Wed, 20 Apr 2011 18:29:35 -0700
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
Summary: memory slices should be always created for non-static fields after allocation
Reviewed-by: never
1 /*
2 * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
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 assert(stride_con != 0, "missed some peephole opt");
264 if (!xphi->is_Phi())
265 return false; // Too much math on the trip counter
266 if (phi_incr != NULL && phi_incr != xphi)
267 return false;
268 PhiNode *phi = xphi->as_Phi();
270 // Phi must be of loop header; backedge must wrap to increment
271 if (phi->region() != x)
272 return false;
273 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
274 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
275 return false;
276 }
277 Node *init_trip = phi->in(LoopNode::EntryControl);
279 // If iv trunc type is smaller than int, check for possible wrap.
280 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
281 assert(trunc1 != NULL, "must have found some truncation");
283 // Get a better type for the phi (filtered thru if's)
284 const TypeInt* phi_ft = filtered_type(phi);
286 // Can iv take on a value that will wrap?
287 //
288 // Ensure iv's limit is not within "stride" of the wrap value.
289 //
290 // Example for "short" type
291 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
292 // If the stride is +10, then the last value of the induction
293 // variable before the increment (phi_ft->_hi) must be
294 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
295 // ensure no truncation occurs after the increment.
297 if (stride_con > 0) {
298 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
299 iv_trunc_t->_lo > phi_ft->_lo) {
300 return false; // truncation may occur
301 }
302 } else if (stride_con < 0) {
303 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
304 iv_trunc_t->_hi < phi_ft->_hi) {
305 return false; // truncation may occur
306 }
307 }
308 // No possibility of wrap so truncation can be discarded
309 // Promote iv type to Int
310 } else {
311 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
312 }
314 // If the condition is inverted and we will be rolling
315 // through MININT to MAXINT, then bail out.
316 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
317 // Odd stride
318 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
319 // Count down loop rolls through MAXINT
320 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
321 // Count up loop rolls through MININT
322 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0 ) {
323 return false; // Bail out
324 }
326 const TypeInt* init_t = gvn->type(init_trip)->is_int();
327 const TypeInt* limit_t = gvn->type(limit)->is_int();
329 if (stride_con > 0) {
330 long init_p = (long)init_t->_lo + stride_con;
331 if (init_p > (long)max_jint || init_p > (long)limit_t->_hi)
332 return false; // cyclic loop or this loop trips only once
333 } else {
334 long init_p = (long)init_t->_hi + stride_con;
335 if (init_p < (long)min_jint || init_p < (long)limit_t->_lo)
336 return false; // cyclic loop or this loop trips only once
337 }
339 // =================================================
340 // ---- SUCCESS! Found A Trip-Counted Loop! -----
341 //
342 assert(x->Opcode() == Op_Loop, "regular loops only");
343 C->print_method("Before CountedLoop", 3);
344 #ifndef PRODUCT
345 if (TraceLoopOpts) {
346 tty->print("Counted ");
347 loop->dump_head();
348 }
349 #endif
350 // If compare points to incr, we are ok. Otherwise the compare
351 // can directly point to the phi; in this case adjust the compare so that
352 // it points to the incr by adjusting the limit.
353 if (cmp->in(1) == phi || cmp->in(2) == phi)
354 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
356 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
357 // Final value for iterator should be: trip_count * stride + init_trip.
358 Node *one_p = gvn->intcon( 1);
359 Node *one_m = gvn->intcon(-1);
361 Node *trip_count = NULL;
362 Node *hook = new (C, 6) Node(6);
363 switch( bt ) {
364 case BoolTest::eq:
365 ShouldNotReachHere();
366 case BoolTest::ne: // Ahh, the case we desire
367 if (stride_con == 1)
368 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
369 else if (stride_con == -1)
370 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
371 else
372 ShouldNotReachHere();
373 set_subtree_ctrl(trip_count);
374 //_loop.map(trip_count->_idx,loop(limit));
375 break;
376 case BoolTest::le: // Maybe convert to '<' case
377 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
378 set_subtree_ctrl( limit );
379 hook->init_req(4, limit);
381 bt = BoolTest::lt;
382 // Make the new limit be in the same loop nest as the old limit
383 //_loop.map(limit->_idx,limit_loop);
384 // Fall into next case
385 case BoolTest::lt: { // Maybe convert to '!=' case
386 if (stride_con < 0) // Count down loop rolls through MAXINT
387 ShouldNotReachHere();
388 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
389 set_subtree_ctrl( range );
390 hook->init_req(0, range);
392 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
393 set_subtree_ctrl( bias );
394 hook->init_req(1, bias);
396 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
397 set_subtree_ctrl( bias1 );
398 hook->init_req(2, bias1);
400 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
401 set_subtree_ctrl( trip_count );
402 hook->init_req(3, trip_count);
403 break;
404 }
406 case BoolTest::ge: // Maybe convert to '>' case
407 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
408 set_subtree_ctrl( limit );
409 hook->init_req(4 ,limit);
411 bt = BoolTest::gt;
412 // Make the new limit be in the same loop nest as the old limit
413 //_loop.map(limit->_idx,limit_loop);
414 // Fall into next case
415 case BoolTest::gt: { // Maybe convert to '!=' case
416 if (stride_con > 0) // count up loop rolls through MININT
417 ShouldNotReachHere();
418 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
419 set_subtree_ctrl( range );
420 hook->init_req(0, range);
422 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
423 set_subtree_ctrl( bias );
424 hook->init_req(1, bias);
426 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
427 set_subtree_ctrl( bias1 );
428 hook->init_req(2, bias1);
430 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
431 set_subtree_ctrl( trip_count );
432 hook->init_req(3, trip_count);
433 break;
434 }
435 } // switch( bt )
437 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
438 set_subtree_ctrl( span );
439 hook->init_req(5, span);
441 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
442 set_subtree_ctrl( limit );
444 // Check for SafePoint on backedge and remove
445 Node *sfpt = x->in(LoopNode::LoopBackControl);
446 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
447 lazy_replace( sfpt, iftrue );
448 loop->_tail = iftrue;
449 }
451 // Build a canonical trip test.
452 // Clone code, as old values may be in use.
453 Node* nphi = PhiNode::make(x, init_trip, TypeInt::INT);
454 nphi = _igvn.register_new_node_with_optimizer(nphi);
455 set_ctrl(nphi, get_ctrl(phi));
457 incr = incr->clone();
458 incr->set_req(1,nphi);
459 incr->set_req(2,stride);
460 incr = _igvn.register_new_node_with_optimizer(incr);
461 set_early_ctrl( incr );
463 nphi->set_req(LoopNode::LoopBackControl, incr);
464 _igvn.replace_node(phi, nphi);
465 phi = nphi->as_Phi();
467 cmp = cmp->clone();
468 cmp->set_req(1,incr);
469 cmp->set_req(2,limit);
470 cmp = _igvn.register_new_node_with_optimizer(cmp);
471 set_ctrl(cmp, iff->in(0));
473 test = test->clone()->as_Bool();
474 (*(BoolTest*)&test->_test)._test = bt;
475 test->set_req(1,cmp);
476 _igvn.register_new_node_with_optimizer(test);
477 set_ctrl(test, iff->in(0));
479 // Replace the old IfNode with a new LoopEndNode
480 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
481 IfNode *le = lex->as_If();
482 uint dd = dom_depth(iff);
483 set_idom(le, le->in(0), dd); // Update dominance for loop exit
484 set_loop(le, loop);
486 // Get the loop-exit control
487 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
489 // Need to swap loop-exit and loop-back control?
490 if (iftrue_op == Op_IfFalse) {
491 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
492 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
494 loop->_tail = back_control = ift2;
495 set_loop(ift2, loop);
496 set_loop(iff2, get_loop(iffalse));
498 // Lazy update of 'get_ctrl' mechanism.
499 lazy_replace_proj( iffalse, iff2 );
500 lazy_replace_proj( iftrue, ift2 );
502 // Swap names
503 iffalse = iff2;
504 iftrue = ift2;
505 } else {
506 _igvn.hash_delete(iffalse);
507 _igvn.hash_delete(iftrue);
508 iffalse->set_req_X( 0, le, &_igvn );
509 iftrue ->set_req_X( 0, le, &_igvn );
510 }
512 set_idom(iftrue, le, dd+1);
513 set_idom(iffalse, le, dd+1);
514 assert(iff->outcnt() == 0, "should be dead now");
515 lazy_replace( iff, le ); // fix 'get_ctrl'
517 // Now setup a new CountedLoopNode to replace the existing LoopNode
518 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
519 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
520 // The following assert is approximately true, and defines the intention
521 // of can_be_counted_loop. It fails, however, because phase->type
522 // is not yet initialized for this loop and its parts.
523 //assert(l->can_be_counted_loop(this), "sanity");
524 _igvn.register_new_node_with_optimizer(l);
525 set_loop(l, loop);
526 loop->_head = l;
527 // Fix all data nodes placed at the old loop head.
528 // Uses the lazy-update mechanism of 'get_ctrl'.
529 lazy_replace( x, l );
530 set_idom(l, init_control, dom_depth(x));
532 // Check for immediately preceding SafePoint and remove
533 Node *sfpt2 = le->in(0);
534 if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
535 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
537 // Free up intermediate goo
538 _igvn.remove_dead_node(hook);
540 #ifdef ASSERT
541 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
542 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
543 #endif
545 C->print_method("After CountedLoop", 3);
547 return true;
548 }
551 //------------------------------Ideal------------------------------------------
552 // Return a node which is more "ideal" than the current node.
553 // Attempt to convert into a counted-loop.
554 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
555 if (!can_be_counted_loop(phase)) {
556 phase->C->set_major_progress();
557 }
558 return RegionNode::Ideal(phase, can_reshape);
559 }
562 //=============================================================================
563 //------------------------------Ideal------------------------------------------
564 // Return a node which is more "ideal" than the current node.
565 // Attempt to convert into a counted-loop.
566 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
567 return RegionNode::Ideal(phase, can_reshape);
568 }
570 //------------------------------dump_spec--------------------------------------
571 // Dump special per-node info
572 #ifndef PRODUCT
573 void CountedLoopNode::dump_spec(outputStream *st) const {
574 LoopNode::dump_spec(st);
575 if( stride_is_con() ) {
576 st->print("stride: %d ",stride_con());
577 } else {
578 st->print("stride: not constant ");
579 }
580 if( is_pre_loop () ) st->print("pre of N%d" , _main_idx );
581 if( is_main_loop() ) st->print("main of N%d", _idx );
582 if( is_post_loop() ) st->print("post of N%d", _main_idx );
583 }
584 #endif
586 //=============================================================================
587 int CountedLoopEndNode::stride_con() const {
588 return stride()->bottom_type()->is_int()->get_con();
589 }
592 //----------------------match_incr_with_optional_truncation--------------------
593 // Match increment with optional truncation:
594 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
595 // Return NULL for failure. Success returns the increment node.
596 Node* CountedLoopNode::match_incr_with_optional_truncation(
597 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
598 // Quick cutouts:
599 if (expr == NULL || expr->req() != 3) return false;
601 Node *t1 = NULL;
602 Node *t2 = NULL;
603 const TypeInt* trunc_t = TypeInt::INT;
604 Node* n1 = expr;
605 int n1op = n1->Opcode();
607 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
608 if (n1op == Op_AndI &&
609 n1->in(2)->is_Con() &&
610 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
611 // %%% This check should match any mask of 2**K-1.
612 t1 = n1;
613 n1 = t1->in(1);
614 n1op = n1->Opcode();
615 trunc_t = TypeInt::CHAR;
616 } else if (n1op == Op_RShiftI &&
617 n1->in(1) != NULL &&
618 n1->in(1)->Opcode() == Op_LShiftI &&
619 n1->in(2) == n1->in(1)->in(2) &&
620 n1->in(2)->is_Con()) {
621 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
622 // %%% This check should match any shift in [1..31].
623 if (shift == 16 || shift == 8) {
624 t1 = n1;
625 t2 = t1->in(1);
626 n1 = t2->in(1);
627 n1op = n1->Opcode();
628 if (shift == 16) {
629 trunc_t = TypeInt::SHORT;
630 } else if (shift == 8) {
631 trunc_t = TypeInt::BYTE;
632 }
633 }
634 }
636 // If (maybe after stripping) it is an AddI, we won:
637 if (n1op == Op_AddI) {
638 *trunc1 = t1;
639 *trunc2 = t2;
640 *trunc_type = trunc_t;
641 return n1;
642 }
644 // failed
645 return NULL;
646 }
649 //------------------------------filtered_type--------------------------------
650 // Return a type based on condition control flow
651 // A successful return will be a type that is restricted due
652 // to a series of dominating if-tests, such as:
653 // if (i < 10) {
654 // if (i > 0) {
655 // here: "i" type is [1..10)
656 // }
657 // }
658 // or a control flow merge
659 // if (i < 10) {
660 // do {
661 // phi( , ) -- at top of loop type is [min_int..10)
662 // i = ?
663 // } while ( i < 10)
664 //
665 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
666 assert(n && n->bottom_type()->is_int(), "must be int");
667 const TypeInt* filtered_t = NULL;
668 if (!n->is_Phi()) {
669 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
670 filtered_t = filtered_type_from_dominators(n, n_ctrl);
672 } else {
673 Node* phi = n->as_Phi();
674 Node* region = phi->in(0);
675 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
676 if (region && region != C->top()) {
677 for (uint i = 1; i < phi->req(); i++) {
678 Node* val = phi->in(i);
679 Node* use_c = region->in(i);
680 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
681 if (val_t != NULL) {
682 if (filtered_t == NULL) {
683 filtered_t = val_t;
684 } else {
685 filtered_t = filtered_t->meet(val_t)->is_int();
686 }
687 }
688 }
689 }
690 }
691 const TypeInt* n_t = _igvn.type(n)->is_int();
692 if (filtered_t != NULL) {
693 n_t = n_t->join(filtered_t)->is_int();
694 }
695 return n_t;
696 }
699 //------------------------------filtered_type_from_dominators--------------------------------
700 // Return a possibly more restrictive type for val based on condition control flow of dominators
701 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
702 if (val->is_Con()) {
703 return val->bottom_type()->is_int();
704 }
705 uint if_limit = 10; // Max number of dominating if's visited
706 const TypeInt* rtn_t = NULL;
708 if (use_ctrl && use_ctrl != C->top()) {
709 Node* val_ctrl = get_ctrl(val);
710 uint val_dom_depth = dom_depth(val_ctrl);
711 Node* pred = use_ctrl;
712 uint if_cnt = 0;
713 while (if_cnt < if_limit) {
714 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
715 if_cnt++;
716 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
717 if (if_t != NULL) {
718 if (rtn_t == NULL) {
719 rtn_t = if_t;
720 } else {
721 rtn_t = rtn_t->join(if_t)->is_int();
722 }
723 }
724 }
725 pred = idom(pred);
726 if (pred == NULL || pred == C->top()) {
727 break;
728 }
729 // Stop if going beyond definition block of val
730 if (dom_depth(pred) < val_dom_depth) {
731 break;
732 }
733 }
734 }
735 return rtn_t;
736 }
739 //------------------------------dump_spec--------------------------------------
740 // Dump special per-node info
741 #ifndef PRODUCT
742 void CountedLoopEndNode::dump_spec(outputStream *st) const {
743 if( in(TestValue)->is_Bool() ) {
744 BoolTest bt( test_trip()); // Added this for g++.
746 st->print("[");
747 bt.dump_on(st);
748 st->print("]");
749 }
750 st->print(" ");
751 IfNode::dump_spec(st);
752 }
753 #endif
755 //=============================================================================
756 //------------------------------is_member--------------------------------------
757 // Is 'l' a member of 'this'?
758 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
759 while( l->_nest > _nest ) l = l->_parent;
760 return l == this;
761 }
763 //------------------------------set_nest---------------------------------------
764 // Set loop tree nesting depth. Accumulate _has_call bits.
765 int IdealLoopTree::set_nest( uint depth ) {
766 _nest = depth;
767 int bits = _has_call;
768 if( _child ) bits |= _child->set_nest(depth+1);
769 if( bits ) _has_call = 1;
770 if( _next ) bits |= _next ->set_nest(depth );
771 return bits;
772 }
774 //------------------------------split_fall_in----------------------------------
775 // Split out multiple fall-in edges from the loop header. Move them to a
776 // private RegionNode before the loop. This becomes the loop landing pad.
777 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
778 PhaseIterGVN &igvn = phase->_igvn;
779 uint i;
781 // Make a new RegionNode to be the landing pad.
782 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
783 phase->set_loop(landing_pad,_parent);
784 // Gather all the fall-in control paths into the landing pad
785 uint icnt = fall_in_cnt;
786 uint oreq = _head->req();
787 for( i = oreq-1; i>0; i-- )
788 if( !phase->is_member( this, _head->in(i) ) )
789 landing_pad->set_req(icnt--,_head->in(i));
791 // Peel off PhiNode edges as well
792 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
793 Node *oj = _head->fast_out(j);
794 if( oj->is_Phi() ) {
795 PhiNode* old_phi = oj->as_Phi();
796 assert( old_phi->region() == _head, "" );
797 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
798 Node *p = PhiNode::make_blank(landing_pad, old_phi);
799 uint icnt = fall_in_cnt;
800 for( i = oreq-1; i>0; i-- ) {
801 if( !phase->is_member( this, _head->in(i) ) ) {
802 p->init_req(icnt--, old_phi->in(i));
803 // Go ahead and clean out old edges from old phi
804 old_phi->del_req(i);
805 }
806 }
807 // Search for CSE's here, because ZKM.jar does a lot of
808 // loop hackery and we need to be a little incremental
809 // with the CSE to avoid O(N^2) node blow-up.
810 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
811 if( p2 ) { // Found CSE
812 p->destruct(); // Recover useless new node
813 p = p2; // Use old node
814 } else {
815 igvn.register_new_node_with_optimizer(p, old_phi);
816 }
817 // Make old Phi refer to new Phi.
818 old_phi->add_req(p);
819 // Check for the special case of making the old phi useless and
820 // disappear it. In JavaGrande I have a case where this useless
821 // Phi is the loop limit and prevents recognizing a CountedLoop
822 // which in turn prevents removing an empty loop.
823 Node *id_old_phi = old_phi->Identity( &igvn );
824 if( id_old_phi != old_phi ) { // Found a simple identity?
825 // Note that I cannot call 'replace_node' here, because
826 // that will yank the edge from old_phi to the Region and
827 // I'm mid-iteration over the Region's uses.
828 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
829 Node* use = old_phi->last_out(i);
830 igvn.hash_delete(use);
831 igvn._worklist.push(use);
832 uint uses_found = 0;
833 for (uint j = 0; j < use->len(); j++) {
834 if (use->in(j) == old_phi) {
835 if (j < use->req()) use->set_req (j, id_old_phi);
836 else use->set_prec(j, id_old_phi);
837 uses_found++;
838 }
839 }
840 i -= uses_found; // we deleted 1 or more copies of this edge
841 }
842 }
843 igvn._worklist.push(old_phi);
844 }
845 }
846 // Finally clean out the fall-in edges from the RegionNode
847 for( i = oreq-1; i>0; i-- ) {
848 if( !phase->is_member( this, _head->in(i) ) ) {
849 _head->del_req(i);
850 }
851 }
852 // Transform landing pad
853 igvn.register_new_node_with_optimizer(landing_pad, _head);
854 // Insert landing pad into the header
855 _head->add_req(landing_pad);
856 }
858 //------------------------------split_outer_loop-------------------------------
859 // Split out the outermost loop from this shared header.
860 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
861 PhaseIterGVN &igvn = phase->_igvn;
863 // Find index of outermost loop; it should also be my tail.
864 uint outer_idx = 1;
865 while( _head->in(outer_idx) != _tail ) outer_idx++;
867 // Make a LoopNode for the outermost loop.
868 Node *ctl = _head->in(LoopNode::EntryControl);
869 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
870 outer = igvn.register_new_node_with_optimizer(outer, _head);
871 phase->set_created_loop_node();
873 Node* pred = phase->clone_loop_predicates(ctl, outer);
874 // Outermost loop falls into '_head' loop
875 _head->set_req(LoopNode::EntryControl, pred);
876 _head->del_req(outer_idx);
877 // Split all the Phis up between '_head' loop and 'outer' loop.
878 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
879 Node *out = _head->fast_out(j);
880 if( out->is_Phi() ) {
881 PhiNode *old_phi = out->as_Phi();
882 assert( old_phi->region() == _head, "" );
883 Node *phi = PhiNode::make_blank(outer, old_phi);
884 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
885 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
886 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
887 // Make old Phi point to new Phi on the fall-in path
888 igvn.hash_delete(old_phi);
889 old_phi->set_req(LoopNode::EntryControl, phi);
890 old_phi->del_req(outer_idx);
891 igvn._worklist.push(old_phi);
892 }
893 }
895 // Use the new loop head instead of the old shared one
896 _head = outer;
897 phase->set_loop(_head, this);
898 }
900 //------------------------------fix_parent-------------------------------------
901 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
902 loop->_parent = parent;
903 if( loop->_child ) fix_parent( loop->_child, loop );
904 if( loop->_next ) fix_parent( loop->_next , parent );
905 }
907 //------------------------------estimate_path_freq-----------------------------
908 static float estimate_path_freq( Node *n ) {
909 // Try to extract some path frequency info
910 IfNode *iff;
911 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
912 uint nop = n->Opcode();
913 if( nop == Op_SafePoint ) { // Skip any safepoint
914 n = n->in(0);
915 continue;
916 }
917 if( nop == Op_CatchProj ) { // Get count from a prior call
918 // Assume call does not always throw exceptions: means the call-site
919 // count is also the frequency of the fall-through path.
920 assert( n->is_CatchProj(), "" );
921 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
922 return 0.0f; // Assume call exception path is rare
923 Node *call = n->in(0)->in(0)->in(0);
924 assert( call->is_Call(), "expect a call here" );
925 const JVMState *jvms = ((CallNode*)call)->jvms();
926 ciMethodData* methodData = jvms->method()->method_data();
927 if (!methodData->is_mature()) return 0.0f; // No call-site data
928 ciProfileData* data = methodData->bci_to_data(jvms->bci());
929 if ((data == NULL) || !data->is_CounterData()) {
930 // no call profile available, try call's control input
931 n = n->in(0);
932 continue;
933 }
934 return data->as_CounterData()->count()/FreqCountInvocations;
935 }
936 // See if there's a gating IF test
937 Node *n_c = n->in(0);
938 if( !n_c->is_If() ) break; // No estimate available
939 iff = n_c->as_If();
940 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
941 // Compute how much count comes on this path
942 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
943 // Have no count info. Skip dull uncommon-trap like branches.
944 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
945 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
946 break;
947 // Skip through never-taken branch; look for a real loop exit.
948 n = iff->in(0);
949 }
950 return 0.0f; // No estimate available
951 }
953 //------------------------------merge_many_backedges---------------------------
954 // Merge all the backedges from the shared header into a private Region.
955 // Feed that region as the one backedge to this loop.
956 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
957 uint i;
959 // Scan for the top 2 hottest backedges
960 float hotcnt = 0.0f;
961 float warmcnt = 0.0f;
962 uint hot_idx = 0;
963 // Loop starts at 2 because slot 1 is the fall-in path
964 for( i = 2; i < _head->req(); i++ ) {
965 float cnt = estimate_path_freq(_head->in(i));
966 if( cnt > hotcnt ) { // Grab hottest path
967 warmcnt = hotcnt;
968 hotcnt = cnt;
969 hot_idx = i;
970 } else if( cnt > warmcnt ) { // And 2nd hottest path
971 warmcnt = cnt;
972 }
973 }
975 // See if the hottest backedge is worthy of being an inner loop
976 // by being much hotter than the next hottest backedge.
977 if( hotcnt <= 0.0001 ||
978 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
980 // Peel out the backedges into a private merge point; peel
981 // them all except optionally hot_idx.
982 PhaseIterGVN &igvn = phase->_igvn;
984 Node *hot_tail = NULL;
985 // Make a Region for the merge point
986 Node *r = new (phase->C, 1) RegionNode(1);
987 for( i = 2; i < _head->req(); i++ ) {
988 if( i != hot_idx )
989 r->add_req( _head->in(i) );
990 else hot_tail = _head->in(i);
991 }
992 igvn.register_new_node_with_optimizer(r, _head);
993 // Plug region into end of loop _head, followed by hot_tail
994 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
995 _head->set_req(2, r);
996 if( hot_idx ) _head->add_req(hot_tail);
998 // Split all the Phis up between '_head' loop and the Region 'r'
999 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1000 Node *out = _head->fast_out(j);
1001 if( out->is_Phi() ) {
1002 PhiNode* n = out->as_Phi();
1003 igvn.hash_delete(n); // Delete from hash before hacking edges
1004 Node *hot_phi = NULL;
1005 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
1006 // Check all inputs for the ones to peel out
1007 uint j = 1;
1008 for( uint i = 2; i < n->req(); i++ ) {
1009 if( i != hot_idx )
1010 phi->set_req( j++, n->in(i) );
1011 else hot_phi = n->in(i);
1012 }
1013 // Register the phi but do not transform until whole place transforms
1014 igvn.register_new_node_with_optimizer(phi, n);
1015 // Add the merge phi to the old Phi
1016 while( n->req() > 3 ) n->del_req( n->req()-1 );
1017 n->set_req(2, phi);
1018 if( hot_idx ) n->add_req(hot_phi);
1019 }
1020 }
1023 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1024 // of self loop tree. Turn self into a loop headed by _head and with
1025 // tail being the new merge point.
1026 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1027 phase->set_loop(_tail,ilt); // Adjust tail
1028 _tail = r; // Self's tail is new merge point
1029 phase->set_loop(r,this);
1030 ilt->_child = _child; // New guy has my children
1031 _child = ilt; // Self has new guy as only child
1032 ilt->_parent = this; // new guy has self for parent
1033 ilt->_nest = _nest; // Same nesting depth (for now)
1035 // Starting with 'ilt', look for child loop trees using the same shared
1036 // header. Flatten these out; they will no longer be loops in the end.
1037 IdealLoopTree **pilt = &_child;
1038 while( ilt ) {
1039 if( ilt->_head == _head ) {
1040 uint i;
1041 for( i = 2; i < _head->req(); i++ )
1042 if( _head->in(i) == ilt->_tail )
1043 break; // Still a loop
1044 if( i == _head->req() ) { // No longer a loop
1045 // Flatten ilt. Hang ilt's "_next" list from the end of
1046 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1047 IdealLoopTree **cp = &ilt->_child;
1048 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1049 *cp = ilt->_next; // Hang next list at end of child list
1050 *pilt = ilt->_child; // Move child up to replace ilt
1051 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1052 ilt = ilt->_child; // Repeat using new ilt
1053 continue; // do not advance over ilt->_child
1054 }
1055 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1056 phase->set_loop(_head,ilt);
1057 }
1058 pilt = &ilt->_child; // Advance to next
1059 ilt = *pilt;
1060 }
1062 if( _child ) fix_parent( _child, this );
1063 }
1065 //------------------------------beautify_loops---------------------------------
1066 // Split shared headers and insert loop landing pads.
1067 // Insert a LoopNode to replace the RegionNode.
1068 // Return TRUE if loop tree is structurally changed.
1069 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1070 bool result = false;
1071 // Cache parts in locals for easy
1072 PhaseIterGVN &igvn = phase->_igvn;
1074 igvn.hash_delete(_head); // Yank from hash before hacking edges
1076 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1077 int fall_in_cnt = 0;
1078 for( uint i = 1; i < _head->req(); i++ )
1079 if( !phase->is_member( this, _head->in(i) ) )
1080 fall_in_cnt++;
1081 assert( fall_in_cnt, "at least 1 fall-in path" );
1082 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1083 split_fall_in( phase, fall_in_cnt );
1085 // Swap inputs to the _head and all Phis to move the fall-in edge to
1086 // the left.
1087 fall_in_cnt = 1;
1088 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1089 fall_in_cnt++;
1090 if( fall_in_cnt > 1 ) {
1091 // Since I am just swapping inputs I do not need to update def-use info
1092 Node *tmp = _head->in(1);
1093 _head->set_req( 1, _head->in(fall_in_cnt) );
1094 _head->set_req( fall_in_cnt, tmp );
1095 // Swap also all Phis
1096 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1097 Node* phi = _head->fast_out(i);
1098 if( phi->is_Phi() ) {
1099 igvn.hash_delete(phi); // Yank from hash before hacking edges
1100 tmp = phi->in(1);
1101 phi->set_req( 1, phi->in(fall_in_cnt) );
1102 phi->set_req( fall_in_cnt, tmp );
1103 }
1104 }
1105 }
1106 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1107 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1109 // If I am a shared header (multiple backedges), peel off the many
1110 // backedges into a private merge point and use the merge point as
1111 // the one true backedge.
1112 if( _head->req() > 3 ) {
1113 // Merge the many backedges into a single backedge but leave
1114 // the hottest backedge as separate edge for the following peel.
1115 merge_many_backedges( phase );
1116 result = true;
1117 }
1119 // If I have one hot backedge, peel off myself loop.
1120 // I better be the outermost loop.
1121 if( _head->req() > 3 ) {
1122 split_outer_loop( phase );
1123 result = true;
1125 } else if( !_head->is_Loop() && !_irreducible ) {
1126 // Make a new LoopNode to replace the old loop head
1127 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1128 l = igvn.register_new_node_with_optimizer(l, _head);
1129 phase->set_created_loop_node();
1130 // Go ahead and replace _head
1131 phase->_igvn.replace_node( _head, l );
1132 _head = l;
1133 phase->set_loop(_head, this);
1134 }
1136 // Now recursively beautify nested loops
1137 if( _child ) result |= _child->beautify_loops( phase );
1138 if( _next ) result |= _next ->beautify_loops( phase );
1139 return result;
1140 }
1142 //------------------------------allpaths_check_safepts----------------------------
1143 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1144 // encountered. Helper for check_safepts.
1145 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1146 assert(stack.size() == 0, "empty stack");
1147 stack.push(_tail);
1148 visited.Clear();
1149 visited.set(_tail->_idx);
1150 while (stack.size() > 0) {
1151 Node* n = stack.pop();
1152 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1153 // Terminate this path
1154 } else if (n->Opcode() == Op_SafePoint) {
1155 if (_phase->get_loop(n) != this) {
1156 if (_required_safept == NULL) _required_safept = new Node_List();
1157 _required_safept->push(n); // save the one closest to the tail
1158 }
1159 // Terminate this path
1160 } else {
1161 uint start = n->is_Region() ? 1 : 0;
1162 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1163 for (uint i = start; i < end; i++) {
1164 Node* in = n->in(i);
1165 assert(in->is_CFG(), "must be");
1166 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1167 stack.push(in);
1168 }
1169 }
1170 }
1171 }
1172 }
1174 //------------------------------check_safepts----------------------------
1175 // Given dominators, try to find loops with calls that must always be
1176 // executed (call dominates loop tail). These loops do not need non-call
1177 // safepoints (ncsfpt).
1178 //
1179 // A complication is that a safepoint in a inner loop may be needed
1180 // by an outer loop. In the following, the inner loop sees it has a
1181 // call (block 3) on every path from the head (block 2) to the
1182 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1183 // in block 2, _but_ this leaves the outer loop without a safepoint.
1184 //
1185 // entry 0
1186 // |
1187 // v
1188 // outer 1,2 +->1
1189 // | |
1190 // | v
1191 // | 2<---+ ncsfpt in 2
1192 // |_/|\ |
1193 // | v |
1194 // inner 2,3 / 3 | call in 3
1195 // / | |
1196 // v +--+
1197 // exit 4
1198 //
1199 //
1200 // This method creates a list (_required_safept) of ncsfpt nodes that must
1201 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1202 // is first looked for in the lists for the outer loops of the current loop.
1203 //
1204 // The insights into the problem:
1205 // A) counted loops are okay
1206 // B) innermost loops are okay (only an inner loop can delete
1207 // a ncsfpt needed by an outer loop)
1208 // C) a loop is immune from an inner loop deleting a safepoint
1209 // if the loop has a call on the idom-path
1210 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1211 // idom-path that is not in a nested loop
1212 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1213 // loop needs to be prevented from deletion by an inner loop
1214 //
1215 // There are two analyses:
1216 // 1) The first, and cheaper one, scans the loop body from
1217 // tail to head following the idom (immediate dominator)
1218 // chain, looking for the cases (C,D,E) above.
1219 // Since inner loops are scanned before outer loops, there is summary
1220 // information about inner loops. Inner loops can be skipped over
1221 // when the tail of an inner loop is encountered.
1222 //
1223 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1224 // the idom path (which is rare), scans all predecessor control paths
1225 // from the tail to the head, terminating a path when a call or sfpt
1226 // is encountered, to find the ncsfpt's that are closest to the tail.
1227 //
1228 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1229 // Bottom up traversal
1230 IdealLoopTree* ch = _child;
1231 while (ch != NULL) {
1232 ch->check_safepts(visited, stack);
1233 ch = ch->_next;
1234 }
1236 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1237 bool has_call = false; // call on dom-path
1238 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1239 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1240 // Scan the dom-path nodes from tail to head
1241 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1242 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1243 has_call = true;
1244 _has_sfpt = 1; // Then no need for a safept!
1245 break;
1246 } else if (n->Opcode() == Op_SafePoint) {
1247 if (_phase->get_loop(n) == this) {
1248 has_local_ncsfpt = true;
1249 break;
1250 }
1251 if (nonlocal_ncsfpt == NULL) {
1252 nonlocal_ncsfpt = n; // save the one closest to the tail
1253 }
1254 } else {
1255 IdealLoopTree* nlpt = _phase->get_loop(n);
1256 if (this != nlpt) {
1257 // If at an inner loop tail, see if the inner loop has already
1258 // recorded seeing a call on the dom-path (and stop.) If not,
1259 // jump to the head of the inner loop.
1260 assert(is_member(nlpt), "nested loop");
1261 Node* tail = nlpt->_tail;
1262 if (tail->in(0)->is_If()) tail = tail->in(0);
1263 if (n == tail) {
1264 // If inner loop has call on dom-path, so does outer loop
1265 if (nlpt->_has_sfpt) {
1266 has_call = true;
1267 _has_sfpt = 1;
1268 break;
1269 }
1270 // Skip to head of inner loop
1271 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1272 n = nlpt->_head;
1273 }
1274 }
1275 }
1276 }
1277 // Record safept's that this loop needs preserved when an
1278 // inner loop attempts to delete it's safepoints.
1279 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1280 if (nonlocal_ncsfpt != NULL) {
1281 if (_required_safept == NULL) _required_safept = new Node_List();
1282 _required_safept->push(nonlocal_ncsfpt);
1283 } else {
1284 // Failed to find a suitable safept on the dom-path. Now use
1285 // an all paths walk from tail to head, looking for safepoints to preserve.
1286 allpaths_check_safepts(visited, stack);
1287 }
1288 }
1289 }
1290 }
1292 //---------------------------is_deleteable_safept----------------------------
1293 // Is safept not required by an outer loop?
1294 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1295 assert(sfpt->Opcode() == Op_SafePoint, "");
1296 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1297 while (lp != NULL) {
1298 Node_List* sfpts = lp->_required_safept;
1299 if (sfpts != NULL) {
1300 for (uint i = 0; i < sfpts->size(); i++) {
1301 if (sfpt == sfpts->at(i))
1302 return false;
1303 }
1304 }
1305 lp = lp->_parent;
1306 }
1307 return true;
1308 }
1310 //---------------------------replace_parallel_iv-------------------------------
1311 // Replace parallel induction variable (parallel to trip counter)
1312 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1313 assert(loop->_head->is_CountedLoop(), "");
1314 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1315 Node *incr = cl->incr();
1316 if (incr == NULL)
1317 return; // Dead loop?
1318 Node *init = cl->init_trip();
1319 Node *phi = cl->phi();
1320 // protect against stride not being a constant
1321 if (!cl->stride_is_con())
1322 return;
1323 int stride_con = cl->stride_con();
1325 PhaseGVN *gvn = &_igvn;
1327 // Visit all children, looking for Phis
1328 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1329 Node *out = cl->out(i);
1330 // Look for other phis (secondary IVs). Skip dead ones
1331 if (!out->is_Phi() || out == phi || !has_node(out))
1332 continue;
1333 PhiNode* phi2 = out->as_Phi();
1334 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1335 // Look for induction variables of the form: X += constant
1336 if (phi2->region() != loop->_head ||
1337 incr2->req() != 3 ||
1338 incr2->in(1) != phi2 ||
1339 incr2 == incr ||
1340 incr2->Opcode() != Op_AddI ||
1341 !incr2->in(2)->is_Con())
1342 continue;
1344 // Check for parallel induction variable (parallel to trip counter)
1345 // via an affine function. In particular, count-down loops with
1346 // count-up array indices are common. We only RCE references off
1347 // the trip-counter, so we need to convert all these to trip-counter
1348 // expressions.
1349 Node *init2 = phi2->in( LoopNode::EntryControl );
1350 int stride_con2 = incr2->in(2)->get_int();
1352 // The general case here gets a little tricky. We want to find the
1353 // GCD of all possible parallel IV's and make a new IV using this
1354 // GCD for the loop. Then all possible IVs are simple multiples of
1355 // the GCD. In practice, this will cover very few extra loops.
1356 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1357 // where +/-1 is the common case, but other integer multiples are
1358 // also easy to handle.
1359 int ratio_con = stride_con2/stride_con;
1361 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1362 // Convert to using the trip counter. The parallel induction
1363 // variable differs from the trip counter by a loop-invariant
1364 // amount, the difference between their respective initial values.
1365 // It is scaled by the 'ratio_con'.
1366 // Perform local Ideal transformation since in most cases ratio == 1.
1367 Node* ratio = _igvn.intcon(ratio_con);
1368 set_ctrl(ratio, C->root());
1369 Node* hook = new (C, 3) Node(3);
1370 Node* ratio_init = gvn->transform(new (C, 3) MulINode(init, ratio));
1371 hook->init_req(0, ratio_init);
1372 Node* diff = gvn->transform(new (C, 3) SubINode(init2, ratio_init));
1373 hook->init_req(1, diff);
1374 Node* ratio_idx = gvn->transform(new (C, 3) MulINode(phi, ratio));
1375 hook->init_req(2, ratio_idx);
1376 Node* add = gvn->transform(new (C, 3) AddINode(ratio_idx, diff));
1377 set_subtree_ctrl(add);
1378 _igvn.replace_node( phi2, add );
1379 // Free up intermediate goo
1380 _igvn.remove_dead_node(hook);
1381 // Sometimes an induction variable is unused
1382 if (add->outcnt() == 0) {
1383 _igvn.remove_dead_node(add);
1384 }
1385 --i; // deleted this phi; rescan starting with next position
1386 continue;
1387 }
1388 }
1389 }
1391 //------------------------------counted_loop-----------------------------------
1392 // Convert to counted loops where possible
1393 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1395 // For grins, set the inner-loop flag here
1396 if (!_child) {
1397 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1398 }
1400 if (_head->is_CountedLoop() ||
1401 phase->is_counted_loop(_head, this)) {
1402 _has_sfpt = 1; // Indicate we do not need a safepoint here
1404 // Look for a safepoint to remove
1405 for (Node* n = tail(); n != _head; n = phase->idom(n))
1406 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1407 phase->is_deleteable_safept(n))
1408 phase->lazy_replace(n,n->in(TypeFunc::Control));
1410 // Look for induction variables
1411 phase->replace_parallel_iv(this);
1413 } else if (_parent != NULL && !_irreducible) {
1414 // Not a counted loop.
1415 // Look for a safepoint on the idom-path to remove, preserving the first one
1416 bool found = false;
1417 Node* n = tail();
1418 for (; n != _head && !found; n = phase->idom(n)) {
1419 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1420 found = true; // Found one
1421 }
1422 // Skip past it and delete the others
1423 for (; n != _head; n = phase->idom(n)) {
1424 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1425 phase->is_deleteable_safept(n))
1426 phase->lazy_replace(n,n->in(TypeFunc::Control));
1427 }
1428 }
1430 // Recursively
1431 if (_child) _child->counted_loop( phase );
1432 if (_next) _next ->counted_loop( phase );
1433 }
1435 #ifndef PRODUCT
1436 //------------------------------dump_head--------------------------------------
1437 // Dump 1 liner for loop header info
1438 void IdealLoopTree::dump_head( ) const {
1439 for (uint i=0; i<_nest; i++)
1440 tty->print(" ");
1441 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1442 if (_irreducible) tty->print(" IRREDUCIBLE");
1443 if (UseLoopPredicate) {
1444 Node* entry = PhaseIdealLoop::find_predicate_insertion_point(_head->in(LoopNode::EntryControl),
1445 Deoptimization::Reason_predicate);
1446 if (entry != NULL) {
1447 tty->print(" predicated");
1448 }
1449 }
1450 if (_head->is_CountedLoop()) {
1451 CountedLoopNode *cl = _head->as_CountedLoop();
1452 tty->print(" counted");
1454 Node* init_n = cl->init_trip();
1455 if (init_n != NULL && init_n->is_Con())
1456 tty->print(" [%d,", cl->init_trip()->get_int());
1457 else
1458 tty->print(" [int,");
1459 Node* limit_n = cl->limit();
1460 if (limit_n != NULL && limit_n->is_Con())
1461 tty->print("%d),", cl->limit()->get_int());
1462 else
1463 tty->print("int),");
1464 int stride_con = cl->stride_con();
1465 if (stride_con > 0) tty->print("+");
1466 tty->print("%d", stride_con);
1468 if (cl->is_pre_loop ()) tty->print(" pre" );
1469 if (cl->is_main_loop()) tty->print(" main");
1470 if (cl->is_post_loop()) tty->print(" post");
1471 }
1472 tty->cr();
1473 }
1475 //------------------------------dump-------------------------------------------
1476 // Dump loops by loop tree
1477 void IdealLoopTree::dump( ) const {
1478 dump_head();
1479 if (_child) _child->dump();
1480 if (_next) _next ->dump();
1481 }
1483 #endif
1485 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1486 if (loop == root) {
1487 if (loop->_child != NULL) {
1488 log->begin_head("loop_tree");
1489 log->end_head();
1490 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1491 log->tail("loop_tree");
1492 assert(loop->_next == NULL, "what?");
1493 }
1494 } else {
1495 Node* head = loop->_head;
1496 log->begin_head("loop");
1497 log->print(" idx='%d' ", head->_idx);
1498 if (loop->_irreducible) log->print("irreducible='1' ");
1499 if (head->is_Loop()) {
1500 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1501 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1502 }
1503 if (head->is_CountedLoop()) {
1504 CountedLoopNode* cl = head->as_CountedLoop();
1505 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1506 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1507 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1508 }
1509 log->end_head();
1510 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1511 log->tail("loop");
1512 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1513 }
1514 }
1516 //---------------------collect_potentially_useful_predicates-----------------------
1517 // Helper function to collect potentially useful predicates to prevent them from
1518 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1519 void PhaseIdealLoop::collect_potentially_useful_predicates(
1520 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1521 if (loop->_child) { // child
1522 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1523 }
1525 // self (only loops that we can apply loop predication may use their predicates)
1526 if (loop->_head->is_Loop() &&
1527 !loop->_irreducible &&
1528 !loop->tail()->is_top()) {
1529 LoopNode* lpn = loop->_head->as_Loop();
1530 Node* entry = lpn->in(LoopNode::EntryControl);
1531 Node* predicate_proj = find_predicate(entry);
1532 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1533 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1534 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1535 }
1536 }
1538 if (loop->_next) { // sibling
1539 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1540 }
1541 }
1543 //------------------------eliminate_useless_predicates-----------------------------
1544 // Eliminate all inserted predicates if they could not be used by loop predication.
1545 void PhaseIdealLoop::eliminate_useless_predicates() {
1546 if (C->predicate_count() == 0)
1547 return; // no predicate left
1549 Unique_Node_List useful_predicates; // to store useful predicates
1550 if (C->has_loops()) {
1551 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1552 }
1554 for (int i = C->predicate_count(); i > 0; i--) {
1555 Node * n = C->predicate_opaque1_node(i-1);
1556 assert(n->Opcode() == Op_Opaque1, "must be");
1557 if (!useful_predicates.member(n)) { // not in the useful list
1558 _igvn.replace_node(n, n->in(1));
1559 }
1560 }
1561 }
1563 //=============================================================================
1564 //----------------------------build_and_optimize-------------------------------
1565 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1566 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1567 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs) {
1568 ResourceMark rm;
1570 int old_progress = C->major_progress();
1571 uint orig_worklist_size = _igvn._worklist.size();
1573 // Reset major-progress flag for the driver's heuristics
1574 C->clear_major_progress();
1576 #ifndef PRODUCT
1577 // Capture for later assert
1578 uint unique = C->unique();
1579 _loop_invokes++;
1580 _loop_work += unique;
1581 #endif
1583 // True if the method has at least 1 irreducible loop
1584 _has_irreducible_loops = false;
1586 _created_loop_node = false;
1588 Arena *a = Thread::current()->resource_area();
1589 VectorSet visited(a);
1590 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1591 _nodes.map(C->unique(), NULL);
1592 memset(_nodes.adr(), 0, wordSize * C->unique());
1594 // Pre-build the top-level outermost loop tree entry
1595 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1596 // Do not need a safepoint at the top level
1597 _ltree_root->_has_sfpt = 1;
1599 // Initialize Dominators.
1600 // Checked in clone_loop_predicate() during beautify_loops().
1601 _idom_size = 0;
1602 _idom = NULL;
1603 _dom_depth = NULL;
1604 _dom_stk = NULL;
1606 // Empty pre-order array
1607 allocate_preorders();
1609 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1610 // IdealLoopTree entries. Data nodes are NOT walked.
1611 build_loop_tree();
1612 // Check for bailout, and return
1613 if (C->failing()) {
1614 return;
1615 }
1617 // No loops after all
1618 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
1620 // There should always be an outer loop containing the Root and Return nodes.
1621 // If not, we have a degenerate empty program. Bail out in this case.
1622 if (!has_node(C->root())) {
1623 if (!_verify_only) {
1624 C->clear_major_progress();
1625 C->record_method_not_compilable("empty program detected during loop optimization");
1626 }
1627 return;
1628 }
1630 // Nothing to do, so get out
1631 if( !C->has_loops() && !do_split_ifs && !_verify_me && !_verify_only ) {
1632 _igvn.optimize(); // Cleanup NeverBranches
1633 return;
1634 }
1636 // Set loop nesting depth
1637 _ltree_root->set_nest( 0 );
1639 // Split shared headers and insert loop landing pads.
1640 // Do not bother doing this on the Root loop of course.
1641 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
1642 C->print_method("Before beautify loops", 3);
1643 if( _ltree_root->_child->beautify_loops( this ) ) {
1644 // Re-build loop tree!
1645 _ltree_root->_child = NULL;
1646 _nodes.clear();
1647 reallocate_preorders();
1648 build_loop_tree();
1649 // Check for bailout, and return
1650 if (C->failing()) {
1651 return;
1652 }
1653 // Reset loop nesting depth
1654 _ltree_root->set_nest( 0 );
1656 C->print_method("After beautify loops", 3);
1657 }
1658 }
1660 // Build Dominators for elision of NULL checks & loop finding.
1661 // Since nodes do not have a slot for immediate dominator, make
1662 // a persistent side array for that info indexed on node->_idx.
1663 _idom_size = C->unique();
1664 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1665 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1666 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1667 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1669 Dominators();
1671 if (!_verify_only) {
1672 // As a side effect, Dominators removed any unreachable CFG paths
1673 // into RegionNodes. It doesn't do this test against Root, so
1674 // we do it here.
1675 for( uint i = 1; i < C->root()->req(); i++ ) {
1676 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1677 _igvn.hash_delete(C->root());
1678 C->root()->del_req(i);
1679 _igvn._worklist.push(C->root());
1680 i--; // Rerun same iteration on compressed edges
1681 }
1682 }
1684 // Given dominators, try to find inner loops with calls that must
1685 // always be executed (call dominates loop tail). These loops do
1686 // not need a separate safepoint.
1687 Node_List cisstack(a);
1688 _ltree_root->check_safepts(visited, cisstack);
1689 }
1691 // Walk the DATA nodes and place into loops. Find earliest control
1692 // node. For CFG nodes, the _nodes array starts out and remains
1693 // holding the associated IdealLoopTree pointer. For DATA nodes, the
1694 // _nodes array holds the earliest legal controlling CFG node.
1696 // Allocate stack with enough space to avoid frequent realloc
1697 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
1698 Node_Stack nstack( a, stack_size );
1700 visited.Clear();
1701 Node_List worklist(a);
1702 // Don't need C->root() on worklist since
1703 // it will be processed among C->top() inputs
1704 worklist.push( C->top() );
1705 visited.set( C->top()->_idx ); // Set C->top() as visited now
1706 build_loop_early( visited, worklist, nstack );
1708 // Given early legal placement, try finding counted loops. This placement
1709 // is good enough to discover most loop invariants.
1710 if( !_verify_me && !_verify_only )
1711 _ltree_root->counted_loop( this );
1713 // Find latest loop placement. Find ideal loop placement.
1714 visited.Clear();
1715 init_dom_lca_tags();
1716 // Need C->root() on worklist when processing outs
1717 worklist.push( C->root() );
1718 NOT_PRODUCT( C->verify_graph_edges(); )
1719 worklist.push( C->top() );
1720 build_loop_late( visited, worklist, nstack );
1722 if (_verify_only) {
1723 // restore major progress flag
1724 for (int i = 0; i < old_progress; i++)
1725 C->set_major_progress();
1726 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
1727 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
1728 return;
1729 }
1731 // Some parser-inserted loop predicates could never be used by loop
1732 // predication or they were moved away from loop during some optimizations.
1733 // For example, peeling. Eliminate them before next loop optimizations.
1734 if (UseLoopPredicate) {
1735 eliminate_useless_predicates();
1736 }
1738 // clear out the dead code
1739 while(_deadlist.size()) {
1740 _igvn.remove_globally_dead_node(_deadlist.pop());
1741 }
1743 #ifndef PRODUCT
1744 C->verify_graph_edges();
1745 if (_verify_me) { // Nested verify pass?
1746 // Check to see if the verify mode is broken
1747 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
1748 return;
1749 }
1750 if(VerifyLoopOptimizations) verify();
1751 if(TraceLoopOpts && C->has_loops()) {
1752 _ltree_root->dump();
1753 }
1754 #endif
1756 if (ReassociateInvariants) {
1757 // Reassociate invariants and prep for split_thru_phi
1758 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1759 IdealLoopTree* lpt = iter.current();
1760 if (!lpt->is_counted() || !lpt->is_inner()) continue;
1762 lpt->reassociate_invariants(this);
1764 // Because RCE opportunities can be masked by split_thru_phi,
1765 // look for RCE candidates and inhibit split_thru_phi
1766 // on just their loop-phi's for this pass of loop opts
1767 if (SplitIfBlocks && do_split_ifs) {
1768 if (lpt->policy_range_check(this)) {
1769 lpt->_rce_candidate = 1; // = true
1770 }
1771 }
1772 }
1773 }
1775 // Check for aggressive application of split-if and other transforms
1776 // that require basic-block info (like cloning through Phi's)
1777 if( SplitIfBlocks && do_split_ifs ) {
1778 visited.Clear();
1779 split_if_with_blocks( visited, nstack );
1780 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
1781 }
1783 // Perform loop predication before iteration splitting
1784 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
1785 _ltree_root->_child->loop_predication(this);
1786 }
1788 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
1789 if (do_intrinsify_fill()) {
1790 C->set_major_progress();
1791 }
1792 }
1794 // Perform iteration-splitting on inner loops. Split iterations to avoid
1795 // range checks or one-shot null checks.
1797 // If split-if's didn't hack the graph too bad (no CFG changes)
1798 // then do loop opts.
1799 if (C->has_loops() && !C->major_progress()) {
1800 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
1801 _ltree_root->_child->iteration_split( this, worklist );
1802 // No verify after peeling! GCM has hoisted code out of the loop.
1803 // After peeling, the hoisted code could sink inside the peeled area.
1804 // The peeling code does not try to recompute the best location for
1805 // all the code before the peeled area, so the verify pass will always
1806 // complain about it.
1807 }
1808 // Do verify graph edges in any case
1809 NOT_PRODUCT( C->verify_graph_edges(); );
1811 if (!do_split_ifs) {
1812 // We saw major progress in Split-If to get here. We forced a
1813 // pass with unrolling and not split-if, however more split-if's
1814 // might make progress. If the unrolling didn't make progress
1815 // then the major-progress flag got cleared and we won't try
1816 // another round of Split-If. In particular the ever-common
1817 // instance-of/check-cast pattern requires at least 2 rounds of
1818 // Split-If to clear out.
1819 C->set_major_progress();
1820 }
1822 // Repeat loop optimizations if new loops were seen
1823 if (created_loop_node()) {
1824 C->set_major_progress();
1825 }
1827 // Keep loop predicates and perform optimizations with them
1828 // until no more loop optimizations could be done.
1829 // After that switch predicates off and do more loop optimizations.
1830 if (!C->major_progress() && (C->predicate_count() > 0)) {
1831 C->cleanup_loop_predicates(_igvn);
1832 #ifndef PRODUCT
1833 if (TraceLoopOpts) {
1834 tty->print_cr("PredicatesOff");
1835 }
1836 #endif
1837 C->set_major_progress();
1838 }
1840 // Convert scalar to superword operations at the end of all loop opts.
1841 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
1842 // SuperWord transform
1843 SuperWord sw(this);
1844 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1845 IdealLoopTree* lpt = iter.current();
1846 if (lpt->is_counted()) {
1847 sw.transform_loop(lpt);
1848 }
1849 }
1850 }
1852 // Cleanup any modified bits
1853 _igvn.optimize();
1855 // disable assert until issue with split_flow_path is resolved (6742111)
1856 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
1857 // "shouldn't introduce irreducible loops");
1859 if (C->log() != NULL) {
1860 log_loop_tree(_ltree_root, _ltree_root, C->log());
1861 }
1862 }
1864 #ifndef PRODUCT
1865 //------------------------------print_statistics-------------------------------
1866 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
1867 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
1868 void PhaseIdealLoop::print_statistics() {
1869 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
1870 }
1872 //------------------------------verify-----------------------------------------
1873 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
1874 static int fail; // debug only, so its multi-thread dont care
1875 void PhaseIdealLoop::verify() const {
1876 int old_progress = C->major_progress();
1877 ResourceMark rm;
1878 PhaseIdealLoop loop_verify( _igvn, this );
1879 VectorSet visited(Thread::current()->resource_area());
1881 fail = 0;
1882 verify_compare( C->root(), &loop_verify, visited );
1883 assert( fail == 0, "verify loops failed" );
1884 // Verify loop structure is the same
1885 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
1886 // Reset major-progress. It was cleared by creating a verify version of
1887 // PhaseIdealLoop.
1888 for( int i=0; i<old_progress; i++ )
1889 C->set_major_progress();
1890 }
1892 //------------------------------verify_compare---------------------------------
1893 // Make sure me and the given PhaseIdealLoop agree on key data structures
1894 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
1895 if( !n ) return;
1896 if( visited.test_set( n->_idx ) ) return;
1897 if( !_nodes[n->_idx] ) { // Unreachable
1898 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
1899 return;
1900 }
1902 uint i;
1903 for( i = 0; i < n->req(); i++ )
1904 verify_compare( n->in(i), loop_verify, visited );
1906 // Check the '_nodes' block/loop structure
1907 i = n->_idx;
1908 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
1909 if( _nodes[i] != loop_verify->_nodes[i] &&
1910 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
1911 tty->print("Mismatched control setting for: ");
1912 n->dump();
1913 if( fail++ > 10 ) return;
1914 Node *c = get_ctrl_no_update(n);
1915 tty->print("We have it as: ");
1916 if( c->in(0) ) c->dump();
1917 else tty->print_cr("N%d",c->_idx);
1918 tty->print("Verify thinks: ");
1919 if( loop_verify->has_ctrl(n) )
1920 loop_verify->get_ctrl_no_update(n)->dump();
1921 else
1922 loop_verify->get_loop_idx(n)->dump();
1923 tty->cr();
1924 }
1925 } else { // We have a loop
1926 IdealLoopTree *us = get_loop_idx(n);
1927 if( loop_verify->has_ctrl(n) ) {
1928 tty->print("Mismatched loop setting for: ");
1929 n->dump();
1930 if( fail++ > 10 ) return;
1931 tty->print("We have it as: ");
1932 us->dump();
1933 tty->print("Verify thinks: ");
1934 loop_verify->get_ctrl_no_update(n)->dump();
1935 tty->cr();
1936 } else if (!C->major_progress()) {
1937 // Loop selection can be messed up if we did a major progress
1938 // operation, like split-if. Do not verify in that case.
1939 IdealLoopTree *them = loop_verify->get_loop_idx(n);
1940 if( us->_head != them->_head || us->_tail != them->_tail ) {
1941 tty->print("Unequals loops for: ");
1942 n->dump();
1943 if( fail++ > 10 ) return;
1944 tty->print("We have it as: ");
1945 us->dump();
1946 tty->print("Verify thinks: ");
1947 them->dump();
1948 tty->cr();
1949 }
1950 }
1951 }
1953 // Check for immediate dominators being equal
1954 if( i >= _idom_size ) {
1955 if( !n->is_CFG() ) return;
1956 tty->print("CFG Node with no idom: ");
1957 n->dump();
1958 return;
1959 }
1960 if( !n->is_CFG() ) return;
1961 if( n == C->root() ) return; // No IDOM here
1963 assert(n->_idx == i, "sanity");
1964 Node *id = idom_no_update(n);
1965 if( id != loop_verify->idom_no_update(n) ) {
1966 tty->print("Unequals idoms for: ");
1967 n->dump();
1968 if( fail++ > 10 ) return;
1969 tty->print("We have it as: ");
1970 id->dump();
1971 tty->print("Verify thinks: ");
1972 loop_verify->idom_no_update(n)->dump();
1973 tty->cr();
1974 }
1976 }
1978 //------------------------------verify_tree------------------------------------
1979 // Verify that tree structures match. Because the CFG can change, siblings
1980 // within the loop tree can be reordered. We attempt to deal with that by
1981 // reordering the verify's loop tree if possible.
1982 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
1983 assert( _parent == parent, "Badly formed loop tree" );
1985 // Siblings not in same order? Attempt to re-order.
1986 if( _head != loop->_head ) {
1987 // Find _next pointer to update
1988 IdealLoopTree **pp = &loop->_parent->_child;
1989 while( *pp != loop )
1990 pp = &((*pp)->_next);
1991 // Find proper sibling to be next
1992 IdealLoopTree **nn = &loop->_next;
1993 while( (*nn) && (*nn)->_head != _head )
1994 nn = &((*nn)->_next);
1996 // Check for no match.
1997 if( !(*nn) ) {
1998 // Annoyingly, irreducible loops can pick different headers
1999 // after a major_progress operation, so the rest of the loop
2000 // tree cannot be matched.
2001 if (_irreducible && Compile::current()->major_progress()) return;
2002 assert( 0, "failed to match loop tree" );
2003 }
2005 // Move (*nn) to (*pp)
2006 IdealLoopTree *hit = *nn;
2007 *nn = hit->_next;
2008 hit->_next = loop;
2009 *pp = loop;
2010 loop = hit;
2011 // Now try again to verify
2012 }
2014 assert( _head == loop->_head , "mismatched loop head" );
2015 Node *tail = _tail; // Inline a non-updating version of
2016 while( !tail->in(0) ) // the 'tail()' call.
2017 tail = tail->in(1);
2018 assert( tail == loop->_tail, "mismatched loop tail" );
2020 // Counted loops that are guarded should be able to find their guards
2021 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2022 CountedLoopNode *cl = _head->as_CountedLoop();
2023 Node *init = cl->init_trip();
2024 Node *ctrl = cl->in(LoopNode::EntryControl);
2025 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2026 Node *iff = ctrl->in(0);
2027 assert( iff->Opcode() == Op_If, "" );
2028 Node *bol = iff->in(1);
2029 assert( bol->Opcode() == Op_Bool, "" );
2030 Node *cmp = bol->in(1);
2031 assert( cmp->Opcode() == Op_CmpI, "" );
2032 Node *add = cmp->in(1);
2033 Node *opaq;
2034 if( add->Opcode() == Op_Opaque1 ) {
2035 opaq = add;
2036 } else {
2037 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2038 assert( add == init, "" );
2039 opaq = cmp->in(2);
2040 }
2041 assert( opaq->Opcode() == Op_Opaque1, "" );
2043 }
2045 if (_child != NULL) _child->verify_tree(loop->_child, this);
2046 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2047 // Innermost loops need to verify loop bodies,
2048 // but only if no 'major_progress'
2049 int fail = 0;
2050 if (!Compile::current()->major_progress() && _child == NULL) {
2051 for( uint i = 0; i < _body.size(); i++ ) {
2052 Node *n = _body.at(i);
2053 if (n->outcnt() == 0) continue; // Ignore dead
2054 uint j;
2055 for( j = 0; j < loop->_body.size(); j++ )
2056 if( loop->_body.at(j) == n )
2057 break;
2058 if( j == loop->_body.size() ) { // Not found in loop body
2059 // Last ditch effort to avoid assertion: Its possible that we
2060 // have some users (so outcnt not zero) but are still dead.
2061 // Try to find from root.
2062 if (Compile::current()->root()->find(n->_idx)) {
2063 fail++;
2064 tty->print("We have that verify does not: ");
2065 n->dump();
2066 }
2067 }
2068 }
2069 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2070 Node *n = loop->_body.at(i2);
2071 if (n->outcnt() == 0) continue; // Ignore dead
2072 uint j;
2073 for( j = 0; j < _body.size(); j++ )
2074 if( _body.at(j) == n )
2075 break;
2076 if( j == _body.size() ) { // Not found in loop body
2077 // Last ditch effort to avoid assertion: Its possible that we
2078 // have some users (so outcnt not zero) but are still dead.
2079 // Try to find from root.
2080 if (Compile::current()->root()->find(n->_idx)) {
2081 fail++;
2082 tty->print("Verify has that we do not: ");
2083 n->dump();
2084 }
2085 }
2086 }
2087 assert( !fail, "loop body mismatch" );
2088 }
2089 }
2091 #endif
2093 //------------------------------set_idom---------------------------------------
2094 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2095 uint idx = d->_idx;
2096 if (idx >= _idom_size) {
2097 uint newsize = _idom_size<<1;
2098 while( idx >= newsize ) {
2099 newsize <<= 1;
2100 }
2101 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2102 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2103 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2104 _idom_size = newsize;
2105 }
2106 _idom[idx] = n;
2107 _dom_depth[idx] = dom_depth;
2108 }
2110 //------------------------------recompute_dom_depth---------------------------------------
2111 // The dominator tree is constructed with only parent pointers.
2112 // This recomputes the depth in the tree by first tagging all
2113 // nodes as "no depth yet" marker. The next pass then runs up
2114 // the dom tree from each node marked "no depth yet", and computes
2115 // the depth on the way back down.
2116 void PhaseIdealLoop::recompute_dom_depth() {
2117 uint no_depth_marker = C->unique();
2118 uint i;
2119 // Initialize depth to "no depth yet"
2120 for (i = 0; i < _idom_size; i++) {
2121 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2122 _dom_depth[i] = no_depth_marker;
2123 }
2124 }
2125 if (_dom_stk == NULL) {
2126 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
2127 if (init_size < 10) init_size = 10;
2128 _dom_stk = new GrowableArray<uint>(init_size);
2129 }
2130 // Compute new depth for each node.
2131 for (i = 0; i < _idom_size; i++) {
2132 uint j = i;
2133 // Run up the dom tree to find a node with a depth
2134 while (_dom_depth[j] == no_depth_marker) {
2135 _dom_stk->push(j);
2136 j = _idom[j]->_idx;
2137 }
2138 // Compute the depth on the way back down this tree branch
2139 uint dd = _dom_depth[j] + 1;
2140 while (_dom_stk->length() > 0) {
2141 uint j = _dom_stk->pop();
2142 _dom_depth[j] = dd;
2143 dd++;
2144 }
2145 }
2146 }
2148 //------------------------------sort-------------------------------------------
2149 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2150 // loop tree, not the root.
2151 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2152 if( !innermost ) return loop; // New innermost loop
2154 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2155 assert( loop_preorder, "not yet post-walked loop" );
2156 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2157 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2159 // Insert at start of list
2160 while( l ) { // Insertion sort based on pre-order
2161 if( l == loop ) return innermost; // Already on list!
2162 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2163 assert( l_preorder, "not yet post-walked l" );
2164 // Check header pre-order number to figure proper nesting
2165 if( loop_preorder > l_preorder )
2166 break; // End of insertion
2167 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2168 // Since I split shared headers, you'd think this could not happen.
2169 // BUT: I must first do the preorder numbering before I can discover I
2170 // have shared headers, so the split headers all get the same preorder
2171 // number as the RegionNode they split from.
2172 if( loop_preorder == l_preorder &&
2173 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2174 break; // Also check for shared headers (same pre#)
2175 pp = &l->_parent; // Chain up list
2176 l = *pp;
2177 }
2178 // Link into list
2179 // Point predecessor to me
2180 *pp = loop;
2181 // Point me to successor
2182 IdealLoopTree *p = loop->_parent;
2183 loop->_parent = l; // Point me to successor
2184 if( p ) sort( p, innermost ); // Insert my parents into list as well
2185 return innermost;
2186 }
2188 //------------------------------build_loop_tree--------------------------------
2189 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2190 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2191 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2192 // tightest enclosing IdealLoopTree for post-walked.
2193 //
2194 // During my forward walk I do a short 1-layer lookahead to see if I can find
2195 // a loop backedge with that doesn't have any work on the backedge. This
2196 // helps me construct nested loops with shared headers better.
2197 //
2198 // Once I've done the forward recursion, I do the post-work. For each child
2199 // I check to see if there is a backedge. Backedges define a loop! I
2200 // insert an IdealLoopTree at the target of the backedge.
2201 //
2202 // During the post-work I also check to see if I have several children
2203 // belonging to different loops. If so, then this Node is a decision point
2204 // where control flow can choose to change loop nests. It is at this
2205 // decision point where I can figure out how loops are nested. At this
2206 // time I can properly order the different loop nests from my children.
2207 // Note that there may not be any backedges at the decision point!
2208 //
2209 // Since the decision point can be far removed from the backedges, I can't
2210 // order my loops at the time I discover them. Thus at the decision point
2211 // I need to inspect loop header pre-order numbers to properly nest my
2212 // loops. This means I need to sort my childrens' loops by pre-order.
2213 // The sort is of size number-of-control-children, which generally limits
2214 // it to size 2 (i.e., I just choose between my 2 target loops).
2215 void PhaseIdealLoop::build_loop_tree() {
2216 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2217 GrowableArray <Node *> bltstack(C->unique() >> 1);
2218 Node *n = C->root();
2219 bltstack.push(n);
2220 int pre_order = 1;
2221 int stack_size;
2223 while ( ( stack_size = bltstack.length() ) != 0 ) {
2224 n = bltstack.top(); // Leave node on stack
2225 if ( !is_visited(n) ) {
2226 // ---- Pre-pass Work ----
2227 // Pre-walked but not post-walked nodes need a pre_order number.
2229 set_preorder_visited( n, pre_order ); // set as visited
2231 // ---- Scan over children ----
2232 // Scan first over control projections that lead to loop headers.
2233 // This helps us find inner-to-outer loops with shared headers better.
2235 // Scan children's children for loop headers.
2236 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2237 Node* m = n->raw_out(i); // Child
2238 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2239 // Scan over children's children to find loop
2240 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2241 Node* l = m->fast_out(j);
2242 if( is_visited(l) && // Been visited?
2243 !is_postvisited(l) && // But not post-visited
2244 get_preorder(l) < pre_order ) { // And smaller pre-order
2245 // Found! Scan the DFS down this path before doing other paths
2246 bltstack.push(m);
2247 break;
2248 }
2249 }
2250 }
2251 }
2252 pre_order++;
2253 }
2254 else if ( !is_postvisited(n) ) {
2255 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2256 // such as com.sun.rsasign.am::a.
2257 // For non-recursive version, first, process current children.
2258 // On next iteration, check if additional children were added.
2259 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2260 Node* u = n->raw_out(k);
2261 if ( u->is_CFG() && !is_visited(u) ) {
2262 bltstack.push(u);
2263 }
2264 }
2265 if ( bltstack.length() == stack_size ) {
2266 // There were no additional children, post visit node now
2267 (void)bltstack.pop(); // Remove node from stack
2268 pre_order = build_loop_tree_impl( n, pre_order );
2269 // Check for bailout
2270 if (C->failing()) {
2271 return;
2272 }
2273 // Check to grow _preorders[] array for the case when
2274 // build_loop_tree_impl() adds new nodes.
2275 check_grow_preorders();
2276 }
2277 }
2278 else {
2279 (void)bltstack.pop(); // Remove post-visited node from stack
2280 }
2281 }
2282 }
2284 //------------------------------build_loop_tree_impl---------------------------
2285 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2286 // ---- Post-pass Work ----
2287 // Pre-walked but not post-walked nodes need a pre_order number.
2289 // Tightest enclosing loop for this Node
2290 IdealLoopTree *innermost = NULL;
2292 // For all children, see if any edge is a backedge. If so, make a loop
2293 // for it. Then find the tightest enclosing loop for the self Node.
2294 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2295 Node* m = n->fast_out(i); // Child
2296 if( n == m ) continue; // Ignore control self-cycles
2297 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2299 IdealLoopTree *l; // Child's loop
2300 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2301 // Found a backedge
2302 assert( get_preorder(m) < pre_order, "should be backedge" );
2303 // Check for the RootNode, which is already a LoopNode and is allowed
2304 // to have multiple "backedges".
2305 if( m == C->root()) { // Found the root?
2306 l = _ltree_root; // Root is the outermost LoopNode
2307 } else { // Else found a nested loop
2308 // Insert a LoopNode to mark this loop.
2309 l = new IdealLoopTree(this, m, n);
2310 } // End of Else found a nested loop
2311 if( !has_loop(m) ) // If 'm' does not already have a loop set
2312 set_loop(m, l); // Set loop header to loop now
2314 } else { // Else not a nested loop
2315 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2316 l = get_loop(m); // Get previously determined loop
2317 // If successor is header of a loop (nest), move up-loop till it
2318 // is a member of some outer enclosing loop. Since there are no
2319 // shared headers (I've split them already) I only need to go up
2320 // at most 1 level.
2321 while( l && l->_head == m ) // Successor heads loop?
2322 l = l->_parent; // Move up 1 for me
2323 // If this loop is not properly parented, then this loop
2324 // has no exit path out, i.e. its an infinite loop.
2325 if( !l ) {
2326 // Make loop "reachable" from root so the CFG is reachable. Basically
2327 // insert a bogus loop exit that is never taken. 'm', the loop head,
2328 // points to 'n', one (of possibly many) fall-in paths. There may be
2329 // many backedges as well.
2331 // Here I set the loop to be the root loop. I could have, after
2332 // inserting a bogus loop exit, restarted the recursion and found my
2333 // new loop exit. This would make the infinite loop a first-class
2334 // loop and it would then get properly optimized. What's the use of
2335 // optimizing an infinite loop?
2336 l = _ltree_root; // Oops, found infinite loop
2338 if (!_verify_only) {
2339 // Insert the NeverBranch between 'm' and it's control user.
2340 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2341 _igvn.register_new_node_with_optimizer(iff);
2342 set_loop(iff, l);
2343 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2344 _igvn.register_new_node_with_optimizer(if_t);
2345 set_loop(if_t, l);
2347 Node* cfg = NULL; // Find the One True Control User of m
2348 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2349 Node* x = m->fast_out(j);
2350 if (x->is_CFG() && x != m && x != iff)
2351 { cfg = x; break; }
2352 }
2353 assert(cfg != NULL, "must find the control user of m");
2354 uint k = 0; // Probably cfg->in(0)
2355 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2356 cfg->set_req( k, if_t ); // Now point to NeverBranch
2358 // Now create the never-taken loop exit
2359 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2360 _igvn.register_new_node_with_optimizer(if_f);
2361 set_loop(if_f, l);
2362 // Find frame ptr for Halt. Relies on the optimizer
2363 // V-N'ing. Easier and quicker than searching through
2364 // the program structure.
2365 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2366 _igvn.register_new_node_with_optimizer(frame);
2367 // Halt & Catch Fire
2368 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2369 _igvn.register_new_node_with_optimizer(halt);
2370 set_loop(halt, l);
2371 C->root()->add_req(halt);
2372 }
2373 set_loop(C->root(), _ltree_root);
2374 }
2375 }
2376 // Weeny check for irreducible. This child was already visited (this
2377 // IS the post-work phase). Is this child's loop header post-visited
2378 // as well? If so, then I found another entry into the loop.
2379 if (!_verify_only) {
2380 while( is_postvisited(l->_head) ) {
2381 // found irreducible
2382 l->_irreducible = 1; // = true
2383 l = l->_parent;
2384 _has_irreducible_loops = true;
2385 // Check for bad CFG here to prevent crash, and bailout of compile
2386 if (l == NULL) {
2387 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2388 return pre_order;
2389 }
2390 }
2391 }
2393 // This Node might be a decision point for loops. It is only if
2394 // it's children belong to several different loops. The sort call
2395 // does a trivial amount of work if there is only 1 child or all
2396 // children belong to the same loop. If however, the children
2397 // belong to different loops, the sort call will properly set the
2398 // _parent pointers to show how the loops nest.
2399 //
2400 // In any case, it returns the tightest enclosing loop.
2401 innermost = sort( l, innermost );
2402 }
2404 // Def-use info will have some dead stuff; dead stuff will have no
2405 // loop decided on.
2407 // Am I a loop header? If so fix up my parent's child and next ptrs.
2408 if( innermost && innermost->_head == n ) {
2409 assert( get_loop(n) == innermost, "" );
2410 IdealLoopTree *p = innermost->_parent;
2411 IdealLoopTree *l = innermost;
2412 while( p && l->_head == n ) {
2413 l->_next = p->_child; // Put self on parents 'next child'
2414 p->_child = l; // Make self as first child of parent
2415 l = p; // Now walk up the parent chain
2416 p = l->_parent;
2417 }
2418 } else {
2419 // Note that it is possible for a LoopNode to reach here, if the
2420 // backedge has been made unreachable (hence the LoopNode no longer
2421 // denotes a Loop, and will eventually be removed).
2423 // Record tightest enclosing loop for self. Mark as post-visited.
2424 set_loop(n, innermost);
2425 // Also record has_call flag early on
2426 if( innermost ) {
2427 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2428 // Do not count uncommon calls
2429 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2430 Node *iff = n->in(0)->in(0);
2431 if( !iff->is_If() ||
2432 (n->in(0)->Opcode() == Op_IfFalse &&
2433 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2434 (iff->as_If()->_prob >= 0.01) )
2435 innermost->_has_call = 1;
2436 }
2437 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2438 // Disable loop optimizations if the loop has a scalar replaceable
2439 // allocation. This disabling may cause a potential performance lost
2440 // if the allocation is not eliminated for some reason.
2441 innermost->_allow_optimizations = false;
2442 innermost->_has_call = 1; // = true
2443 }
2444 }
2445 }
2447 // Flag as post-visited now
2448 set_postvisited(n);
2449 return pre_order;
2450 }
2453 //------------------------------build_loop_early-------------------------------
2454 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2455 // First pass computes the earliest controlling node possible. This is the
2456 // controlling input with the deepest dominating depth.
2457 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2458 while (worklist.size() != 0) {
2459 // Use local variables nstack_top_n & nstack_top_i to cache values
2460 // on nstack's top.
2461 Node *nstack_top_n = worklist.pop();
2462 uint nstack_top_i = 0;
2463 //while_nstack_nonempty:
2464 while (true) {
2465 // Get parent node and next input's index from stack's top.
2466 Node *n = nstack_top_n;
2467 uint i = nstack_top_i;
2468 uint cnt = n->req(); // Count of inputs
2469 if (i == 0) { // Pre-process the node.
2470 if( has_node(n) && // Have either loop or control already?
2471 !has_ctrl(n) ) { // Have loop picked out already?
2472 // During "merge_many_backedges" we fold up several nested loops
2473 // into a single loop. This makes the members of the original
2474 // loop bodies pointing to dead loops; they need to move up
2475 // to the new UNION'd larger loop. I set the _head field of these
2476 // dead loops to NULL and the _parent field points to the owning
2477 // loop. Shades of UNION-FIND algorithm.
2478 IdealLoopTree *ilt;
2479 while( !(ilt = get_loop(n))->_head ) {
2480 // Normally I would use a set_loop here. But in this one special
2481 // case, it is legal (and expected) to change what loop a Node
2482 // belongs to.
2483 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2484 }
2485 // Remove safepoints ONLY if I've already seen I don't need one.
2486 // (the old code here would yank a 2nd safepoint after seeing a
2487 // first one, even though the 1st did not dominate in the loop body
2488 // and thus could be avoided indefinitely)
2489 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2490 is_deleteable_safept(n)) {
2491 Node *in = n->in(TypeFunc::Control);
2492 lazy_replace(n,in); // Pull safepoint now
2493 // Carry on with the recursion "as if" we are walking
2494 // only the control input
2495 if( !visited.test_set( in->_idx ) ) {
2496 worklist.push(in); // Visit this guy later, using worklist
2497 }
2498 // Get next node from nstack:
2499 // - skip n's inputs processing by setting i > cnt;
2500 // - we also will not call set_early_ctrl(n) since
2501 // has_node(n) == true (see the condition above).
2502 i = cnt + 1;
2503 }
2504 }
2505 } // if (i == 0)
2507 // Visit all inputs
2508 bool done = true; // Assume all n's inputs will be processed
2509 while (i < cnt) {
2510 Node *in = n->in(i);
2511 ++i;
2512 if (in == NULL) continue;
2513 if (in->pinned() && !in->is_CFG())
2514 set_ctrl(in, in->in(0));
2515 int is_visited = visited.test_set( in->_idx );
2516 if (!has_node(in)) { // No controlling input yet?
2517 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2518 assert( !is_visited, "visit only once" );
2519 nstack.push(n, i); // Save parent node and next input's index.
2520 nstack_top_n = in; // Process current input now.
2521 nstack_top_i = 0;
2522 done = false; // Not all n's inputs processed.
2523 break; // continue while_nstack_nonempty;
2524 } else if (!is_visited) {
2525 // This guy has a location picked out for him, but has not yet
2526 // been visited. Happens to all CFG nodes, for instance.
2527 // Visit him using the worklist instead of recursion, to break
2528 // cycles. Since he has a location already we do not need to
2529 // find his location before proceeding with the current Node.
2530 worklist.push(in); // Visit this guy later, using worklist
2531 }
2532 }
2533 if (done) {
2534 // All of n's inputs have been processed, complete post-processing.
2536 // Compute earliest point this Node can go.
2537 // CFG, Phi, pinned nodes already know their controlling input.
2538 if (!has_node(n)) {
2539 // Record earliest legal location
2540 set_early_ctrl( n );
2541 }
2542 if (nstack.is_empty()) {
2543 // Finished all nodes on stack.
2544 // Process next node on the worklist.
2545 break;
2546 }
2547 // Get saved parent node and next input's index.
2548 nstack_top_n = nstack.node();
2549 nstack_top_i = nstack.index();
2550 nstack.pop();
2551 }
2552 } // while (true)
2553 }
2554 }
2556 //------------------------------dom_lca_internal--------------------------------
2557 // Pair-wise LCA
2558 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2559 if( !n1 ) return n2; // Handle NULL original LCA
2560 assert( n1->is_CFG(), "" );
2561 assert( n2->is_CFG(), "" );
2562 // find LCA of all uses
2563 uint d1 = dom_depth(n1);
2564 uint d2 = dom_depth(n2);
2565 while (n1 != n2) {
2566 if (d1 > d2) {
2567 n1 = idom(n1);
2568 d1 = dom_depth(n1);
2569 } else if (d1 < d2) {
2570 n2 = idom(n2);
2571 d2 = dom_depth(n2);
2572 } else {
2573 // Here d1 == d2. Due to edits of the dominator-tree, sections
2574 // of the tree might have the same depth. These sections have
2575 // to be searched more carefully.
2577 // Scan up all the n1's with equal depth, looking for n2.
2578 Node *t1 = idom(n1);
2579 while (dom_depth(t1) == d1) {
2580 if (t1 == n2) return n2;
2581 t1 = idom(t1);
2582 }
2583 // Scan up all the n2's with equal depth, looking for n1.
2584 Node *t2 = idom(n2);
2585 while (dom_depth(t2) == d2) {
2586 if (t2 == n1) return n1;
2587 t2 = idom(t2);
2588 }
2589 // Move up to a new dominator-depth value as well as up the dom-tree.
2590 n1 = t1;
2591 n2 = t2;
2592 d1 = dom_depth(n1);
2593 d2 = dom_depth(n2);
2594 }
2595 }
2596 return n1;
2597 }
2599 //------------------------------compute_idom-----------------------------------
2600 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2601 // IDOMs are correct.
2602 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2603 assert( region->is_Region(), "" );
2604 Node *LCA = NULL;
2605 for( uint i = 1; i < region->req(); i++ ) {
2606 if( region->in(i) != C->top() )
2607 LCA = dom_lca( LCA, region->in(i) );
2608 }
2609 return LCA;
2610 }
2612 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
2613 bool had_error = false;
2614 #ifdef ASSERT
2615 if (early != C->root()) {
2616 // Make sure that there's a dominance path from use to LCA
2617 Node* d = use;
2618 while (d != LCA) {
2619 d = idom(d);
2620 if (d == C->root()) {
2621 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
2622 n->dump();
2623 use->dump();
2624 had_error = true;
2625 break;
2626 }
2627 }
2628 }
2629 #endif
2630 return had_error;
2631 }
2634 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
2635 // Compute LCA over list of uses
2636 bool had_error = false;
2637 Node *LCA = NULL;
2638 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2639 Node* c = n->fast_out(i);
2640 if (_nodes[c->_idx] == NULL)
2641 continue; // Skip the occasional dead node
2642 if( c->is_Phi() ) { // For Phis, we must land above on the path
2643 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2644 if( c->in(j) == n ) { // Found matching input?
2645 Node *use = c->in(0)->in(j);
2646 if (_verify_only && use->is_top()) continue;
2647 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2648 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2649 }
2650 }
2651 } else {
2652 // For CFG data-users, use is in the block just prior
2653 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2654 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2655 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2656 }
2657 }
2658 assert(!had_error, "bad dominance");
2659 return LCA;
2660 }
2662 //------------------------------get_late_ctrl----------------------------------
2663 // Compute latest legal control.
2664 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2665 assert(early != NULL, "early control should not be NULL");
2667 Node* LCA = compute_lca_of_uses(n, early);
2668 #ifdef ASSERT
2669 if (LCA == C->root() && LCA != early) {
2670 // def doesn't dominate uses so print some useful debugging output
2671 compute_lca_of_uses(n, early, true);
2672 }
2673 #endif
2675 // if this is a load, check for anti-dependent stores
2676 // We use a conservative algorithm to identify potential interfering
2677 // instructions and for rescheduling the load. The users of the memory
2678 // input of this load are examined. Any use which is not a load and is
2679 // dominated by early is considered a potentially interfering store.
2680 // This can produce false positives.
2681 if (n->is_Load() && LCA != early) {
2682 Node_List worklist;
2684 Node *mem = n->in(MemNode::Memory);
2685 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2686 Node* s = mem->fast_out(i);
2687 worklist.push(s);
2688 }
2689 while(worklist.size() != 0 && LCA != early) {
2690 Node* s = worklist.pop();
2691 if (s->is_Load()) {
2692 continue;
2693 } else if (s->is_MergeMem()) {
2694 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
2695 Node* s1 = s->fast_out(i);
2696 worklist.push(s1);
2697 }
2698 } else {
2699 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
2700 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
2701 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
2702 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
2703 }
2704 }
2705 }
2706 }
2708 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
2709 return LCA;
2710 }
2712 // true if CFG node d dominates CFG node n
2713 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
2714 if (d == n)
2715 return true;
2716 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
2717 uint dd = dom_depth(d);
2718 while (dom_depth(n) >= dd) {
2719 if (n == d)
2720 return true;
2721 n = idom(n);
2722 }
2723 return false;
2724 }
2726 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
2727 // Pair-wise LCA with tags.
2728 // Tag each index with the node 'tag' currently being processed
2729 // before advancing up the dominator chain using idom().
2730 // Later calls that find a match to 'tag' know that this path has already
2731 // been considered in the current LCA (which is input 'n1' by convention).
2732 // Since get_late_ctrl() is only called once for each node, the tag array
2733 // does not need to be cleared between calls to get_late_ctrl().
2734 // Algorithm trades a larger constant factor for better asymptotic behavior
2735 //
2736 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
2737 uint d1 = dom_depth(n1);
2738 uint d2 = dom_depth(n2);
2740 do {
2741 if (d1 > d2) {
2742 // current lca is deeper than n2
2743 _dom_lca_tags.map(n1->_idx, tag);
2744 n1 = idom(n1);
2745 d1 = dom_depth(n1);
2746 } else if (d1 < d2) {
2747 // n2 is deeper than current lca
2748 Node *memo = _dom_lca_tags[n2->_idx];
2749 if( memo == tag ) {
2750 return n1; // Return the current LCA
2751 }
2752 _dom_lca_tags.map(n2->_idx, tag);
2753 n2 = idom(n2);
2754 d2 = dom_depth(n2);
2755 } else {
2756 // Here d1 == d2. Due to edits of the dominator-tree, sections
2757 // of the tree might have the same depth. These sections have
2758 // to be searched more carefully.
2760 // Scan up all the n1's with equal depth, looking for n2.
2761 _dom_lca_tags.map(n1->_idx, tag);
2762 Node *t1 = idom(n1);
2763 while (dom_depth(t1) == d1) {
2764 if (t1 == n2) return n2;
2765 _dom_lca_tags.map(t1->_idx, tag);
2766 t1 = idom(t1);
2767 }
2768 // Scan up all the n2's with equal depth, looking for n1.
2769 _dom_lca_tags.map(n2->_idx, tag);
2770 Node *t2 = idom(n2);
2771 while (dom_depth(t2) == d2) {
2772 if (t2 == n1) return n1;
2773 _dom_lca_tags.map(t2->_idx, tag);
2774 t2 = idom(t2);
2775 }
2776 // Move up to a new dominator-depth value as well as up the dom-tree.
2777 n1 = t1;
2778 n2 = t2;
2779 d1 = dom_depth(n1);
2780 d2 = dom_depth(n2);
2781 }
2782 } while (n1 != n2);
2783 return n1;
2784 }
2786 //------------------------------init_dom_lca_tags------------------------------
2787 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2788 // Intended use does not involve any growth for the array, so it could
2789 // be of fixed size.
2790 void PhaseIdealLoop::init_dom_lca_tags() {
2791 uint limit = C->unique() + 1;
2792 _dom_lca_tags.map( limit, NULL );
2793 #ifdef ASSERT
2794 for( uint i = 0; i < limit; ++i ) {
2795 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2796 }
2797 #endif // ASSERT
2798 }
2800 //------------------------------clear_dom_lca_tags------------------------------
2801 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2802 // Intended use does not involve any growth for the array, so it could
2803 // be of fixed size.
2804 void PhaseIdealLoop::clear_dom_lca_tags() {
2805 uint limit = C->unique() + 1;
2806 _dom_lca_tags.map( limit, NULL );
2807 _dom_lca_tags.clear();
2808 #ifdef ASSERT
2809 for( uint i = 0; i < limit; ++i ) {
2810 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2811 }
2812 #endif // ASSERT
2813 }
2815 //------------------------------build_loop_late--------------------------------
2816 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2817 // Second pass finds latest legal placement, and ideal loop placement.
2818 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2819 while (worklist.size() != 0) {
2820 Node *n = worklist.pop();
2821 // Only visit once
2822 if (visited.test_set(n->_idx)) continue;
2823 uint cnt = n->outcnt();
2824 uint i = 0;
2825 while (true) {
2826 assert( _nodes[n->_idx], "no dead nodes" );
2827 // Visit all children
2828 if (i < cnt) {
2829 Node* use = n->raw_out(i);
2830 ++i;
2831 // Check for dead uses. Aggressively prune such junk. It might be
2832 // dead in the global sense, but still have local uses so I cannot
2833 // easily call 'remove_dead_node'.
2834 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
2835 // Due to cycles, we might not hit the same fixed point in the verify
2836 // pass as we do in the regular pass. Instead, visit such phis as
2837 // simple uses of the loop head.
2838 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
2839 if( !visited.test(use->_idx) )
2840 worklist.push(use);
2841 } else if( !visited.test_set(use->_idx) ) {
2842 nstack.push(n, i); // Save parent and next use's index.
2843 n = use; // Process all children of current use.
2844 cnt = use->outcnt();
2845 i = 0;
2846 }
2847 } else {
2848 // Do not visit around the backedge of loops via data edges.
2849 // push dead code onto a worklist
2850 _deadlist.push(use);
2851 }
2852 } else {
2853 // All of n's children have been processed, complete post-processing.
2854 build_loop_late_post(n);
2855 if (nstack.is_empty()) {
2856 // Finished all nodes on stack.
2857 // Process next node on the worklist.
2858 break;
2859 }
2860 // Get saved parent node and next use's index. Visit the rest of uses.
2861 n = nstack.node();
2862 cnt = n->outcnt();
2863 i = nstack.index();
2864 nstack.pop();
2865 }
2866 }
2867 }
2868 }
2870 //------------------------------build_loop_late_post---------------------------
2871 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2872 // Second pass finds latest legal placement, and ideal loop placement.
2873 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
2875 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
2876 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
2877 }
2879 // CFG and pinned nodes already handled
2880 if( n->in(0) ) {
2881 if( n->in(0)->is_top() ) return; // Dead?
2883 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
2884 // _must_ be pinned (they have to observe their control edge of course).
2885 // Unlike Stores (which modify an unallocable resource, the memory
2886 // state), Mods/Loads can float around. So free them up.
2887 bool pinned = true;
2888 switch( n->Opcode() ) {
2889 case Op_DivI:
2890 case Op_DivF:
2891 case Op_DivD:
2892 case Op_ModI:
2893 case Op_ModF:
2894 case Op_ModD:
2895 case Op_LoadB: // Same with Loads; they can sink
2896 case Op_LoadUS: // during loop optimizations.
2897 case Op_LoadD:
2898 case Op_LoadF:
2899 case Op_LoadI:
2900 case Op_LoadKlass:
2901 case Op_LoadNKlass:
2902 case Op_LoadL:
2903 case Op_LoadS:
2904 case Op_LoadP:
2905 case Op_LoadN:
2906 case Op_LoadRange:
2907 case Op_LoadD_unaligned:
2908 case Op_LoadL_unaligned:
2909 case Op_StrComp: // Does a bunch of load-like effects
2910 case Op_StrEquals:
2911 case Op_StrIndexOf:
2912 case Op_AryEq:
2913 pinned = false;
2914 }
2915 if( pinned ) {
2916 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
2917 if( !chosen_loop->_child ) // Inner loop?
2918 chosen_loop->_body.push(n); // Collect inner loops
2919 return;
2920 }
2921 } else { // No slot zero
2922 if( n->is_CFG() ) { // CFG with no slot 0 is dead
2923 _nodes.map(n->_idx,0); // No block setting, it's globally dead
2924 return;
2925 }
2926 assert(!n->is_CFG() || n->outcnt() == 0, "");
2927 }
2929 // Do I have a "safe range" I can select over?
2930 Node *early = get_ctrl(n);// Early location already computed
2932 // Compute latest point this Node can go
2933 Node *LCA = get_late_ctrl( n, early );
2934 // LCA is NULL due to uses being dead
2935 if( LCA == NULL ) {
2936 #ifdef ASSERT
2937 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
2938 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
2939 }
2940 #endif
2941 _nodes.map(n->_idx, 0); // This node is useless
2942 _deadlist.push(n);
2943 return;
2944 }
2945 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
2947 Node *legal = LCA; // Walk 'legal' up the IDOM chain
2948 Node *least = legal; // Best legal position so far
2949 while( early != legal ) { // While not at earliest legal
2950 #ifdef ASSERT
2951 if (legal->is_Start() && !early->is_Root()) {
2952 // Bad graph. Print idom path and fail.
2953 tty->print_cr( "Bad graph detected in build_loop_late");
2954 tty->print("n: ");n->dump(); tty->cr();
2955 tty->print("early: ");early->dump(); tty->cr();
2956 int ct = 0;
2957 Node *dbg_legal = LCA;
2958 while(!dbg_legal->is_Start() && ct < 100) {
2959 tty->print("idom[%d] ",ct); dbg_legal->dump(); tty->cr();
2960 ct++;
2961 dbg_legal = idom(dbg_legal);
2962 }
2963 assert(false, "Bad graph detected in build_loop_late");
2964 }
2965 #endif
2966 // Find least loop nesting depth
2967 legal = idom(legal); // Bump up the IDOM tree
2968 // Check for lower nesting depth
2969 if( get_loop(legal)->_nest < get_loop(least)->_nest )
2970 least = legal;
2971 }
2972 assert(early == legal || legal != C->root(), "bad dominance of inputs");
2974 // Try not to place code on a loop entry projection
2975 // which can inhibit range check elimination.
2976 if (least != early) {
2977 Node* ctrl_out = least->unique_ctrl_out();
2978 if (ctrl_out && ctrl_out->is_CountedLoop() &&
2979 least == ctrl_out->in(LoopNode::EntryControl)) {
2980 Node* least_dom = idom(least);
2981 if (get_loop(least_dom)->is_member(get_loop(least))) {
2982 least = least_dom;
2983 }
2984 }
2985 }
2987 #ifdef ASSERT
2988 // If verifying, verify that 'verify_me' has a legal location
2989 // and choose it as our location.
2990 if( _verify_me ) {
2991 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
2992 Node *legal = LCA;
2993 while( early != legal ) { // While not at earliest legal
2994 if( legal == v_ctrl ) break; // Check for prior good location
2995 legal = idom(legal) ;// Bump up the IDOM tree
2996 }
2997 // Check for prior good location
2998 if( legal == v_ctrl ) least = legal; // Keep prior if found
2999 }
3000 #endif
3002 // Assign discovered "here or above" point
3003 least = find_non_split_ctrl(least);
3004 set_ctrl(n, least);
3006 // Collect inner loop bodies
3007 IdealLoopTree *chosen_loop = get_loop(least);
3008 if( !chosen_loop->_child ) // Inner loop?
3009 chosen_loop->_body.push(n);// Collect inner loops
3010 }
3012 #ifndef PRODUCT
3013 //------------------------------dump-------------------------------------------
3014 void PhaseIdealLoop::dump( ) const {
3015 ResourceMark rm;
3016 Arena* arena = Thread::current()->resource_area();
3017 Node_Stack stack(arena, C->unique() >> 2);
3018 Node_List rpo_list;
3019 VectorSet visited(arena);
3020 visited.set(C->top()->_idx);
3021 rpo( C->root(), stack, visited, rpo_list );
3022 // Dump root loop indexed by last element in PO order
3023 dump( _ltree_root, rpo_list.size(), rpo_list );
3024 }
3026 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3027 loop->dump_head();
3029 // Now scan for CFG nodes in the same loop
3030 for( uint j=idx; j > 0; j-- ) {
3031 Node *n = rpo_list[j-1];
3032 if( !_nodes[n->_idx] ) // Skip dead nodes
3033 continue;
3034 if( get_loop(n) != loop ) { // Wrong loop nest
3035 if( get_loop(n)->_head == n && // Found nested loop?
3036 get_loop(n)->_parent == loop )
3037 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3038 continue;
3039 }
3041 // Dump controlling node
3042 for( uint x = 0; x < loop->_nest; x++ )
3043 tty->print(" ");
3044 tty->print("C");
3045 if( n == C->root() ) {
3046 n->dump();
3047 } else {
3048 Node* cached_idom = idom_no_update(n);
3049 Node *computed_idom = n->in(0);
3050 if( n->is_Region() ) {
3051 computed_idom = compute_idom(n);
3052 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3053 // any MultiBranch ctrl node), so apply a similar transform to
3054 // the cached idom returned from idom_no_update.
3055 cached_idom = find_non_split_ctrl(cached_idom);
3056 }
3057 tty->print(" ID:%d",computed_idom->_idx);
3058 n->dump();
3059 if( cached_idom != computed_idom ) {
3060 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3061 computed_idom->_idx, cached_idom->_idx);
3062 }
3063 }
3064 // Dump nodes it controls
3065 for( uint k = 0; k < _nodes.Size(); k++ ) {
3066 // (k < C->unique() && get_ctrl(find(k)) == n)
3067 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3068 Node *m = C->root()->find(k);
3069 if( m && m->outcnt() > 0 ) {
3070 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3071 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3072 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3073 }
3074 for( uint j = 0; j < loop->_nest; j++ )
3075 tty->print(" ");
3076 tty->print(" ");
3077 m->dump();
3078 }
3079 }
3080 }
3081 }
3082 }
3084 // Collect a R-P-O for the whole CFG.
3085 // Result list is in post-order (scan backwards for RPO)
3086 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3087 stk.push(start, 0);
3088 visited.set(start->_idx);
3090 while (stk.is_nonempty()) {
3091 Node* m = stk.node();
3092 uint idx = stk.index();
3093 if (idx < m->outcnt()) {
3094 stk.set_index(idx + 1);
3095 Node* n = m->raw_out(idx);
3096 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3097 stk.push(n, 0);
3098 }
3099 } else {
3100 rpo_list.push(m);
3101 stk.pop();
3102 }
3103 }
3104 }
3105 #endif
3108 //=============================================================================
3109 //------------------------------LoopTreeIterator-----------------------------------
3111 // Advance to next loop tree using a preorder, left-to-right traversal.
3112 void LoopTreeIterator::next() {
3113 assert(!done(), "must not be done.");
3114 if (_curnt->_child != NULL) {
3115 _curnt = _curnt->_child;
3116 } else if (_curnt->_next != NULL) {
3117 _curnt = _curnt->_next;
3118 } else {
3119 while (_curnt != _root && _curnt->_next == NULL) {
3120 _curnt = _curnt->_parent;
3121 }
3122 if (_curnt == _root) {
3123 _curnt = NULL;
3124 assert(done(), "must be done.");
3125 } else {
3126 assert(_curnt->_next != NULL, "must be more to do");
3127 _curnt = _curnt->_next;
3128 }
3129 }
3130 }