limit.
703 return cl->limit();
704 }
705 Node *limit = NULL;
706 #ifdef ASSERT
707 BoolTest::mask bt = cl->loopexit()->test_trip();
708 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
709 #endif
710 if (cl->has_exact_trip_count()) {
711 // Simple case: loop has constant boundaries.
712 // Use longs to avoid integer overflow.
713 int stride_con = cl->stride_con();
714 long init_con = cl->init_trip()->get_int();
715 long limit_con = cl->limit()->get_int();
716 julong trip_cnt = cl->trip_count();
717 long final_con = init_con + trip_cnt*stride_con;
718 int final_int = (int)final_con;
719 // The final value should be in integer range since the loop
720 // is counted and the limit was checked for overflow.
721 assert(final_con == (long)final_int, "final value should be integer");
722 limit = _igvn.intcon(final_int);
723 } else {
724 // Create new LoopLimit node to get exact limit (final iv value).
725 limit = new (C, 4) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
726 register_new_node(limit, cl->in(LoopNode::EntryControl));
727 }
728 assert(limit != NULL, "sanity");
729 return limit;
730 }
732 //------------------------------Ideal------------------------------------------
733 // Return a node which is more "ideal" than the current node.
734 // Attempt to convert into a counted-loop.
735 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
736 if (!can_be_counted_loop(phase)) {
737 phase->C->set_major_progress();
738 }
739 return RegionNode::Ideal(phase, can_reshape);
740 }
743 //=============================================================================
744 //------------------------------Ideal------------------------------------------
745 // Return a node which is more "ideal" than the current node.
746 // Attempt to convert into a counted-loop.
747 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
748 return RegionNode::Ideal(phase, can_reshape);
749 }
751 //------------------------------dump_spec--------------------------------------
752 // Dump special per-node info
753 #ifndef PRODUCT
754 void CountedLoopNode::dump_spec(outputStream *st) const {
755 LoopNode::dump_spec(st);
756 if (stride_is_con()) {
757 st->print("stride: %d ",stride_con());
758 }
759 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
760 if (is_main_loop()) st->print("main of N%d", _idx);
761 if (is_post_loop()) st->print("post of N%d", _main_idx);
762 }
763 #endif
765 //=============================================================================
766 int CountedLoopEndNode::stride_con() const {
767 return stride()->bottom_type()->is_int()->get_con();
768 }
770 //=============================================================================
771 //------------------------------Value-----------------------------------------
772 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
773 const Type* init_t = phase->type(in(Init));
774 const Type* limit_t = phase->type(in(Limit));
775 const Type* stride_t = phase->type(in(Stride));
776 // Either input is TOP ==> the result is TOP
777 if (init_t == Type::TOP) return Type::TOP;
778 if (limit_t == Type::TOP) return Type::TOP;
779 if (stride_t == Type::TOP) return Type::TOP;
781 int stride_con = stride_t->is_int()->get_con();
782 if (stride_con == 1)
783 return NULL; // Identity
785 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
786 // Use longs to avoid integer overflow.
787 long init_con = init_t->is_int()->get_con();
788 long limit_con = limit_t->is_int()->get_con();
789 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
790 long trip_count = (limit_con - init_con + stride_m)/stride_con;
791 long final_con = init_con + stride_con*trip_count;
792 int final_int = (int)final_con;
793 // The final value should be in integer range since the loop
794 // is counted and the limit was checked for overflow.
795 assert(final_con == (long)final_int, "final value should be integer");
796 return TypeInt::make(final_int);
797 }
799 return bottom_type(); // TypeInt::INT
800 }
802 //------------------------------Ideal------------------------------------------
803 // Return a node which is more "ideal" than the current node.
804 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
805 if (phase->type(in(Init)) == Type::TOP ||
806 phase->type(in(Limit)) == Type::TOP ||
807 phase->type(in(Stride)) == Type::TOP)
808 return NULL; // Dead
810 int stride_con = phase->type(in(Stride))->is_int()->get_con();
811 if (stride_con == 1)
812 return NULL; // Identity
814 if (in(Init)->is_Con() && in(Limit)->is_Con())
815 return NULL; // Value
817 // Delay following optimizations until all loop optimizations
818 // done to keep Ideal graph simple.
819 if (!can_reshape || phase->C->major_progress())
820 return NULL;
822 const TypeInt* init_t = phase->type(in(Init) )->is_int();
823 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
824 int stride_p;
825 long lim, ini;
826 julong max;
827 if (stride_con > 0) {
828 stride_p = stride_con;
829 lim = limit_t->_hi;
830 ini = init_t->_lo;
831 max = (julong)max_jint;
832 } else {
833 stride_p = -stride_con;
834 lim = init_t->_hi;
835 ini = limit_t->_lo;
836 max = (julong)min_jint;
837 }
838 julong range = lim - ini + stride_p;
839 if (range <= max) {
840 // Convert to integer expression if it is not overflow.
841 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
842 Node *range = phase->transform(new (phase->C, 3) SubINode(in(Limit), in(Init)));
843 Node *bias = phase->transform(new (phase->C, 3) AddINode(range, stride_m));
844 Node *trip = phase->transform(new (phase->C, 3) DivINode(0, bias, in(Stride)));
845 Node *span = phase->transform(new (phase->C, 3) MulINode(trip, in(Stride)));
846 return new (phase->C, 3) AddINode(span, in(Init)); // exact limit
847 }
849 if (is_power_of_2(stride_p) || // divisor is 2^n
850 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
851 // Convert to long expression to avoid integer overflow
852 // and let igvn optimizer convert this division.
853 //
854 Node* init = phase->transform( new (phase->C, 2) ConvI2LNode(in(Init)));
855 Node* limit = phase->transform( new (phase->C, 2) ConvI2LNode(in(Limit)));
856 Node* stride = phase->longcon(stride_con);
857 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
859 Node *range = phase->transform(new (phase->C, 3) SubLNode(limit, init));
860 Node *bias = phase->transform(new (phase->C, 3) AddLNode(range, stride_m));
861 Node *span;
862 if (stride_con > 0 && is_power_of_2(stride_p)) {
863 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
864 // and avoid generating rounding for division. Zero trip guard should
865 // guarantee that init < limit but sometimes the guard is missing and
866 // we can get situation when init > limit. Note, for the empty loop
867 // optimization zero trip guard is generated explicitly which leaves
868 // only RCE predicate where exact limit is used and the predicate
869 // will simply fail forcing recompilation.
870 Node* neg_stride = phase->longcon(-stride_con);
871 span = phase->transform(new (phase->C, 3) AndLNode(bias, neg_stride));
872 } else {
873 Node *trip = phase->transform(new (phase->C, 3) DivLNode(0, bias, stride));
874 span = phase->transform(new (phase->C, 3) MulLNode(trip, stride));
875 }
876 // Convert back to int
877 Node *span_int = phase->transform(new (phase->C, 2) ConvL2INode(span));
878 return new (phase->C, 3) AddINode(span_int, in(Init)); // exact limit
879 }
881 return NULL; // No progress
882 }
884 //------------------------------Identity---------------------------------------
885 // If stride == 1 return limit node.
886 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
887 int stride_con = phase->type(in(Stride))->is_int()->get_con();
888 if (stride_con == 1 || stride_con == -1)
889 return in(Limit);
890 return this;
891 }
893 //=============================================================================
894 //----------------------match_incr_with_optional_truncation--------------------
895 // Match increment with optional truncation:
896 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
897 // Return NULL for failure. Success returns the increment node.
898 Node* CountedLoopNode::match_incr_with_optional_truncation(
899 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
900 // Quick cutouts:
901 if (expr == NULL || expr->req() != 3) return NULL;
903 Node *t1 = NULL;
904 Node *t2 = NULL;
905 const TypeInt* trunc_t = TypeInt::INT;
906 Node* n1 = expr;
907 int n1op = n1->Opcode();
909 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
910 if (n1op == Op_AndI &&
911 n1->in(2)->is_Con() &&
912 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
913 // %%% This check should match any mask of 2**K-1.
914 t1 = n1;
915 n1 = t1->in(1);
916 n1op = n1->Opcode();
917 trunc_t = TypeInt::CHAR;
918 } else if (n1op == Op_RShiftI &&
919 n1->in(1) != NULL &&
920 n1->in(1)->Opcode() == Op_LShiftI &&
921 n1->in(2) == n1->in(1)->in(2) &&
922 n1->in(2)->is_Con()) {
923 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
924 // %%% This check should match any shift in [1..31].
925 if (shift == 16 || shift == 8) {
926 t1 = n1;
927 t2 = t1->in(1);
928 n1 = t2->in(1);
929 n1op = n1->Opcode();
930 if (shift == 16) {
931 trunc_t = TypeInt::SHORT;
932 } else if (shift == 8) {
933 trunc_t = TypeInt::BYTE;
934 }
935 }
936 }
938 // If (maybe after stripping) it is an AddI, we won:
939 if (n1op == Op_AddI) {
940 *trunc1 = t1;
941 *trunc2 = t2;
942 *trunc_type = trunc_t;
943 return n1;
944 }
946 // failed
947 return NULL;
948 }
951 //------------------------------filtered_type--------------------------------
952 // Return a type based on condition control flow
953 // A successful return will be a type that is restricted due
954 // to a series of dominating if-tests, such as:
955 // if (i < 10) {
956 // if (i > 0) {
957 // here: "i" type is [1..10)
958 // }
959 // }
960 // or a control flow merge
961 // if (i < 10) {
962 // do {
963 // phi( , ) -- at top of loop type is [min_int..10)
964 // i = ?
965 // } while ( i < 10)
966 //
967 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
968 assert(n && n->bottom_type()->is_int(), "must be int");
969 const TypeInt* filtered_t = NULL;
970 if (!n->is_Phi()) {
971 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
972 filtered_t = filtered_type_from_dominators(n, n_ctrl);
974 } else {
975 Node* phi = n->as_Phi();
976 Node* region = phi->in(0);
977 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
978 if (region && region != C->top()) {
979 for (uint i = 1; i < phi->req(); i++) {
980 Node* val = phi->in(i);
981 Node* use_c = region->in(i);
982 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
983 if (val_t != NULL) {
984 if (filtered_t == NULL) {
985 filtered_t = val_t;
986 } else {
987 filtered_t = filtered_t->meet(val_t)->is_int();
988 }
989 }
990 }
991 }
992 }
993 const TypeInt* n_t = _igvn.type(n)->is_int();
994 if (filtered_t != NULL) {
995 n_t = n_t->join(filtered_t)->is_int();
996 }
997 return n_t;
998 }
1001 //------------------------------filtered_type_from_dominators--------------------------------
1002 // Return a possibly more restrictive type for val based on condition control flow of dominators
1003 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1004 if (val->is_Con()) {
1005 return val->bottom_type()->is_int();
1006 }
1007 uint if_limit = 10; // Max number of dominating if's visited
1008 const TypeInt* rtn_t = NULL;
1010 if (use_ctrl && use_ctrl != C->top()) {
1011 Node* val_ctrl = get_ctrl(val);
1012 uint val_dom_depth = dom_depth(val_ctrl);
1013 Node* pred = use_ctrl;
1014 uint if_cnt = 0;
1015 while (if_cnt < if_limit) {
1016 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1017 if_cnt++;
1018 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1019 if (if_t != NULL) {
1020 if (rtn_t == NULL) {
1021 rtn_t = if_t;
1022 } else {
1023 rtn_t = rtn_t->join(if_t)->is_int();
1024 }
1025 }
1026 }
1027 pred = idom(pred);
1028 if (pred == NULL || pred == C->top()) {
1029 break;
1030 }
1031 // Stop if going beyond definition block of val
1032 if (dom_depth(pred) < val_dom_depth) {
1033 break;
1034 }
1035 }
1036 }
1037 return rtn_t;
1038 }
1041 //------------------------------dump_spec--------------------------------------
1042 // Dump special per-node info
1043 #ifndef PRODUCT
1044 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1045 if( in(TestValue)->is_Bool() ) {
1046 BoolTest bt( test_trip()); // Added this for g++.
1048 st->print("[");
1049 bt.dump_on(st);
1050 st->print("]");
1051 }
1052 st->print(" ");
1053 IfNode::dump_spec(st);
1054 }
1055 #endif
1057 //=============================================================================
1058 //------------------------------is_member--------------------------------------
1059 // Is 'l' a member of 'this'?
1060 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1061 while( l->_nest > _nest ) l = l->_parent;
1062 return l == this;
1063 }
1065 //------------------------------set_nest---------------------------------------
1066 // Set loop tree nesting depth. Accumulate _has_call bits.
1067 int IdealLoopTree::set_nest( uint depth ) {
1068 _nest = depth;
1069 int bits = _has_call;
1070 if( _child ) bits |= _child->set_nest(depth+1);
1071 if( bits ) _has_call = 1;
1072 if( _next ) bits |= _next ->set_nest(depth );
1073 return bits;
1074 }
1076 //------------------------------split_fall_in----------------------------------
1077 // Split out multiple fall-in edges from the loop header. Move them to a
1078 // private RegionNode before the loop. This becomes the loop landing pad.
1079 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1080 PhaseIterGVN &igvn = phase->_igvn;
1081 uint i;
1083 // Make a new RegionNode to be the landing pad.
1084 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
1085 phase->set_loop(landing_pad,_parent);
1086 // Gather all the fall-in control paths into the landing pad
1087 uint icnt = fall_in_cnt;
1088 uint oreq = _head->req();
1089 for( i = oreq-1; i>0; i-- )
1090 if( !phase->is_member( this, _head->in(i) ) )
1091 landing_pad->set_req(icnt--,_head->in(i));
1093 // Peel off PhiNode edges as well
1094 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1095 Node *oj = _head->fast_out(j);
1096 if( oj->is_Phi() ) {
1097 PhiNode* old_phi = oj->as_Phi();
1098 assert( old_phi->region() == _head, "" );
1099 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1100 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1101 uint icnt = fall_in_cnt;
1102 for( i = oreq-1; i>0; i-- ) {
1103 if( !phase->is_member( this, _head->in(i) ) ) {
1104 p->init_req(icnt--, old_phi->in(i));
1105 // Go ahead and clean out old edges from old phi
1106 old_phi->del_req(i);
1107 }
1108 }
1109 // Search for CSE's here, because ZKM.jar does a lot of
1110 // loop hackery and we need to be a little incremental
1111 // with the CSE to avoid O(N^2) node blow-up.
1112 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1113 if( p2 ) { // Found CSE
1114 p->destruct(); // Recover useless new node
1115 p = p2; // Use old node
1116 } else {
1117 igvn.register_new_node_with_optimizer(p, old_phi);
1118 }
1119 // Make old Phi refer to new Phi.
1120 old_phi->add_req(p);
1121 // Check for the special case of making the old phi useless and
1122 // disappear it. In JavaGrande I have a case where this useless
1123 // Phi is the loop limit and prevents recognizing a CountedLoop
1124 // which in turn prevents removing an empty loop.
1125 Node *id_old_phi = old_phi->Identity( &igvn );
1126 if( id_old_phi != old_phi ) { // Found a simple identity?
1127 // Note that I cannot call 'replace_node' here, because
1128 // that will yank the edge from old_phi to the Region and
1129 // I'm mid-iteration over the Region's uses.
1130 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1131 Node* use = old_phi->last_out(i);
1132 igvn.rehash_node_delayed(use);
1133 uint uses_found = 0;
1134 for (uint j = 0; j < use->len(); j++) {
1135 if (use->in(j) == old_phi) {
1136 if (j < use->req()) use->set_req (j, id_old_phi);
1137 else use->set_prec(j, id_old_phi);
1138 uses_found++;
1139 }
1140 }
1141 i -= uses_found; // we deleted 1 or more copies of this edge
1142 }
1143 }
1144 igvn._worklist.push(old_phi);
1145 }
1146 }
1147 // Finally clean out the fall-in edges from the RegionNode
1148 for( i = oreq-1; i>0; i-- ) {
1149 if( !phase->is_member( this, _head->in(i) ) ) {
1150 _head->del_req(i);
1151 }
1152 }
1153 // Transform landing pad
1154 igvn.register_new_node_with_optimizer(landing_pad, _head);
1155 // Insert landing pad into the header
1156 _head->add_req(landing_pad);
1157 }
1159 //------------------------------split_outer_loop-------------------------------
1160 // Split out the outermost loop from this shared header.
1161 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1162 PhaseIterGVN &igvn = phase->_igvn;
1164 // Find index of outermost loop; it should also be my tail.
1165 uint outer_idx = 1;
1166 while( _head->in(outer_idx) != _tail ) outer_idx++;
1168 // Make a LoopNode for the outermost loop.
1169 Node *ctl = _head->in(LoopNode::EntryControl);
1170 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
1171 outer = igvn.register_new_node_with_optimizer(outer, _head);
1172 phase->set_created_loop_node();
1174 // Outermost loop falls into '_head' loop
1175 _head->set_req(LoopNode::EntryControl, outer);
1176 _head->del_req(outer_idx);
1177 // Split all the Phis up between '_head' loop and 'outer' loop.
1178 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1179 Node *out = _head->fast_out(j);
1180 if( out->is_Phi() ) {
1181 PhiNode *old_phi = out->as_Phi();
1182 assert( old_phi->region() == _head, "" );
1183 Node *phi = PhiNode::make_blank(outer, old_phi);
1184 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1185 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1186 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1187 // Make old Phi point to new Phi on the fall-in path
1188 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1189 old_phi->del_req(outer_idx);
1190 }
1191 }
1193 // Use the new loop head instead of the old shared one
1194 _head = outer;
1195 phase->set_loop(_head, this);
1196 }
1198 //------------------------------fix_parent-------------------------------------
1199 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1200 loop->_parent = parent;
1201 if( loop->_child ) fix_parent( loop->_child, loop );
1202 if( loop->_next ) fix_parent( loop->_next , parent );
1203 }
1205 //------------------------------estimate_path_freq-----------------------------
1206 static float estimate_path_freq( Node *n ) {
1207 // Try to extract some path frequency info
1208 IfNode *iff;
1209 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1210 uint nop = n->Opcode();
1211 if( nop == Op_SafePoint ) { // Skip any safepoint
1212 n = n->in(0);
1213 continue;
1214 }
1215 if( nop == Op_CatchProj ) { // Get count from a prior call
1216 // Assume call does not always throw exceptions: means the call-site
1217 // count is also the frequency of the fall-through path.
1218 assert( n->is_CatchProj(), "" );
1219 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1220 return 0.0f; // Assume call exception path is rare
1221 Node *call = n->in(0)->in(0)->in(0);
1222 assert( call->is_Call(), "expect a call here" );
1223 const JVMState *jvms = ((CallNode*)call)->jvms();
1224 ciMethodData* methodData = jvms->method()->method_data();
1225 if (!methodData->is_mature()) return 0.0f; // No call-site data
1226 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1227 if ((data == NULL) || !data->is_CounterData()) {
1228 // no call profile available, try call's control input
1229 n = n->in(0);
1230 continue;
1231 }
1232 return data->as_CounterData()->count()/FreqCountInvocations;
1233 }
1234 // See if there's a gating IF test
1235 Node *n_c = n->in(0);
1236 if( !n_c->is_If() ) break; // No estimate available
1237 iff = n_c->as_If();
1238 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1239 // Compute how much count comes on this path
1240 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1241 // Have no count info. Skip dull uncommon-trap like branches.
1242 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1243 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1244 break;
1245 // Skip through never-taken branch; look for a real loop exit.
1246 n = iff->in(0);
1247 }
1248 return 0.0f; // No estimate available
1249 }
1251 //------------------------------merge_many_backedges---------------------------
1252 // Merge all the backedges from the shared header into a private Region.
1253 // Feed that region as the one backedge to this loop.
1254 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1255 uint i;
1257 // Scan for the top 2 hottest backedges
1258 float hotcnt = 0.0f;
1259 float warmcnt = 0.0f;
1260 uint hot_idx = 0;
1261 // Loop starts at 2 because slot 1 is the fall-in path
1262 for( i = 2; i < _head->req(); i++ ) {
1263 float cnt = estimate_path_freq(_head->in(i));
1264 if( cnt > hotcnt ) { // Grab hottest path
1265 warmcnt = hotcnt;
1266 hotcnt = cnt;
1267 hot_idx = i;
1268 } else if( cnt > warmcnt ) { // And 2nd hottest path
1269 warmcnt = cnt;
1270 }
1271 }
1273 // See if the hottest backedge is worthy of being an inner loop
1274 // by being much hotter than the next hottest backedge.
1275 if( hotcnt <= 0.0001 ||
1276 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1278 // Peel out the backedges into a private merge point; peel
1279 // them all except optionally hot_idx.
1280 PhaseIterGVN &igvn = phase->_igvn;
1282 Node *hot_tail = NULL;
1283 // Make a Region for the merge point
1284 Node *r = new (phase->C, 1) RegionNode(1);
1285 for( i = 2; i < _head->req(); i++ ) {
1286 if( i != hot_idx )
1287 r->add_req( _head->in(i) );
1288 else hot_tail = _head->in(i);
1289 }
1290 igvn.register_new_node_with_optimizer(r, _head);
1291 // Plug region into end of loop _head, followed by hot_tail
1292 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1293 _head->set_req(2, r);
1294 if( hot_idx ) _head->add_req(hot_tail);
1296 // Split all the Phis up between '_head' loop and the Region 'r'
1297 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1298 Node *out = _head->fast_out(j);
1299 if( out->is_Phi() ) {
1300 PhiNode* n = out->as_Phi();
1301 igvn.hash_delete(n); // Delete from hash before hacking edges
1302 Node *hot_phi = NULL;
1303 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
1304 // Check all inputs for the ones to peel out
1305 uint j = 1;
1306 for( uint i = 2; i < n->req(); i++ ) {
1307 if( i != hot_idx )
1308 phi->set_req( j++, n->in(i) );
1309 else hot_phi = n->in(i);
1310 }
1311 // Register the phi but do not transform until whole place transforms
1312 igvn.register_new_node_with_optimizer(phi, n);
1313 // Add the merge phi to the old Phi
1314 while( n->req() > 3 ) n->del_req( n->req()-1 );
1315 n->set_req(2, phi);
1316 if( hot_idx ) n->add_req(hot_phi);
1317 }
1318 }
1321 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1322 // of self loop tree. Turn self into a loop headed by _head and with
1323 // tail being the new merge point.
1324 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1325 phase->set_loop(_tail,ilt); // Adjust tail
1326 _tail = r; // Self's tail is new merge point
1327 phase->set_loop(r,this);
1328 ilt->_child = _child; // New guy has my children
1329 _child = ilt; // Self has new guy as only child
1330 ilt->_parent = this; // new guy has self for parent
1331 ilt->_nest = _nest; // Same nesting depth (for now)
1333 // Starting with 'ilt', look for child loop trees using the same shared
1334 // header. Flatten these out; they will no longer be loops in the end.
1335 IdealLoopTree **pilt = &_child;
1336 while( ilt ) {
1337 if( ilt->_head == _head ) {
1338 uint i;
1339 for( i = 2; i < _head->req(); i++ )
1340 if( _head->in(i) == ilt->_tail )
1341 break; // Still a loop
1342 if( i == _head->req() ) { // No longer a loop
1343 // Flatten ilt. Hang ilt's "_next" list from the end of
1344 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1345 IdealLoopTree **cp = &ilt->_child;
1346 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1347 *cp = ilt->_next; // Hang next list at end of child list
1348 *pilt = ilt->_child; // Move child up to replace ilt
1349 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1350 ilt = ilt->_child; // Repeat using new ilt
1351 continue; // do not advance over ilt->_child
1352 }
1353 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1354 phase->set_loop(_head,ilt);
1355 }
1356 pilt = &ilt->_child; // Advance to next
1357 ilt = *pilt;
1358 }
1360 if( _child ) fix_parent( _child, this );
1361 }
1363 //------------------------------beautify_loops---------------------------------
1364 // Split shared headers and insert loop landing pads.
1365 // Insert a LoopNode to replace the RegionNode.
1366 // Return TRUE if loop tree is structurally changed.
1367 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1368 bool result = false;
1369 // Cache parts in locals for easy
1370 PhaseIterGVN &igvn = phase->_igvn;
1372 igvn.hash_delete(_head); // Yank from hash before hacking edges
1374 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1375 int fall_in_cnt = 0;
1376 for( uint i = 1; i < _head->req(); i++ )
1377 if( !phase->is_member( this, _head->in(i) ) )
1378 fall_in_cnt++;
1379 assert( fall_in_cnt, "at least 1 fall-in path" );
1380 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1381 split_fall_in( phase, fall_in_cnt );
1383 // Swap inputs to the _head and all Phis to move the fall-in edge to
1384 // the left.
1385 fall_in_cnt = 1;
1386 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1387 fall_in_cnt++;
1388 if( fall_in_cnt > 1 ) {
1389 // Since I am just swapping inputs I do not need to update def-use info
1390 Node *tmp = _head->in(1);
1391 _head->set_req( 1, _head->in(fall_in_cnt) );
1392 _head->set_req( fall_in_cnt, tmp );
1393 // Swap also all Phis
1394 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1395 Node* phi = _head->fast_out(i);
1396 if( phi->is_Phi() ) {
1397 igvn.hash_delete(phi); // Yank from hash before hacking edges
1398 tmp = phi->in(1);
1399 phi->set_req( 1, phi->in(fall_in_cnt) );
1400 phi->set_req( fall_in_cnt, tmp );
1401 }
1402 }
1403 }
1404 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1405 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1407 // If I am a shared header (multiple backedges), peel off the many
1408 // backedges into a private merge point and use the merge point as
1409 // the one true backedge.
1410 if( _head->req() > 3 ) {
1411 // Merge the many backedges into a single backedge but leave
1412 // the hottest backedge as separate edge for the following peel.
1413 merge_many_backedges( phase );
1414 result = true;
1415 }
1417 // If I have one hot backedge, peel off myself loop.
1418 // I better be the outermost loop.
1419 if( _head->req() > 3 ) {
1420 split_outer_loop( phase );
1421 result = true;
1423 } else if( !_head->is_Loop() && !_irreducible ) {
1424 // Make a new LoopNode to replace the old loop head
1425 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1426 l = igvn.register_new_node_with_optimizer(l, _head);
1427 phase->set_created_loop_node();
1428 // Go ahead and replace _head
1429 phase->_igvn.replace_node( _head, l );
1430 _head = l;
1431 phase->set_loop(_head, this);
1432 }
1434 // Now recursively beautify nested loops
1435 if( _child ) result |= _child->beautify_loops( phase );
1436 if( _next ) result |= _next ->beautify_loops( phase );
1437 return result;
1438 }
1440 //------------------------------allpaths_check_safepts----------------------------
1441 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1442 // encountered. Helper for check_safepts.
1443 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1444 assert(stack.size() == 0, "empty stack");
1445 stack.push(_tail);
1446 visited.Clear();
1447 visited.set(_tail->_idx);
1448 while (stack.size() > 0) {
1449 Node* n = stack.pop();
1450 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1451 // Terminate this path
1452 } else if (n->Opcode() == Op_SafePoint) {
1453 if (_phase->get_loop(n) != this) {
1454 if (_required_safept == NULL) _required_safept = new Node_List();
1455 _required_safept->push(n); // save the one closest to the tail
1456 }
1457 // Terminate this path
1458 } else {
1459 uint start = n->is_Region() ? 1 : 0;
1460 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1461 for (uint i = start; i < end; i++) {
1462 Node* in = n->in(i);
1463 assert(in->is_CFG(), "must be");
1464 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1465 stack.push(in);
1466 }
1467 }
1468 }
1469 }
1470 }
1472 //------------------------------check_safepts----------------------------
1473 // Given dominators, try to find loops with calls that must always be
1474 // executed (call dominates loop tail). These loops do not need non-call
1475 // safepoints (ncsfpt).
1476 //
1477 // A complication is that a safepoint in a inner loop may be needed
1478 // by an outer loop. In the following, the inner loop sees it has a
1479 // call (block 3) on every path from the head (block 2) to the
1480 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1481 // in block 2, _but_ this leaves the outer loop without a safepoint.
1482 //
1483 // entry 0
1484 // |
1485 // v
1486 // outer 1,2 +->1
1487 // | |
1488 // | v
1489 // | 2<---+ ncsfpt in 2
1490 // |_/|\ |
1491 // | v |
1492 // inner 2,3 / 3 | call in 3
1493 // / | |
1494 // v +--+
1495 // exit 4
1496 //
1497 //
1498 // This method creates a list (_required_safept) of ncsfpt nodes that must
1499 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1500 // is first looked for in the lists for the outer loops of the current loop.
1501 //
1502 // The insights into the problem:
1503 // A) counted loops are okay
1504 // B) innermost loops are okay (only an inner loop can delete
1505 // a ncsfpt needed by an outer loop)
1506 // C) a loop is immune from an inner loop deleting a safepoint
1507 // if the loop has a call on the idom-path
1508 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1509 // idom-path that is not in a nested loop
1510 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1511 // loop needs to be prevented from deletion by an inner loop
1512 //
1513 // There are two analyses:
1514 // 1) The first, and cheaper one, scans the loop body from
1515 // tail to head following the idom (immediate dominator)
1516 // chain, looking for the cases (C,D,E) above.
1517 // Since inner loops are scanned before outer loops, there is summary
1518 // information about inner loops. Inner loops can be skipped over
1519 // when the tail of an inner loop is encountered.
1520 //
1521 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1522 // the idom path (which is rare), scans all predecessor control paths
1523 // from the tail to the head, terminating a path when a call or sfpt
1524 // is encountered, to find the ncsfpt's that are closest to the tail.
1525 //
1526 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1527 // Bottom up traversal
1528 IdealLoopTree* ch = _child;
1529 while (ch != NULL) {
1530 ch->check_safepts(visited, stack);
1531 ch = ch->_next;
1532 }
1534 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1535 bool has_call = false; // call on dom-path
1536 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1537 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1538 // Scan the dom-path nodes from tail to head
1539 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1540 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1541 has_call = true;
1542 _has_sfpt = 1; // Then no need for a safept!
1543 break;
1544 } else if (n->Opcode() == Op_SafePoint) {
1545 if (_phase->get_loop(n) == this) {
1546 has_local_ncsfpt = true;
1547 break;
1548 }
1549 if (nonlocal_ncsfpt == NULL) {
1550 nonlocal_ncsfpt = n; // save the one closest to the tail
1551 }
1552 } else {
1553 IdealLoopTree* nlpt = _phase->get_loop(n);
1554 if (this != nlpt) {
1555 // If at an inner loop tail, see if the inner loop has already
1556 // recorded seeing a call on the dom-path (and stop.) If not,
1557 // jump to the head of the inner loop.
1558 assert(is_member(nlpt), "nested loop");
1559 Node* tail = nlpt->_tail;
1560 if (tail->in(0)->is_If()) tail = tail->in(0);
1561 if (n == tail) {
1562 // If inner loop has call on dom-path, so does outer loop
1563 if (nlpt->_has_sfpt) {
1564 has_call = true;
1565 _has_sfpt = 1;
1566 break;
1567 }
1568 // Skip to head of inner loop
1569 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1570 n = nlpt->_head;
1571 }
1572 }
1573 }
1574 }
1575 // Record safept's that this loop needs preserved when an
1576 // inner loop attempts to delete it's safepoints.
1577 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1578 if (nonlocal_ncsfpt != NULL) {
1579 if (_required_safept == NULL) _required_safept = new Node_List();
1580 _required_safept->push(nonlocal_ncsfpt);
1581 } else {
1582 // Failed to find a suitable safept on the dom-path. Now use
1583 // an all paths walk from tail to head, looking for safepoints to preserve.
1584 allpaths_check_safepts(visited, stack);
1585 }
1586 }
1587 }
1588 }
1590 //---------------------------is_deleteable_safept----------------------------
1591 // Is safept not required by an outer loop?
1592 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1593 assert(sfpt->Opcode() == Op_SafePoint, "");
1594 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1595 while (lp != NULL) {
1596 Node_List* sfpts = lp->_required_safept;
1597 if (sfpts != NULL) {
1598 for (uint i = 0; i < sfpts->size(); i++) {
1599 if (sfpt == sfpts->at(i))
1600 return false;
1601 }
1602 }
1603 lp = lp->_parent;
1604 }
1605 return true;
1606 }
1608 //---------------------------replace_parallel_iv-------------------------------
1609 // Replace parallel induction variable (parallel to trip counter)
1610 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1611 assert(loop->_head->is_CountedLoop(), "");
1612 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1613 if (!cl->is_valid_counted_loop())
1614 return; // skip malformed counted loop
1615 Node *incr = cl->incr();
1616 if (incr == NULL)
1617 return; // Dead loop?
1618 Node *init = cl->init_trip();
1619 Node *phi = cl->phi();
1620 int stride_con = cl->stride_con();
1622 // Visit all children, looking for Phis
1623 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1624 Node *out = cl->out(i);
1625 // Look for other phis (secondary IVs). Skip dead ones
1626 if (!out->is_Phi() || out == phi || !has_node(out))
1627 continue;
1628 PhiNode* phi2 = out->as_Phi();
1629 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1630 // Look for induction variables of the form: X += constant
1631 if (phi2->region() != loop->_head ||
1632 incr2->req() != 3 ||
1633 incr2->in(1) != phi2 ||
1634 incr2 == incr ||
1635 incr2->Opcode() != Op_AddI ||
1636 !incr2->in(2)->is_Con())
1637 continue;
1639 // Check for parallel induction variable (parallel to trip counter)
1640 // via an affine function. In particular, count-down loops with
1641 // count-up array indices are common. We only RCE references off
1642 // the trip-counter, so we need to convert all these to trip-counter
1643 // expressions.
1644 Node *init2 = phi2->in( LoopNode::EntryControl );
1645 int stride_con2 = incr2->in(2)->get_int();
1647 // The general case here gets a little tricky. We want to find the
1648 // GCD of all possible parallel IV's and make a new IV using this
1649 // GCD for the loop. Then all possible IVs are simple multiples of
1650 // the GCD. In practice, this will cover very few extra loops.
1651 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1652 // where +/-1 is the common case, but other integer multiples are
1653 // also easy to handle.
1654 int ratio_con = stride_con2/stride_con;
1656 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1657 #ifndef PRODUCT
1658 if (TraceLoopOpts) {
1659 tty->print("Parallel IV: %d ", phi2->_idx);
1660 loop->dump_head();
1661 }
1662 #endif
1663 // Convert to using the trip counter. The parallel induction
1664 // variable differs from the trip counter by a loop-invariant
1665 // amount, the difference between their respective initial values.
1666 // It is scaled by the 'ratio_con'.
1667 Node* ratio = _igvn.intcon(ratio_con);
1668 set_ctrl(ratio, C->root());
1669 Node* ratio_init = new (C, 3) MulINode(init, ratio);
1670 _igvn.register_new_node_with_optimizer(ratio_init, init);
1671 set_early_ctrl(ratio_init);
1672 Node* diff = new (C, 3) SubINode(init2, ratio_init);
1673 _igvn.register_new_node_with_optimizer(diff, init2);
1674 set_early_ctrl(diff);
1675 Node* ratio_idx = new (C, 3) MulINode(phi, ratio);
1676 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1677 set_ctrl(ratio_idx, cl);
1678 Node* add = new (C, 3) AddINode(ratio_idx, diff);
1679 _igvn.register_new_node_with_optimizer(add);
1680 set_ctrl(add, cl);
1681 _igvn.replace_node( phi2, add );
1682 // Sometimes an induction variable is unused
1683 if (add->outcnt() == 0) {
1684 _igvn.remove_dead_node(add);
1685 }
1686 --i; // deleted this phi; rescan starting with next position
1687 continue;
1688 }
1689 }
1690 }
1692 //------------------------------counted_loop-----------------------------------
1693 // Convert to counted loops where possible
1694 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1696 // For grins, set the inner-loop flag here
1697 if (!_child) {
1698 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1699 }
1701 if (_head->is_CountedLoop() ||
1702 phase->is_counted_loop(_head, this)) {
1703 _has_sfpt = 1; // Indicate we do not need a safepoint here
1705 // Look for a safepoint to remove
1706 for (Node* n = tail(); n != _head; n = phase->idom(n))
1707 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1708 phase->is_deleteable_safept(n))
1709 phase->lazy_replace(n,n->in(TypeFunc::Control));
1711 // Look for induction variables
1712 phase->replace_parallel_iv(this);
1714 } else if (_parent != NULL && !_irreducible) {
1715 // Not a counted loop.
1716 // Look for a safepoint on the idom-path to remove, preserving the first one
1717 bool found = false;
1718 Node* n = tail();
1719 for (; n != _head && !found; n = phase->idom(n)) {
1720 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1721 found = true; // Found one
1722 }
1723 // Skip past it and delete the others
1724 for (; n != _head; n = phase->idom(n)) {
1725 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1726 phase->is_deleteable_safept(n))
1727 phase->lazy_replace(n,n->in(TypeFunc::Control));
1728 }
1729 }
1731 // Recursively
1732 if (_child) _child->counted_loop( phase );
1733 if (_next) _next ->counted_loop( phase );
1734 }
1736 #ifndef PRODUCT
1737 //------------------------------dump_head--------------------------------------
1738 // Dump 1 liner for loop header info
1739 void IdealLoopTree::dump_head( ) const {
1740 for (uint i=0; i<_nest; i++)
1741 tty->print(" ");
1742 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1743 if (_irreducible) tty->print(" IRREDUCIBLE");
1744 Node* entry = _head->in(LoopNode::EntryControl);
1745 if (LoopLimitCheck) {
1746 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1747 if (predicate != NULL ) {
1748 tty->print(" limit_check");
1749 entry = entry->in(0)->in(0);
1750 }
1751 }
1752 if (UseLoopPredicate) {
1753 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1754 if (entry != NULL) {
1755 tty->print(" predicated");
1756 }
1757 }
1758 if (_head->is_CountedLoop()) {
1759 CountedLoopNode *cl = _head->as_CountedLoop();
1760 tty->print(" counted");
1762 Node* init_n = cl->init_trip();
1763 if (init_n != NULL && init_n->is_Con())
1764 tty->print(" [%d,", cl->init_trip()->get_int());
1765 else
1766 tty->print(" [int,");
1767 Node* limit_n = cl->limit();
1768 if (limit_n != NULL && limit_n->is_Con())
1769 tty->print("%d),", cl->limit()->get_int());
1770 else
1771 tty->print("int),");
1772 int stride_con = cl->stride_con();
1773 if (stride_con > 0) tty->print("+");
1774 tty->print("%d", stride_con);
1776 tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
1778 if (cl->is_pre_loop ()) tty->print(" pre" );
1779 if (cl->is_main_loop()) tty->print(" main");
1780 if (cl->is_post_loop()) tty->print(" post");
1781 }
1782 tty->cr();
1783 }
1785 //------------------------------dump-------------------------------------------
1786 // Dump loops by loop tree
1787 void IdealLoopTree::dump( ) const {
1788 dump_head();
1789 if (_child) _child->dump();
1790 if (_next) _next ->dump();
1791 }
1793 #endif
1795 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1796 if (loop == root) {
1797 if (loop->_child != NULL) {
1798 log->begin_head("loop_tree");
1799 log->end_head();
1800 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1801 log->tail("loop_tree");
1802 assert(loop->_next == NULL, "what?");
1803 }
1804 } else {
1805 Node* head = loop->_head;
1806 log->begin_head("loop");
1807 log->print(" idx='%d' ", head->_idx);
1808 if (loop->_irreducible) log->print("irreducible='1' ");
1809 if (head->is_Loop()) {
1810 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1811 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1812 }
1813 if (head->is_CountedLoop()) {
1814 CountedLoopNode* cl = head->as_CountedLoop();
1815 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1816 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1817 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1818 }
1819 log->end_head();
1820 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1821 log->tail("loop");
1822 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1823 }
1824 }
1826 //---------------------collect_potentially_useful_predicates-----------------------
1827 // Helper function to collect potentially useful predicates to prevent them from
1828 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1829 void PhaseIdealLoop::collect_potentially_useful_predicates(
1830 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1831 if (loop->_child) { // child
1832 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1833 }
1835 // self (only loops that we can apply loop predication may use their predicates)
1836 if (loop->_head->is_Loop() &&
1837 !loop->_irreducible &&
1838 !loop->tail()->is_top()) {
1839 LoopNode* lpn = loop->_head->as_Loop();
1840 Node* entry = lpn->in(LoopNode::EntryControl);
1841 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1842 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1843 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1844 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1845 entry = entry->in(0)->in(0);
1846 }
1847 predicate_proj = find_predicate(entry); // Predicate
1848 if (predicate_proj != NULL ) {
1849 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1850 }
1851 }
1853 if (loop->_next) { // sibling
1854 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1855 }
1856 }
1858 //------------------------eliminate_useless_predicates-----------------------------
1859 // Eliminate all inserted predicates if they could not be used by loop predication.
1860 // Note: it will also eliminates loop limits check predicate since it also uses
1861 // Opaque1 node (see Parse::add_predicate()).
1862 void PhaseIdealLoop::eliminate_useless_predicates() {
1863 if (C->predicate_count() == 0)
1864 return; // no predicate left
1866 Unique_Node_List useful_predicates; // to store useful predicates
1867 if (C->has_loops()) {
1868 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1869 }
1871 for (int i = C->predicate_count(); i > 0; i--) {
1872 Node * n = C->predicate_opaque1_node(i-1);
1873 assert(n->Opcode() == Op_Opaque1, "must be");
1874 if (!useful_predicates.member(n)) { // not in the useful list
1875 _igvn.replace_node(n, n->in(1));
1876 }
1877 }
1878 }
1880 //=============================================================================
1881 //----------------------------build_and_optimize-------------------------------
1882 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1883 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1884 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
1885 ResourceMark rm;
1887 int old_progress = C->major_progress();
1888 uint orig_worklist_size = _igvn._worklist.size();
1890 // Reset major-progress flag for the driver's heuristics
1891 C->clear_major_progress();
1893 #ifndef PRODUCT
1894 // Capture for later assert
1895 uint unique = C->unique();
1896 _loop_invokes++;
1897 _loop_work += unique;
1898 #endif
1900 // True if the method has at least 1 irreducible loop
1901 _has_irreducible_loops = false;
1903 _created_loop_node = false;
1905 Arena *a = Thread::current()->resource_area();
1906 VectorSet visited(a);
1907 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1908 _nodes.map(C->unique(), NULL);
1909 memset(_nodes.adr(), 0, wordSize * C->unique());
1911 // Pre-build the top-level outermost loop tree entry
1912 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1913 // Do not need a safepoint at the top level
1914 _ltree_root->_has_sfpt = 1;
1916 // Initialize Dominators.
1917 // Checked in clone_loop_predicate() during beautify_loops().
1918 _idom_size = 0;
1919 _idom = NULL;
1920 _dom_depth = NULL;
1921 _dom_stk = NULL;
1923 // Empty pre-order array
1924 allocate_preorders();
1926 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1927 // IdealLoopTree entries. Data nodes are NOT walked.
1928 build_loop_tree();
1929 // Check for bailout, and return
1930 if (C->failing()) {
1931 return;
1932 }
1934 // No loops after all
1935 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
1937 // There should always be an outer loop containing the Root and Return nodes.
1938 // If not, we have a degenerate empty program. Bail out in this case.
1939 if (!has_node(C->root())) {
1940 if (!_verify_only) {
1941 C->clear_major_progress();
1942 C->record_method_not_compilable("empty program detected during loop optimization");
1943 }
1944 return;
1945 }
1947 // Nothing to do, so get out
1948 if( !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only ) {
1949 _igvn.optimize(); // Cleanup NeverBranches
1950 return;
1951 }
1953 // Set loop nesting depth
1954 _ltree_root->set_nest( 0 );
1956 // Split shared headers and insert loop landing pads.
1957 // Do not bother doing this on the Root loop of course.
1958 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
1959 C->print_method("Before beautify loops", 3);
1960 if( _ltree_root->_child->beautify_loops( this ) ) {
1961 // Re-build loop tree!
1962 _ltree_root->_child = NULL;
1963 _nodes.clear();
1964 reallocate_preorders();
1965 build_loop_tree();
1966 // Check for bailout, and return
1967 if (C->failing()) {
1968 return;
1969 }
1970 // Reset loop nesting depth
1971 _ltree_root->set_nest( 0 );
1973 C->print_method("After beautify loops", 3);
1974 }
1975 }
1977 // Build Dominators for elision of NULL checks & loop finding.
1978 // Since nodes do not have a slot for immediate dominator, make
1979 // a persistent side array for that info indexed on node->_idx.
1980 _idom_size = C->unique();
1981 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1982 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1983 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1984 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1986 Dominators();
1988 if (!_verify_only) {
1989 // As a side effect, Dominators removed any unreachable CFG paths
1990 // into RegionNodes. It doesn't do this test against Root, so
1991 // we do it here.
1992 for( uint i = 1; i < C->root()->req(); i++ ) {
1993 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1994 _igvn.delete_input_of(C->root(), i);
1995 i--; // Rerun same iteration on compressed edges
1996 }
1997 }
1999 // Given dominators, try to find inner loops with calls that must
2000 // always be executed (call dominates loop tail). These loops do
2001 // not need a separate safepoint.
2002 Node_List cisstack(a);
2003 _ltree_root->check_safepts(visited, cisstack);
2004 }
2006 // Walk the DATA nodes and place into loops. Find earliest control
2007 // node. For CFG nodes, the _nodes array starts out and remains
2008 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2009 // _nodes array holds the earliest legal controlling CFG node.
2011 // Allocate stack with enough space to avoid frequent realloc
2012 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
2013 Node_Stack nstack( a, stack_size );
2015 visited.Clear();
2016 Node_List worklist(a);
2017 // Don't need C->root() on worklist since
2018 // it will be processed among C->top() inputs
2019 worklist.push( C->top() );
2020 visited.set( C->top()->_idx ); // Set C->top() as visited now
2021 build_loop_early( visited, worklist, nstack );
2023 // Given early legal placement, try finding counted loops. This placement
2024 // is good enough to discover most loop invariants.
2025 if( !_verify_me && !_verify_only )
2026 _ltree_root->counted_loop( this );
2028 // Find latest loop placement. Find ideal loop placement.
2029 visited.Clear();
2030 init_dom_lca_tags();
2031 // Need C->root() on worklist when processing outs
2032 worklist.push( C->root() );
2033 NOT_PRODUCT( C->verify_graph_edges(); )
2034 worklist.push( C->top() );
2035 build_loop_late( visited, worklist, nstack );
2037 if (_verify_only) {
2038 // restore major progress flag
2039 for (int i = 0; i < old_progress; i++)
2040 C->set_major_progress();
2041 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2042 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2043 return;
2044 }
2046 // Some parser-inserted loop predicates could never be used by loop
2047 // predication or they were moved away from loop during some optimizations.
2048 // For example, peeling. Eliminate them before next loop optimizations.
2049 if (UseLoopPredicate || LoopLimitCheck) {
2050 eliminate_useless_predicates();
2051 }
2053 // clear out the dead code
2054 while(_deadlist.size()) {
2055 _igvn.remove_globally_dead_node(_deadlist.pop());
2056 }
2058 #ifndef PRODUCT
2059 C->verify_graph_edges();
2060 if (_verify_me) { // Nested verify pass?
2061 // Check to see if the verify mode is broken
2062 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2063 return;
2064 }
2065 if(VerifyLoopOptimizations) verify();
2066 if(TraceLoopOpts && C->has_loops()) {
2067 _ltree_root->dump();
2068 }
2069 #endif
2071 if (skip_loop_opts) {
2072 // Cleanup any modified bits
2073 _igvn.optimize();
2075 if (C->log() != NULL) {
2076 log_loop_tree(_ltree_root, _ltree_root, C->log());
2077 }
2078 return;
2079 }
2081 if (ReassociateInvariants) {
2082 // Reassociate invariants and prep for split_thru_phi
2083 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2084 IdealLoopTree* lpt = iter.current();
2085 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2087 lpt->reassociate_invariants(this);
2089 // Because RCE opportunities can be masked by split_thru_phi,
2090 // look for RCE candidates and inhibit split_thru_phi
2091 // on just their loop-phi's for this pass of loop opts
2092 if (SplitIfBlocks && do_split_ifs) {
2093 if (lpt->policy_range_check(this)) {
2094 lpt->_rce_candidate = 1; // = true
2095 }
2096 }
2097 }
2098 }
2100 // Check for aggressive application of split-if and other transforms
2101 // that require basic-block info (like cloning through Phi's)
2102 if( SplitIfBlocks && do_split_ifs ) {
2103 visited.Clear();
2104 split_if_with_blocks( visited, nstack );
2105 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2106 }
2108 // Perform loop predication before iteration splitting
2109 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2110 _ltree_root->_child->loop_predication(this);
2111 }
2113 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2114 if (do_intrinsify_fill()) {
2115 C->set_major_progress();
2116 }
2117 }
2119 // Perform iteration-splitting on inner loops. Split iterations to avoid
2120 // range checks or one-shot null checks.
2122 // If split-if's didn't hack the graph too bad (no CFG changes)
2123 // then do loop opts.
2124 if (C->has_loops() && !C->major_progress()) {
2125 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2126 _ltree_root->_child->iteration_split( this, worklist );
2127 // No verify after peeling! GCM has hoisted code out of the loop.
2128 // After peeling, the hoisted code could sink inside the peeled area.
2129 // The peeling code does not try to recompute the best location for
2130 // all the code before the peeled area, so the verify pass will always
2131 // complain about it.
2132 }
2133 // Do verify graph edges in any case
2134 NOT_PRODUCT( C->verify_graph_edges(); );
2136 if (!do_split_ifs) {
2137 // We saw major progress in Split-If to get here. We forced a
2138 // pass with unrolling and not split-if, however more split-if's
2139 // might make progress. If the unrolling didn't make progress
2140 // then the major-progress flag got cleared and we won't try
2141 // another round of Split-If. In particular the ever-common
2142 // instance-of/check-cast pattern requires at least 2 rounds of
2143 // Split-If to clear out.
2144 C->set_major_progress();
2145 }
2147 // Repeat loop optimizations if new loops were seen
2148 if (created_loop_node()) {
2149 C->set_major_progress();
2150 }
2152 // Keep loop predicates and perform optimizations with them
2153 // until no more loop optimizations could be done.
2154 // After that switch predicates off and do more loop optimizations.
2155 if (!C->major_progress() && (C->predicate_count() > 0)) {
2156 C->cleanup_loop_predicates(_igvn);
2157 #ifndef PRODUCT
2158 if (TraceLoopOpts) {
2159 tty->print_cr("PredicatesOff");
2160 }
2161 #endif
2162 C->set_major_progress();
2163 }
2165 // Convert scalar to superword operations at the end of all loop opts.
2166 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2167 // SuperWord transform
2168 SuperWord sw(this);
2169 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2170 IdealLoopTree* lpt = iter.current();
2171 if (lpt->is_counted()) {
2172 sw.transform_loop(lpt);
2173 }
2174 }
2175 }
2177 // Cleanup any modified bits
2178 _igvn.optimize();
2180 // disable assert until issue with split_flow_path is resolved (6742111)
2181 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2182 // "shouldn't introduce irreducible loops");
2184 if (C->log() != NULL) {
2185 log_loop_tree(_ltree_root, _ltree_root, C->log());
2186 }
2187 }
2189 #ifndef PRODUCT
2190 //------------------------------print_statistics-------------------------------
2191 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2192 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2193 void PhaseIdealLoop::print_statistics() {
2194 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2195 }
2197 //------------------------------verify-----------------------------------------
2198 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2199 static int fail; // debug only, so its multi-thread dont care
2200 void PhaseIdealLoop::verify() const {
2201 int old_progress = C->major_progress();
2202 ResourceMark rm;
2203 PhaseIdealLoop loop_verify( _igvn, this );
2204 VectorSet visited(Thread::current()->resource_area());
2206 fail = 0;
2207 verify_compare( C->root(), &loop_verify, visited );
2208 assert( fail == 0, "verify loops failed" );
2209 // Verify loop structure is the same
2210 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2211 // Reset major-progress. It was cleared by creating a verify version of
2212 // PhaseIdealLoop.
2213 for( int i=0; i<old_progress; i++ )
2214 C->set_major_progress();
2215 }
2217 //------------------------------verify_compare---------------------------------
2218 // Make sure me and the given PhaseIdealLoop agree on key data structures
2219 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2220 if( !n ) return;
2221 if( visited.test_set( n->_idx ) ) return;
2222 if( !_nodes[n->_idx] ) { // Unreachable
2223 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2224 return;
2225 }
2227 uint i;
2228 for( i = 0; i < n->req(); i++ )
2229 verify_compare( n->in(i), loop_verify, visited );
2231 // Check the '_nodes' block/loop structure
2232 i = n->_idx;
2233 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2234 if( _nodes[i] != loop_verify->_nodes[i] &&
2235 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2236 tty->print("Mismatched control setting for: ");
2237 n->dump();
2238 if( fail++ > 10 ) return;
2239 Node *c = get_ctrl_no_update(n);
2240 tty->print("We have it as: ");
2241 if( c->in(0) ) c->dump();
2242 else tty->print_cr("N%d",c->_idx);
2243 tty->print("Verify thinks: ");
2244 if( loop_verify->has_ctrl(n) )
2245 loop_verify->get_ctrl_no_update(n)->dump();
2246 else
2247 loop_verify->get_loop_idx(n)->dump();
2248 tty->cr();
2249 }
2250 } else { // We have a loop
2251 IdealLoopTree *us = get_loop_idx(n);
2252 if( loop_verify->has_ctrl(n) ) {
2253 tty->print("Mismatched loop setting for: ");
2254 n->dump();
2255 if( fail++ > 10 ) return;
2256 tty->print("We have it as: ");
2257 us->dump();
2258 tty->print("Verify thinks: ");
2259 loop_verify->get_ctrl_no_update(n)->dump();
2260 tty->cr();
2261 } else if (!C->major_progress()) {
2262 // Loop selection can be messed up if we did a major progress
2263 // operation, like split-if. Do not verify in that case.
2264 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2265 if( us->_head != them->_head || us->_tail != them->_tail ) {
2266 tty->print("Unequals loops for: ");
2267 n->dump();
2268 if( fail++ > 10 ) return;
2269 tty->print("We have it as: ");
2270 us->dump();
2271 tty->print("Verify thinks: ");
2272 them->dump();
2273 tty->cr();
2274 }
2275 }
2276 }
2278 // Check for immediate dominators being equal
2279 if( i >= _idom_size ) {
2280 if( !n->is_CFG() ) return;
2281 tty->print("CFG Node with no idom: ");
2282 n->dump();
2283 return;
2284 }
2285 if( !n->is_CFG() ) return;
2286 if( n == C->root() ) return; // No IDOM here
2288 assert(n->_idx == i, "sanity");
2289 Node *id = idom_no_update(n);
2290 if( id != loop_verify->idom_no_update(n) ) {
2291 tty->print("Unequals idoms for: ");
2292 n->dump();
2293 if( fail++ > 10 ) return;
2294 tty->print("We have it as: ");
2295 id->dump();
2296 tty->print("Verify thinks: ");
2297 loop_verify->idom_no_update(n)->dump();
2298 tty->cr();
2299 }
2301 }
2303 //------------------------------verify_tree------------------------------------
2304 // Verify that tree structures match. Because the CFG can change, siblings
2305 // within the loop tree can be reordered. We attempt to deal with that by
2306 // reordering the verify's loop tree if possible.
2307 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2308 assert( _parent == parent, "Badly formed loop tree" );
2310 // Siblings not in same order? Attempt to re-order.
2311 if( _head != loop->_head ) {
2312 // Find _next pointer to update
2313 IdealLoopTree **pp = &loop->_parent->_child;
2314 while( *pp != loop )
2315 pp = &((*pp)->_next);
2316 // Find proper sibling to be next
2317 IdealLoopTree **nn = &loop->_next;
2318 while( (*nn) && (*nn)->_head != _head )
2319 nn = &((*nn)->_next);
2321 // Check for no match.
2322 if( !(*nn) ) {
2323 // Annoyingly, irreducible loops can pick different headers
2324 // after a major_progress operation, so the rest of the loop
2325 // tree cannot be matched.
2326 if (_irreducible && Compile::current()->major_progress()) return;
2327 assert( 0, "failed to match loop tree" );
2328 }
2330 // Move (*nn) to (*pp)
2331 IdealLoopTree *hit = *nn;
2332 *nn = hit->_next;
2333 hit->_next = loop;
2334 *pp = loop;
2335 loop = hit;
2336 // Now try again to verify
2337 }
2339 assert( _head == loop->_head , "mismatched loop head" );
2340 Node *tail = _tail; // Inline a non-updating version of
2341 while( !tail->in(0) ) // the 'tail()' call.
2342 tail = tail->in(1);
2343 assert( tail == loop->_tail, "mismatched loop tail" );
2345 // Counted loops that are guarded should be able to find their guards
2346 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2347 CountedLoopNode *cl = _head->as_CountedLoop();
2348 Node *init = cl->init_trip();
2349 Node *ctrl = cl->in(LoopNode::EntryControl);
2350 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2351 Node *iff = ctrl->in(0);
2352 assert( iff->Opcode() == Op_If, "" );
2353 Node *bol = iff->in(1);
2354 assert( bol->Opcode() == Op_Bool, "" );
2355 Node *cmp = bol->in(1);
2356 assert( cmp->Opcode() == Op_CmpI, "" );
2357 Node *add = cmp->in(1);
2358 Node *opaq;
2359 if( add->Opcode() == Op_Opaque1 ) {
2360 opaq = add;
2361 } else {
2362 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2363 assert( add == init, "" );
2364 opaq = cmp->in(2);
2365 }
2366 assert( opaq->Opcode() == Op_Opaque1, "" );
2368 }
2370 if (_child != NULL) _child->verify_tree(loop->_child, this);
2371 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2372 // Innermost loops need to verify loop bodies,
2373 // but only if no 'major_progress'
2374 int fail = 0;
2375 if (!Compile::current()->major_progress() && _child == NULL) {
2376 for( uint i = 0; i < _body.size(); i++ ) {
2377 Node *n = _body.at(i);
2378 if (n->outcnt() == 0) continue; // Ignore dead
2379 uint j;
2380 for( j = 0; j < loop->_body.size(); j++ )
2381 if( loop->_body.at(j) == n )
2382 break;
2383 if( j == loop->_body.size() ) { // Not found in loop body
2384 // Last ditch effort to avoid assertion: Its possible that we
2385 // have some users (so outcnt not zero) but are still dead.
2386 // Try to find from root.
2387 if (Compile::current()->root()->find(n->_idx)) {
2388 fail++;
2389 tty->print("We have that verify does not: ");
2390 n->dump();
2391 }
2392 }
2393 }
2394 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2395 Node *n = loop->_body.at(i2);
2396 if (n->outcnt() == 0) continue; // Ignore dead
2397 uint j;
2398 for( j = 0; j < _body.size(); j++ )
2399 if( _body.at(j) == n )
2400 break;
2401 if( j == _body.size() ) { // Not found in loop body
2402 // Last ditch effort to avoid assertion: Its possible that we
2403 // have some users (so outcnt not zero) but are still dead.
2404 // Try to find from root.
2405 if (Compile::current()->root()->find(n->_idx)) {
2406 fail++;
2407 tty->print("Verify has that we do not: ");
2408 n->dump();
2409 }
2410 }
2411 }
2412 assert( !fail, "loop body mismatch" );
2413 }
2414 }
2416 #endif
2418 //------------------------------set_idom---------------------------------------
2419 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2420 uint idx = d->_idx;
2421 if (idx >= _idom_size) {
2422 uint newsize = _idom_size<<1;
2423 while( idx >= newsize ) {
2424 newsize <<= 1;
2425 }
2426 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2427 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2428 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2429 _idom_size = newsize;
2430 }
2431 _idom[idx] = n;
2432 _dom_depth[idx] = dom_depth;
2433 }
2435 //------------------------------recompute_dom_depth---------------------------------------
2436 // The dominator tree is constructed with only parent pointers.
2437 // This recomputes the depth in the tree by first tagging all
2438 // nodes as "no depth yet" marker. The next pass then runs up
2439 // the dom tree from each node marked "no depth yet", and computes
2440 // the depth on the way back down.
2441 void PhaseIdealLoop::recompute_dom_depth() {
2442 uint no_depth_marker = C->unique();
2443 uint i;
2444 // Initialize depth to "no depth yet"
2445 for (i = 0; i < _idom_size; i++) {
2446 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2447 _dom_depth[i] = no_depth_marker;
2448 }
2449 }
2450 if (_dom_stk == NULL) {
2451 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
2452 if (init_size < 10) init_size = 10;
2453 _dom_stk = new GrowableArray<uint>(init_size);
2454 }
2455 // Compute new depth for each node.
2456 for (i = 0; i < _idom_size; i++) {
2457 uint j = i;
2458 // Run up the dom tree to find a node with a depth
2459 while (_dom_depth[j] == no_depth_marker) {
2460 _dom_stk->push(j);
2461 j = _idom[j]->_idx;
2462 }
2463 // Compute the depth on the way back down this tree branch
2464 uint dd = _dom_depth[j] + 1;
2465 while (_dom_stk->length() > 0) {
2466 uint j = _dom_stk->pop();
2467 _dom_depth[j] = dd;
2468 dd++;
2469 }
2470 }
2471 }
2473 //------------------------------sort-------------------------------------------
2474 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2475 // loop tree, not the root.
2476 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2477 if( !innermost ) return loop; // New innermost loop
2479 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2480 assert( loop_preorder, "not yet post-walked loop" );
2481 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2482 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2484 // Insert at start of list
2485 while( l ) { // Insertion sort based on pre-order
2486 if( l == loop ) return innermost; // Already on list!
2487 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2488 assert( l_preorder, "not yet post-walked l" );
2489 // Check header pre-order number to figure proper nesting
2490 if( loop_preorder > l_preorder )
2491 break; // End of insertion
2492 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2493 // Since I split shared headers, you'd think this could not happen.
2494 // BUT: I must first do the preorder numbering before I can discover I
2495 // have shared headers, so the split headers all get the same preorder
2496 // number as the RegionNode they split from.
2497 if( loop_preorder == l_preorder &&
2498 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2499 break; // Also check for shared headers (same pre#)
2500 pp = &l->_parent; // Chain up list
2501 l = *pp;
2502 }
2503 // Link into list
2504 // Point predecessor to me
2505 *pp = loop;
2506 // Point me to successor
2507 IdealLoopTree *p = loop->_parent;
2508 loop->_parent = l; // Point me to successor
2509 if( p ) sort( p, innermost ); // Insert my parents into list as well
2510 return innermost;
2511 }
2513 //------------------------------build_loop_tree--------------------------------
2514 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2515 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2516 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2517 // tightest enclosing IdealLoopTree for post-walked.
2518 //
2519 // During my forward walk I do a short 1-layer lookahead to see if I can find
2520 // a loop backedge with that doesn't have any work on the backedge. This
2521 // helps me construct nested loops with shared headers better.
2522 //
2523 // Once I've done the forward recursion, I do the post-work. For each child
2524 // I check to see if there is a backedge. Backedges define a loop! I
2525 // insert an IdealLoopTree at the target of the backedge.
2526 //
2527 // During the post-work I also check to see if I have several children
2528 // belonging to different loops. If so, then this Node is a decision point
2529 // where control flow can choose to change loop nests. It is at this
2530 // decision point where I can figure out how loops are nested. At this
2531 // time I can properly order the different loop nests from my children.
2532 // Note that there may not be any backedges at the decision point!
2533 //
2534 // Since the decision point can be far removed from the backedges, I can't
2535 // order my loops at the time I discover them. Thus at the decision point
2536 // I need to inspect loop header pre-order numbers to properly nest my
2537 // loops. This means I need to sort my childrens' loops by pre-order.
2538 // The sort is of size number-of-control-children, which generally limits
2539 // it to size 2 (i.e., I just choose between my 2 target loops).
2540 void PhaseIdealLoop::build_loop_tree() {
2541 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2542 GrowableArray <Node *> bltstack(C->unique() >> 1);
2543 Node *n = C->root();
2544 bltstack.push(n);
2545 int pre_order = 1;
2546 int stack_size;
2548 while ( ( stack_size = bltstack.length() ) != 0 ) {
2549 n = bltstack.top(); // Leave node on stack
2550 if ( !is_visited(n) ) {
2551 // ---- Pre-pass Work ----
2552 // Pre-walked but not post-walked nodes need a pre_order number.
2554 set_preorder_visited( n, pre_order ); // set as visited
2556 // ---- Scan over children ----
2557 // Scan first over control projections that lead to loop headers.
2558 // This helps us find inner-to-outer loops with shared headers better.
2560 // Scan children's children for loop headers.
2561 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2562 Node* m = n->raw_out(i); // Child
2563 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2564 // Scan over children's children to find loop
2565 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2566 Node* l = m->fast_out(j);
2567 if( is_visited(l) && // Been visited?
2568 !is_postvisited(l) && // But not post-visited
2569 get_preorder(l) < pre_order ) { // And smaller pre-order
2570 // Found! Scan the DFS down this path before doing other paths
2571 bltstack.push(m);
2572 break;
2573 }
2574 }
2575 }
2576 }
2577 pre_order++;
2578 }
2579 else if ( !is_postvisited(n) ) {
2580 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2581 // such as com.sun.rsasign.am::a.
2582 // For non-recursive version, first, process current children.
2583 // On next iteration, check if additional children were added.
2584 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2585 Node* u = n->raw_out(k);
2586 if ( u->is_CFG() && !is_visited(u) ) {
2587 bltstack.push(u);
2588 }
2589 }
2590 if ( bltstack.length() == stack_size ) {
2591 // There were no additional children, post visit node now
2592 (void)bltstack.pop(); // Remove node from stack
2593 pre_order = build_loop_tree_impl( n, pre_order );
2594 // Check for bailout
2595 if (C->failing()) {
2596 return;
2597 }
2598 // Check to grow _preorders[] array for the case when
2599 // build_loop_tree_impl() adds new nodes.
2600 check_grow_preorders();
2601 }
2602 }
2603 else {
2604 (void)bltstack.pop(); // Remove post-visited node from stack
2605 }
2606 }
2607 }
2609 //------------------------------build_loop_tree_impl---------------------------
2610 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2611 // ---- Post-pass Work ----
2612 // Pre-walked but not post-walked nodes need a pre_order number.
2614 // Tightest enclosing loop for this Node
2615 IdealLoopTree *innermost = NULL;
2617 // For all children, see if any edge is a backedge. If so, make a loop
2618 // for it. Then find the tightest enclosing loop for the self Node.
2619 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2620 Node* m = n->fast_out(i); // Child
2621 if( n == m ) continue; // Ignore control self-cycles
2622 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2624 IdealLoopTree *l; // Child's loop
2625 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2626 // Found a backedge
2627 assert( get_preorder(m) < pre_order, "should be backedge" );
2628 // Check for the RootNode, which is already a LoopNode and is allowed
2629 // to have multiple "backedges".
2630 if( m == C->root()) { // Found the root?
2631 l = _ltree_root; // Root is the outermost LoopNode
2632 } else { // Else found a nested loop
2633 // Insert a LoopNode to mark this loop.
2634 l = new IdealLoopTree(this, m, n);
2635 } // End of Else found a nested loop
2636 if( !has_loop(m) ) // If 'm' does not already have a loop set
2637 set_loop(m, l); // Set loop header to loop now
2639 } else { // Else not a nested loop
2640 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2641 l = get_loop(m); // Get previously determined loop
2642 // If successor is header of a loop (nest), move up-loop till it
2643 // is a member of some outer enclosing loop. Since there are no
2644 // shared headers (I've split them already) I only need to go up
2645 // at most 1 level.
2646 while( l && l->_head == m ) // Successor heads loop?
2647 l = l->_parent; // Move up 1 for me
2648 // If this loop is not properly parented, then this loop
2649 // has no exit path out, i.e. its an infinite loop.
2650 if( !l ) {
2651 // Make loop "reachable" from root so the CFG is reachable. Basically
2652 // insert a bogus loop exit that is never taken. 'm', the loop head,
2653 // points to 'n', one (of possibly many) fall-in paths. There may be
2654 // many backedges as well.
2656 // Here I set the loop to be the root loop. I could have, after
2657 // inserting a bogus loop exit, restarted the recursion and found my
2658 // new loop exit. This would make the infinite loop a first-class
2659 // loop and it would then get properly optimized. What's the use of
2660 // optimizing an infinite loop?
2661 l = _ltree_root; // Oops, found infinite loop
2663 if (!_verify_only) {
2664 // Insert the NeverBranch between 'm' and it's control user.
2665 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2666 _igvn.register_new_node_with_optimizer(iff);
2667 set_loop(iff, l);
2668 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2669 _igvn.register_new_node_with_optimizer(if_t);
2670 set_loop(if_t, l);
2672 Node* cfg = NULL; // Find the One True Control User of m
2673 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2674 Node* x = m->fast_out(j);
2675 if (x->is_CFG() && x != m && x != iff)
2676 { cfg = x; break; }
2677 }
2678 assert(cfg != NULL, "must find the control user of m");
2679 uint k = 0; // Probably cfg->in(0)
2680 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2681 cfg->set_req( k, if_t ); // Now point to NeverBranch
2683 // Now create the never-taken loop exit
2684 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2685 _igvn.register_new_node_with_optimizer(if_f);
2686 set_loop(if_f, l);
2687 // Find frame ptr for Halt. Relies on the optimizer
2688 // V-N'ing. Easier and quicker than searching through
2689 // the program structure.
2690 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2691 _igvn.register_new_node_with_optimizer(frame);
2692 // Halt & Catch Fire
2693 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2694 _igvn.register_new_node_with_optimizer(halt);
2695 set_loop(halt, l);
2696 C->root()->add_req(halt);
2697 }
2698 set_loop(C->root(), _ltree_root);
2699 }
2700 }
2701 // Weeny check for irreducible. This child was already visited (this
2702 // IS the post-work phase). Is this child's loop header post-visited
2703 // as well? If so, then I found another entry into the loop.
2704 if (!_verify_only) {
2705 while( is_postvisited(l->_head) ) {
2706 // found irreducible
2707 l->_irreducible = 1; // = true
2708 l = l->_parent;
2709 _has_irreducible_loops = true;
2710 // Check for bad CFG here to prevent crash, and bailout of compile
2711 if (l == NULL) {
2712 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2713 return pre_order;
2714 }
2715 }
2716 }
2718 // This Node might be a decision point for loops. It is only if
2719 // it's children belong to several different loops. The sort call
2720 // does a trivial amount of work if there is only 1 child or all
2721 // children belong to the same loop. If however, the children
2722 // belong to different loops, the sort call will properly set the
2723 // _parent pointers to show how the loops nest.
2724 //
2725 // In any case, it returns the tightest enclosing loop.
2726 innermost = sort( l, innermost );
2727 }
2729 // Def-use info will have some dead stuff; dead stuff will have no
2730 // loop decided on.
2732 // Am I a loop header? If so fix up my parent's child and next ptrs.
2733 if( innermost && innermost->_head == n ) {
2734 assert( get_loop(n) == innermost, "" );
2735 IdealLoopTree *p = innermost->_parent;
2736 IdealLoopTree *l = innermost;
2737 while( p && l->_head == n ) {
2738 l->_next = p->_child; // Put self on parents 'next child'
2739 p->_child = l; // Make self as first child of parent
2740 l = p; // Now walk up the parent chain
2741 p = l->_parent;
2742 }
2743 } else {
2744 // Note that it is possible for a LoopNode to reach here, if the
2745 // backedge has been made unreachable (hence the LoopNode no longer
2746 // denotes a Loop, and will eventually be removed).
2748 // Record tightest enclosing loop for self. Mark as post-visited.
2749 set_loop(n, innermost);
2750 // Also record has_call flag early on
2751 if( innermost ) {
2752 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2753 // Do not count uncommon calls
2754 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2755 Node *iff = n->in(0)->in(0);
2756 // No any calls for vectorized loops.
2757 if( UseSuperWord || !iff->is_If() ||
2758 (n->in(0)->Opcode() == Op_IfFalse &&
2759 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2760 (iff->as_If()->_prob >= 0.01) )
2761 innermost->_has_call = 1;
2762 }
2763 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2764 // Disable loop optimizations if the loop has a scalar replaceable
2765 // allocation. This disabling may cause a potential performance lost
2766 // if the allocation is not eliminated for some reason.
2767 innermost->_allow_optimizations = false;
2768 innermost->_has_call = 1; // = true
2769 }
2770 }
2771 }
2773 // Flag as post-visited now
2774 set_postvisited(n);
2775 return pre_order;
2776 }
2779 //------------------------------build_loop_early-------------------------------
2780 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2781 // First pass computes the earliest controlling node possible. This is the
2782 // controlling input with the deepest dominating depth.
2783 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2784 while (worklist.size() != 0) {
2785 // Use local variables nstack_top_n & nstack_top_i to cache values
2786 // on nstack's top.
2787 Node *nstack_top_n = worklist.pop();
2788 uint nstack_top_i = 0;
2789 //while_nstack_nonempty:
2790 while (true) {
2791 // Get parent node and next input's index from stack's top.
2792 Node *n = nstack_top_n;
2793 uint i = nstack_top_i;
2794 uint cnt = n->req(); // Count of inputs
2795 if (i == 0) { // Pre-process the node.
2796 if( has_node(n) && // Have either loop or control already?
2797 !has_ctrl(n) ) { // Have loop picked out already?
2798 // During "merge_many_backedges" we fold up several nested loops
2799 // into a single loop. This makes the members of the original
2800 // loop bodies pointing to dead loops; they need to move up
2801 // to the new UNION'd larger loop. I set the _head field of these
2802 // dead loops to NULL and the _parent field points to the owning
2803 // loop. Shades of UNION-FIND algorithm.
2804 IdealLoopTree *ilt;
2805 while( !(ilt = get_loop(n))->_head ) {
2806 // Normally I would use a set_loop here. But in this one special
2807 // case, it is legal (and expected) to change what loop a Node
2808 // belongs to.
2809 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2810 }
2811 // Remove safepoints ONLY if I've already seen I don't need one.
2812 // (the old code here would yank a 2nd safepoint after seeing a
2813 // first one, even though the 1st did not dominate in the loop body
2814 // and thus could be avoided indefinitely)
2815 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2816 is_deleteable_safept(n)) {
2817 Node *in = n->in(TypeFunc::Control);
2818 lazy_replace(n,in); // Pull safepoint now
2819 // Carry on with the recursion "as if" we are walking
2820 // only the control input
2821 if( !visited.test_set( in->_idx ) ) {
2822 worklist.push(in); // Visit this guy later, using worklist
2823 }
2824 // Get next node from nstack:
2825 // - skip n's inputs processing by setting i > cnt;
2826 // - we also will not call set_early_ctrl(n) since
2827 // has_node(n) == true (see the condition above).
2828 i = cnt + 1;
2829 }
2830 }
2831 } // if (i == 0)
2833 // Visit all inputs
2834 bool done = true; // Assume all n's inputs will be processed
2835 while (i < cnt) {
2836 Node *in = n->in(i);
2837 ++i;
2838 if (in == NULL) continue;
2839 if (in->pinned() && !in->is_CFG())
2840 set_ctrl(in, in->in(0));
2841 int is_visited = visited.test_set( in->_idx );
2842 if (!has_node(in)) { // No controlling input yet?
2843 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2844 assert( !is_visited, "visit only once" );
2845 nstack.push(n, i); // Save parent node and next input's index.
2846 nstack_top_n = in; // Process current input now.
2847 nstack_top_i = 0;
2848 done = false; // Not all n's inputs processed.
2849 break; // continue while_nstack_nonempty;
2850 } else if (!is_visited) {
2851 // This guy has a location picked out for him, but has not yet
2852 // been visited. Happens to all CFG nodes, for instance.
2853 // Visit him using the worklist instead of recursion, to break
2854 // cycles. Since he has a location already we do not need to
2855 // find his location before proceeding with the current Node.
2856 worklist.push(in); // Visit this guy later, using worklist
2857 }
2858 }
2859 if (done) {
2860 // All of n's inputs have been processed, complete post-processing.
2862 // Compute earliest point this Node can go.
2863 // CFG, Phi, pinned nodes already know their controlling input.
2864 if (!has_node(n)) {
2865 // Record earliest legal location
2866 set_early_ctrl( n );
2867 }
2868 if (nstack.is_empty()) {
2869 // Finished all nodes on stack.
2870 // Process next node on the worklist.
2871 break;
2872 }
2873 // Get saved parent node and next input's index.
2874 nstack_top_n = nstack.node();
2875 nstack_top_i = nstack.index();
2876 nstack.pop();
2877 }
2878 } // while (true)
2879 }
2880 }
2882 //------------------------------dom_lca_internal--------------------------------
2883 // Pair-wise LCA
2884 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2885 if( !n1 ) return n2; // Handle NULL original LCA
2886 assert( n1->is_CFG(), "" );
2887 assert( n2->is_CFG(), "" );
2888 // find LCA of all uses
2889 uint d1 = dom_depth(n1);
2890 uint d2 = dom_depth(n2);
2891 while (n1 != n2) {
2892 if (d1 > d2) {
2893 n1 = idom(n1);
2894 d1 = dom_depth(n1);
2895 } else if (d1 < d2) {
2896 n2 = idom(n2);
2897 d2 = dom_depth(n2);
2898 } else {
2899 // Here d1 == d2. Due to edits of the dominator-tree, sections
2900 // of the tree might have the same depth. These sections have
2901 // to be searched more carefully.
2903 // Scan up all the n1's with equal depth, looking for n2.
2904 Node *t1 = idom(n1);
2905 while (dom_depth(t1) == d1) {
2906 if (t1 == n2) return n2;
2907 t1 = idom(t1);
2908 }
2909 // Scan up all the n2's with equal depth, looking for n1.
2910 Node *t2 = idom(n2);
2911 while (dom_depth(t2) == d2) {
2912 if (t2 == n1) return n1;
2913 t2 = idom(t2);
2914 }
2915 // Move up to a new dominator-depth value as well as up the dom-tree.
2916 n1 = t1;
2917 n2 = t2;
2918 d1 = dom_depth(n1);
2919 d2 = dom_depth(n2);
2920 }
2921 }
2922 return n1;
2923 }
2925 //------------------------------compute_idom-----------------------------------
2926 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2927 // IDOMs are correct.
2928 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2929 assert( region->is_Region(), "" );
2930 Node *LCA = NULL;
2931 for( uint i = 1; i < region->req(); i++ ) {
2932 if( region->in(i) != C->top() )
2933 LCA = dom_lca( LCA, region->in(i) );
2934 }
2935 return LCA;
2936 }
2938 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
2939 bool had_error = false;
2940 #ifdef ASSERT
2941 if (early != C->root()) {
2942 // Make sure that there's a dominance path from use to LCA
2943 Node* d = use;
2944 while (d != LCA) {
2945 d = idom(d);
2946 if (d == C->root()) {
2947 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
2948 n->dump();
2949 use->dump();
2950 had_error = true;
2951 break;
2952 }
2953 }
2954 }
2955 #endif
2956 return had_error;
2957 }
2960 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
2961 // Compute LCA over list of uses
2962 bool had_error = false;
2963 Node *LCA = NULL;
2964 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2965 Node* c = n->fast_out(i);
2966 if (_nodes[c->_idx] == NULL)
2967 continue; // Skip the occasional dead node
2968 if( c->is_Phi() ) { // For Phis, we must land above on the path
2969 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2970 if( c->in(j) == n ) { // Found matching input?
2971 Node *use = c->in(0)->in(j);
2972 if (_verify_only && use->is_top()) continue;
2973 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2974 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2975 }
2976 }
2977 } else {
2978 // For CFG data-users, use is in the block just prior
2979 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2980 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2981 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2982 }
2983 }
2984 assert(!had_error, "bad dominance");
2985 return LCA;
2986 }
2988 //------------------------------get_late_ctrl----------------------------------
2989 // Compute latest legal control.
2990 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2991 assert(early != NULL, "early control should not be NULL");
2993 Node* LCA = compute_lca_of_uses(n, early);
2994 #ifdef ASSERT
2995 if (LCA == C->root() && LCA != early) {
2996 // def doesn't dominate uses so print some useful debugging output
2997 compute_lca_of_uses(n, early, true);
2998 }
2999 #endif
3001 // if this is a load, check for anti-dependent stores
3002 // We use a conservative algorithm to identify potential interfering
3003 // instructions and for rescheduling the load. The users of the memory
3004 // input of this load are examined. Any use which is not a load and is
3005 // dominated by early is considered a potentially interfering store.
3006 // This can produce false positives.
3007 if (n->is_Load() && LCA != early) {
3008 Node_List worklist;
3010 Node *mem = n->in(MemNode::Memory);
3011 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3012 Node* s = mem->fast_out(i);
3013 worklist.push(s);
3014 }
3015 while(worklist.size() != 0 && LCA != early) {
3016 Node* s = worklist.pop();
3017 if (s->is_Load()) {
3018 continue;
3019 } else if (s->is_MergeMem()) {
3020 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3021 Node* s1 = s->fast_out(i);
3022 worklist.push(s1);
3023 }
3024 } else {
3025 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3026 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3027 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3028 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3029 }
3030 }
3031 }
3032 }
3034 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3035 return LCA;
3036 }
3038 // true if CFG node d dominates CFG node n
3039 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3040 if (d == n)
3041 return true;
3042 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3043 uint dd = dom_depth(d);
3044 while (dom_depth(n) >= dd) {
3045 if (n == d)
3046 return true;
3047 n = idom(n);
3048 }
3049 return false;
3050 }
3052 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3053 // Pair-wise LCA with tags.
3054 // Tag each index with the node 'tag' currently being processed
3055 // before advancing up the dominator chain using idom().
3056 // Later calls that find a match to 'tag' know that this path has already
3057 // been considered in the current LCA (which is input 'n1' by convention).
3058 // Since get_late_ctrl() is only called once for each node, the tag array
3059 // does not need to be cleared between calls to get_late_ctrl().
3060 // Algorithm trades a larger constant factor for better asymptotic behavior
3061 //
3062 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3063 uint d1 = dom_depth(n1);
3064 uint d2 = dom_depth(n2);
3066 do {
3067 if (d1 > d2) {
3068 // current lca is deeper than n2
3069 _dom_lca_tags.map(n1->_idx, tag);
3070 n1 = idom(n1);
3071 d1 = dom_depth(n1);
3072 } else if (d1 < d2) {
3073 // n2 is deeper than current lca
3074 Node *memo = _dom_lca_tags[n2->_idx];
3075 if( memo == tag ) {
3076 return n1; // Return the current LCA
3077 }
3078 _dom_lca_tags.map(n2->_idx, tag);
3079 n2 = idom(n2);
3080 d2 = dom_depth(n2);
3081 } else {
3082 // Here d1 == d2. Due to edits of the dominator-tree, sections
3083 // of the tree might have the same depth. These sections have
3084 // to be searched more carefully.
3086 // Scan up all the n1's with equal depth, looking for n2.
3087 _dom_lca_tags.map(n1->_idx, tag);
3088 Node *t1 = idom(n1);
3089 while (dom_depth(t1) == d1) {
3090 if (t1 == n2) return n2;
3091 _dom_lca_tags.map(t1->_idx, tag);
3092 t1 = idom(t1);
3093 }
3094 // Scan up all the n2's with equal depth, looking for n1.
3095 _dom_lca_tags.map(n2->_idx, tag);
3096 Node *t2 = idom(n2);
3097 while (dom_depth(t2) == d2) {
3098 if (t2 == n1) return n1;
3099 _dom_lca_tags.map(t2->_idx, tag);
3100 t2 = idom(t2);
3101 }
3102 // Move up to a new dominator-depth value as well as up the dom-tree.
3103 n1 = t1;
3104 n2 = t2;
3105 d1 = dom_depth(n1);
3106 d2 = dom_depth(n2);
3107 }
3108 } while (n1 != n2);
3109 return n1;
3110 }
3112 //------------------------------init_dom_lca_tags------------------------------
3113 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3114 // Intended use does not involve any growth for the array, so it could
3115 // be of fixed size.
3116 void PhaseIdealLoop::init_dom_lca_tags() {
3117 uint limit = C->unique() + 1;
3118 _dom_lca_tags.map( limit, NULL );
3119 #ifdef ASSERT
3120 for( uint i = 0; i < limit; ++i ) {
3121 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3122 }
3123 #endif // ASSERT
3124 }
3126 //------------------------------clear_dom_lca_tags------------------------------
3127 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3128 // Intended use does not involve any growth for the array, so it could
3129 // be of fixed size.
3130 void PhaseIdealLoop::clear_dom_lca_tags() {
3131 uint limit = C->unique() + 1;
3132 _dom_lca_tags.map( limit, NULL );
3133 _dom_lca_tags.clear();
3134 #ifdef ASSERT
3135 for( uint i = 0; i < limit; ++i ) {
3136 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3137 }
3138 #endif // ASSERT
3139 }
3141 //------------------------------build_loop_late--------------------------------
3142 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3143 // Second pass finds latest legal placement, and ideal loop placement.
3144 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3145 while (worklist.size() != 0) {
3146 Node *n = worklist.pop();
3147 // Only visit once
3148 if (visited.test_set(n->_idx)) continue;
3149 uint cnt = n->outcnt();
3150 uint i = 0;
3151 while (true) {
3152 assert( _nodes[n->_idx], "no dead nodes" );
3153 // Visit all children
3154 if (i < cnt) {
3155 Node* use = n->raw_out(i);
3156 ++i;
3157 // Check for dead uses. Aggressively prune such junk. It might be
3158 // dead in the global sense, but still have local uses so I cannot
3159 // easily call 'remove_dead_node'.
3160 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3161 // Due to cycles, we might not hit the same fixed point in the verify
3162 // pass as we do in the regular pass. Instead, visit such phis as
3163 // simple uses of the loop head.
3164 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3165 if( !visited.test(use->_idx) )
3166 worklist.push(use);
3167 } else if( !visited.test_set(use->_idx) ) {
3168 nstack.push(n, i); // Save parent and next use's index.
3169 n = use; // Process all children of current use.
3170 cnt = use->outcnt();
3171 i = 0;
3172 }
3173 } else {
3174 // Do not visit around the backedge of loops via data edges.
3175 // push dead code onto a worklist
3176 _deadlist.push(use);
3177 }
3178 } else {
3179 // All of n's children have been processed, complete post-processing.
3180 build_loop_late_post(n);
3181 if (nstack.is_empty()) {
3182 // Finished all nodes on stack.
3183 // Process next node on the worklist.
3184 break;
3185 }
3186 // Get saved parent node and next use's index. Visit the rest of uses.
3187 n = nstack.node();
3188 cnt = n->outcnt();
3189 i = nstack.index();
3190 nstack.pop();
3191 }
3192 }
3193 }
3194 }
3196 //------------------------------build_loop_late_post---------------------------
3197 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3198 // Second pass finds latest legal placement, and ideal loop placement.
3199 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3201 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3202 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3203 }
3205 // CFG and pinned nodes already handled
3206 if( n->in(0) ) {
3207 if( n->in(0)->is_top() ) return; // Dead?
3209 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3210 // _must_ be pinned (they have to observe their control edge of course).
3211 // Unlike Stores (which modify an unallocable resource, the memory
3212 // state), Mods/Loads can float around. So free them up.
3213 bool pinned = true;
3214 switch( n->Opcode() ) {
3215 case Op_DivI:
3216 case Op_DivF:
3217 case Op_DivD:
3218 case Op_ModI:
3219 case Op_ModF:
3220 case Op_ModD:
3221 case Op_LoadB: // Same with Loads; they can sink
3222 case Op_LoadUB: // during loop optimizations.
3223 case Op_LoadUS:
3224 case Op_LoadD:
3225 case Op_LoadF:
3226 case Op_LoadI:
3227 case Op_LoadKlass:
3228 case Op_LoadNKlass:
3229 case Op_LoadL:
3230 case Op_LoadS:
3231 case Op_LoadP:
3232 case Op_LoadN:
3233 case Op_LoadRange:
3234 case Op_LoadD_unaligned:
3235 case Op_LoadL_unaligned:
3236 case Op_StrComp: // Does a bunch of load-like effects
3237 case Op_StrEquals:
3238 case Op_StrIndexOf:
3239 case Op_AryEq:
3240 pinned = false;
3241 }
3242 if( pinned ) {
3243 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3244 if( !chosen_loop->_child ) // Inner loop?
3245 chosen_loop->_body.push(n); // Collect inner loops
3246 return;
3247 }
3248 } else { // No slot zero
3249 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3250 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3251 return;
3252 }
3253 assert(!n->is_CFG() || n->outcnt() == 0, "");
3254 }
3256 // Do I have a "safe range" I can select over?
3257 Node *early = get_ctrl(n);// Early location already computed
3259 // Compute latest point this Node can go
3260 Node *LCA = get_late_ctrl( n, early );
3261 // LCA is NULL due to uses being dead
3262 if( LCA == NULL ) {
3263 #ifdef ASSERT
3264 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3265 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3266 }
3267 #endif
3268 _nodes.map(n->_idx, 0); // This node is useless
3269 _deadlist.push(n);
3270 return;
3271 }
3272 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3274 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3275 Node *least = legal; // Best legal position so far
3276 while( early != legal ) { // While not at earliest legal
3277 #ifdef ASSERT
3278 if (legal->is_Start() && !early->is_Root()) {
3279 // Bad graph. Print idom path and fail.
3280 dump_bad_graph(n, early, LCA);
3281 assert(false, "Bad graph detected in build_loop_late");
3282 }
3283 #endif
3284 // Find least loop nesting depth
3285 legal = idom(legal); // Bump up the IDOM tree
3286 // Check for lower nesting depth
3287 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3288 least = legal;
3289 }
3290 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3292 // Try not to place code on a loop entry projection
3293 // which can inhibit range check elimination.
3294 if (least != early) {
3295 Node* ctrl_out = least->unique_ctrl_out();
3296 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3297 least == ctrl_out->in(LoopNode::EntryControl)) {
3298 Node* least_dom = idom(least);
3299 if (get_loop(least_dom)->is_member(get_loop(least))) {
3300 least = least_dom;
3301 }
3302 }
3303 }
3305 #ifdef ASSERT
3306 // If verifying, verify that 'verify_me' has a legal location
3307 // and choose it as our location.
3308 if( _verify_me ) {
3309 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3310 Node *legal = LCA;
3311 while( early != legal ) { // While not at earliest legal
3312 if( legal == v_ctrl ) break; // Check for prior good location
3313 legal = idom(legal) ;// Bump up the IDOM tree
3314 }
3315 // Check for prior good location
3316 if( legal == v_ctrl ) least = legal; // Keep prior if found
3317 }
3318 #endif
3320 // Assign discovered "here or above" point
3321 least = find_non_split_ctrl(least);
3322 set_ctrl(n, least);
3324 // Collect inner loop bodies
3325 IdealLoopTree *chosen_loop = get_loop(least);
3326 if( !chosen_loop->_child ) // Inner loop?
3327 chosen_loop->_body.push(n);// Collect inner loops
3328 }
3330 #ifdef ASSERT
3331 void PhaseIdealLoop::dump_bad_graph(Node* n, Node* early, Node* LCA) {
3332 tty->print_cr( "Bad graph detected in build_loop_late");
3333 tty->print("n: "); n->dump();
3334 tty->print("early(n): "); early->dump();
3335 if (n->in(0) != NULL && !n->in(0)->is_top() &&
3336 n->in(0) != early && !n->in(0)->is_Root()) {
3337 tty->print("n->in(0): "); n->in(0)->dump();
3338 }
3339 for (uint i = 1; i < n->req(); i++) {
3340 Node* in1 = n->in(i);
3341 if (in1 != NULL && in1 != n && !in1->is_top()) {
3342 tty->print("n->in(%d): ", i); in1->dump();
3343 Node* in1_early = get_ctrl(in1);
3344 tty->print("early(n->in(%d)): ", i); in1_early->dump();
3345 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
3346 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3347 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3348 }
3349 for (uint j = 1; j < in1->req(); j++) {
3350 Node* in2 = in1->in(j);
3351 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3352 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3353 Node* in2_early = get_ctrl(in2);
3354 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3355 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
3356 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3357 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3358 }
3359 }
3360 }
3361 }
3362 }
3363 tty->cr();
3364 tty->print("LCA(n): "); LCA->dump();
3365 for (uint i = 0; i < n->outcnt(); i++) {
3366 Node* u1 = n->raw_out(i);
3367 if (u1 == n)
3368 continue;
3369 tty->print("n->out(%d): ", i); u1->dump();
3370 if (u1->is_CFG()) {
3371 for (uint j = 0; j < u1->outcnt(); j++) {
3372 Node* u2 = u1->raw_out(j);
3373 if (u2 != u1 && u2 != n && u2->is_CFG()) {
3374 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3375 }
3376 }
3377 } else {
3378 Node* u1_later = get_ctrl(u1);
3379 tty->print("later(n->out(%d)): ", i); u1_later->dump();
3380 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
3381 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3382 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3383 }
3384 for (uint j = 0; j < u1->outcnt(); j++) {
3385 Node* u2 = u1->raw_out(j);
3386 if (u2 == n || u2 == u1)
3387 continue;
3388 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3389 if (!u2->is_CFG()) {
3390 Node* u2_later = get_ctrl(u2);
3391 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3392 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
3393 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3394 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3395 }
3396 }
3397 }
3398 }
3399 }
3400 tty->cr();
3401 int ct = 0;
3402 Node *dbg_legal = LCA;
3403 while(!dbg_legal->is_Start() && ct < 100) {
3404 tty->print("idom[%d] ",ct); dbg_legal->dump();
3405 ct++;
3406 dbg_legal = idom(dbg_legal);
3407 }
3408 tty->cr();
3409 }
3410 #endif
3412 #ifndef PRODUCT
3413 //------------------------------dump-------------------------------------------
3414 void PhaseIdealLoop::dump( ) const {
3415 ResourceMark rm;
3416 Arena* arena = Thread::current()->resource_area();
3417 Node_Stack stack(arena, C->unique() >> 2);
3418 Node_List rpo_list;
3419 VectorSet visited(arena);
3420 visited.set(C->top()->_idx);
3421 rpo( C->root(), stack, visited, rpo_list );
3422 // Dump root loop indexed by last element in PO order
3423 dump( _ltree_root, rpo_list.size(), rpo_list );
3424 }
3426 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3427 loop->dump_head();
3429 // Now scan for CFG nodes in the same loop
3430 for( uint j=idx; j > 0; j-- ) {
3431 Node *n = rpo_list[j-1];
3432 if( !_nodes[n->_idx] ) // Skip dead nodes
3433 continue;
3434 if( get_loop(n) != loop ) { // Wrong loop nest
3435 if( get_loop(n)->_head == n && // Found nested loop?
3436 get_loop(n)->_parent == loop )
3437 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3438 continue;
3439 }
3441 // Dump controlling node
3442 for( uint x = 0; x < loop->_nest; x++ )
3443 tty->print(" ");
3444 tty->print("C");
3445 if( n == C->root() ) {
3446 n->dump();
3447 } else {
3448 Node* cached_idom = idom_no_update(n);
3449 Node *computed_idom = n->in(0);
3450 if( n->is_Region() ) {
3451 computed_idom = compute_idom(n);
3452 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3453 // any MultiBranch ctrl node), so apply a similar transform to
3454 // the cached idom returned from idom_no_update.
3455 cached_idom = find_non_split_ctrl(cached_idom);
3456 }
3457 tty->print(" ID:%d",computed_idom->_idx);
3458 n->dump();
3459 if( cached_idom != computed_idom ) {
3460 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3461 computed_idom->_idx, cached_idom->_idx);
3462 }
3463 }
3464 // Dump nodes it controls
3465 for( uint k = 0; k < _nodes.Size(); k++ ) {
3466 // (k < C->unique() && get_ctrl(find(k)) == n)
3467 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3468 Node *m = C->root()->find(k);
3469 if( m && m->outcnt() > 0 ) {
3470 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3471 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3472 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3473 }
3474 for( uint j = 0; j < loop->_nest; j++ )
3475 tty->print(" ");
3476 tty->print(" ");
3477 m->dump();
3478 }
3479 }
3480 }
3481 }
3482 }
3484 // Collect a R-P-O for the whole CFG.
3485 // Result list is in post-order (scan backwards for RPO)
3486 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3487 stk.push(start, 0);
3488 visited.set(start->_idx);
3490 while (stk.is_nonempty()) {
3491 Node* m = stk.node();
3492 uint idx = stk.index();
3493 if (idx < m->outcnt()) {
3494 stk.set_index(idx + 1);
3495 Node* n = m->raw_out(idx);
3496 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3497 stk.push(n, 0);
3498 }
3499 } else {
3500 rpo_list.push(m);
3501 stk.pop();
3502 }
3503 }
3504 }
3505 #endif
3508 //=============================================================================
3509 //------------------------------LoopTreeIterator-----------------------------------
3511 // Advance to next loop tree using a preorder, left-to-right traversal.
3512 void LoopTreeIterator::next() {
3513 assert(!done(), "must not be done.");
3514 if (_curnt->_child != NULL) {
3515 _curnt = _curnt->_child;
3516 } else if (_curnt->_next != NULL) {
3517 _curnt = _curnt->_next;
3518 } else {
3519 while (_curnt != _root && _curnt->_next == NULL) {
3520 _curnt = _curnt->_parent;
3521 }
3522 if (_curnt == _root) {
3523 _curnt = NULL;
3524 assert(done(), "must be done.");
3525 } else {
3526 assert(_curnt->_next != NULL, "must be more to do");
3527 _curnt = _curnt->_next;
3528 }
3529 }
3530 }