Fri, 25 Mar 2011 09:35:39 +0100
7029017: Additional architecture support for c2 compiler
Summary: Enables cross building of a c2 VM. Support masking of shift counts when the processor architecture mandates it.
Reviewed-by: kvn, 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);
345 // If compare points to incr, we are ok. Otherwise the compare
346 // can directly point to the phi; in this case adjust the compare so that
347 // it points to the incr by adjusting the limit.
348 if (cmp->in(1) == phi || cmp->in(2) == phi)
349 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
351 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
352 // Final value for iterator should be: trip_count * stride + init_trip.
353 Node *one_p = gvn->intcon( 1);
354 Node *one_m = gvn->intcon(-1);
356 Node *trip_count = NULL;
357 Node *hook = new (C, 6) Node(6);
358 switch( bt ) {
359 case BoolTest::eq:
360 ShouldNotReachHere();
361 case BoolTest::ne: // Ahh, the case we desire
362 if (stride_con == 1)
363 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
364 else if (stride_con == -1)
365 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
366 else
367 ShouldNotReachHere();
368 set_subtree_ctrl(trip_count);
369 //_loop.map(trip_count->_idx,loop(limit));
370 break;
371 case BoolTest::le: // Maybe convert to '<' case
372 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
373 set_subtree_ctrl( limit );
374 hook->init_req(4, limit);
376 bt = BoolTest::lt;
377 // Make the new limit be in the same loop nest as the old limit
378 //_loop.map(limit->_idx,limit_loop);
379 // Fall into next case
380 case BoolTest::lt: { // Maybe convert to '!=' case
381 if (stride_con < 0) // Count down loop rolls through MAXINT
382 ShouldNotReachHere();
383 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
384 set_subtree_ctrl( range );
385 hook->init_req(0, range);
387 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
388 set_subtree_ctrl( bias );
389 hook->init_req(1, bias);
391 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
392 set_subtree_ctrl( bias1 );
393 hook->init_req(2, bias1);
395 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
396 set_subtree_ctrl( trip_count );
397 hook->init_req(3, trip_count);
398 break;
399 }
401 case BoolTest::ge: // Maybe convert to '>' case
402 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
403 set_subtree_ctrl( limit );
404 hook->init_req(4 ,limit);
406 bt = BoolTest::gt;
407 // Make the new limit be in the same loop nest as the old limit
408 //_loop.map(limit->_idx,limit_loop);
409 // Fall into next case
410 case BoolTest::gt: { // Maybe convert to '!=' case
411 if (stride_con > 0) // count up loop rolls through MININT
412 ShouldNotReachHere();
413 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
414 set_subtree_ctrl( range );
415 hook->init_req(0, range);
417 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
418 set_subtree_ctrl( bias );
419 hook->init_req(1, bias);
421 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
422 set_subtree_ctrl( bias1 );
423 hook->init_req(2, bias1);
425 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
426 set_subtree_ctrl( trip_count );
427 hook->init_req(3, trip_count);
428 break;
429 }
430 } // switch( bt )
432 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
433 set_subtree_ctrl( span );
434 hook->init_req(5, span);
436 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
437 set_subtree_ctrl( limit );
439 // Check for SafePoint on backedge and remove
440 Node *sfpt = x->in(LoopNode::LoopBackControl);
441 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
442 lazy_replace( sfpt, iftrue );
443 loop->_tail = iftrue;
444 }
446 // Build a canonical trip test.
447 // Clone code, as old values may be in use.
448 Node* nphi = PhiNode::make(x, init_trip, TypeInt::INT);
449 nphi = _igvn.register_new_node_with_optimizer(nphi);
450 set_ctrl(nphi, get_ctrl(phi));
452 incr = incr->clone();
453 incr->set_req(1,nphi);
454 incr->set_req(2,stride);
455 incr = _igvn.register_new_node_with_optimizer(incr);
456 set_early_ctrl( incr );
458 nphi->set_req(LoopNode::LoopBackControl, incr);
459 _igvn.replace_node(phi, nphi);
460 phi = nphi->as_Phi();
462 cmp = cmp->clone();
463 cmp->set_req(1,incr);
464 cmp->set_req(2,limit);
465 cmp = _igvn.register_new_node_with_optimizer(cmp);
466 set_ctrl(cmp, iff->in(0));
468 test = test->clone()->as_Bool();
469 (*(BoolTest*)&test->_test)._test = bt;
470 test->set_req(1,cmp);
471 _igvn.register_new_node_with_optimizer(test);
472 set_ctrl(test, iff->in(0));
474 // Replace the old IfNode with a new LoopEndNode
475 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
476 IfNode *le = lex->as_If();
477 uint dd = dom_depth(iff);
478 set_idom(le, le->in(0), dd); // Update dominance for loop exit
479 set_loop(le, loop);
481 // Get the loop-exit control
482 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
484 // Need to swap loop-exit and loop-back control?
485 if (iftrue_op == Op_IfFalse) {
486 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
487 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
489 loop->_tail = back_control = ift2;
490 set_loop(ift2, loop);
491 set_loop(iff2, get_loop(iffalse));
493 // Lazy update of 'get_ctrl' mechanism.
494 lazy_replace_proj( iffalse, iff2 );
495 lazy_replace_proj( iftrue, ift2 );
497 // Swap names
498 iffalse = iff2;
499 iftrue = ift2;
500 } else {
501 _igvn.hash_delete(iffalse);
502 _igvn.hash_delete(iftrue);
503 iffalse->set_req_X( 0, le, &_igvn );
504 iftrue ->set_req_X( 0, le, &_igvn );
505 }
507 set_idom(iftrue, le, dd+1);
508 set_idom(iffalse, le, dd+1);
509 assert(iff->outcnt() == 0, "should be dead now");
510 lazy_replace( iff, le ); // fix 'get_ctrl'
512 // Now setup a new CountedLoopNode to replace the existing LoopNode
513 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
514 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
515 // The following assert is approximately true, and defines the intention
516 // of can_be_counted_loop. It fails, however, because phase->type
517 // is not yet initialized for this loop and its parts.
518 //assert(l->can_be_counted_loop(this), "sanity");
519 _igvn.register_new_node_with_optimizer(l);
520 set_loop(l, loop);
521 loop->_head = l;
522 // Fix all data nodes placed at the old loop head.
523 // Uses the lazy-update mechanism of 'get_ctrl'.
524 lazy_replace( x, l );
525 set_idom(l, init_control, dom_depth(x));
527 // Check for immediately preceding SafePoint and remove
528 Node *sfpt2 = le->in(0);
529 if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
530 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
532 // Free up intermediate goo
533 _igvn.remove_dead_node(hook);
535 #ifdef ASSERT
536 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
537 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
538 #endif
540 C->print_method("After CountedLoop", 3);
542 return true;
543 }
546 //------------------------------Ideal------------------------------------------
547 // Return a node which is more "ideal" than the current node.
548 // Attempt to convert into a counted-loop.
549 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
550 if (!can_be_counted_loop(phase)) {
551 phase->C->set_major_progress();
552 }
553 return RegionNode::Ideal(phase, can_reshape);
554 }
557 //=============================================================================
558 //------------------------------Ideal------------------------------------------
559 // Return a node which is more "ideal" than the current node.
560 // Attempt to convert into a counted-loop.
561 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
562 return RegionNode::Ideal(phase, can_reshape);
563 }
565 //------------------------------dump_spec--------------------------------------
566 // Dump special per-node info
567 #ifndef PRODUCT
568 void CountedLoopNode::dump_spec(outputStream *st) const {
569 LoopNode::dump_spec(st);
570 if( stride_is_con() ) {
571 st->print("stride: %d ",stride_con());
572 } else {
573 st->print("stride: not constant ");
574 }
575 if( is_pre_loop () ) st->print("pre of N%d" , _main_idx );
576 if( is_main_loop() ) st->print("main of N%d", _idx );
577 if( is_post_loop() ) st->print("post of N%d", _main_idx );
578 }
579 #endif
581 //=============================================================================
582 int CountedLoopEndNode::stride_con() const {
583 return stride()->bottom_type()->is_int()->get_con();
584 }
587 //----------------------match_incr_with_optional_truncation--------------------
588 // Match increment with optional truncation:
589 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
590 // Return NULL for failure. Success returns the increment node.
591 Node* CountedLoopNode::match_incr_with_optional_truncation(
592 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
593 // Quick cutouts:
594 if (expr == NULL || expr->req() != 3) return false;
596 Node *t1 = NULL;
597 Node *t2 = NULL;
598 const TypeInt* trunc_t = TypeInt::INT;
599 Node* n1 = expr;
600 int n1op = n1->Opcode();
602 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
603 if (n1op == Op_AndI &&
604 n1->in(2)->is_Con() &&
605 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
606 // %%% This check should match any mask of 2**K-1.
607 t1 = n1;
608 n1 = t1->in(1);
609 n1op = n1->Opcode();
610 trunc_t = TypeInt::CHAR;
611 } else if (n1op == Op_RShiftI &&
612 n1->in(1) != NULL &&
613 n1->in(1)->Opcode() == Op_LShiftI &&
614 n1->in(2) == n1->in(1)->in(2) &&
615 n1->in(2)->is_Con()) {
616 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
617 // %%% This check should match any shift in [1..31].
618 if (shift == 16 || shift == 8) {
619 t1 = n1;
620 t2 = t1->in(1);
621 n1 = t2->in(1);
622 n1op = n1->Opcode();
623 if (shift == 16) {
624 trunc_t = TypeInt::SHORT;
625 } else if (shift == 8) {
626 trunc_t = TypeInt::BYTE;
627 }
628 }
629 }
631 // If (maybe after stripping) it is an AddI, we won:
632 if (n1op == Op_AddI) {
633 *trunc1 = t1;
634 *trunc2 = t2;
635 *trunc_type = trunc_t;
636 return n1;
637 }
639 // failed
640 return NULL;
641 }
644 //------------------------------filtered_type--------------------------------
645 // Return a type based on condition control flow
646 // A successful return will be a type that is restricted due
647 // to a series of dominating if-tests, such as:
648 // if (i < 10) {
649 // if (i > 0) {
650 // here: "i" type is [1..10)
651 // }
652 // }
653 // or a control flow merge
654 // if (i < 10) {
655 // do {
656 // phi( , ) -- at top of loop type is [min_int..10)
657 // i = ?
658 // } while ( i < 10)
659 //
660 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
661 assert(n && n->bottom_type()->is_int(), "must be int");
662 const TypeInt* filtered_t = NULL;
663 if (!n->is_Phi()) {
664 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
665 filtered_t = filtered_type_from_dominators(n, n_ctrl);
667 } else {
668 Node* phi = n->as_Phi();
669 Node* region = phi->in(0);
670 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
671 if (region && region != C->top()) {
672 for (uint i = 1; i < phi->req(); i++) {
673 Node* val = phi->in(i);
674 Node* use_c = region->in(i);
675 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
676 if (val_t != NULL) {
677 if (filtered_t == NULL) {
678 filtered_t = val_t;
679 } else {
680 filtered_t = filtered_t->meet(val_t)->is_int();
681 }
682 }
683 }
684 }
685 }
686 const TypeInt* n_t = _igvn.type(n)->is_int();
687 if (filtered_t != NULL) {
688 n_t = n_t->join(filtered_t)->is_int();
689 }
690 return n_t;
691 }
694 //------------------------------filtered_type_from_dominators--------------------------------
695 // Return a possibly more restrictive type for val based on condition control flow of dominators
696 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
697 if (val->is_Con()) {
698 return val->bottom_type()->is_int();
699 }
700 uint if_limit = 10; // Max number of dominating if's visited
701 const TypeInt* rtn_t = NULL;
703 if (use_ctrl && use_ctrl != C->top()) {
704 Node* val_ctrl = get_ctrl(val);
705 uint val_dom_depth = dom_depth(val_ctrl);
706 Node* pred = use_ctrl;
707 uint if_cnt = 0;
708 while (if_cnt < if_limit) {
709 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
710 if_cnt++;
711 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
712 if (if_t != NULL) {
713 if (rtn_t == NULL) {
714 rtn_t = if_t;
715 } else {
716 rtn_t = rtn_t->join(if_t)->is_int();
717 }
718 }
719 }
720 pred = idom(pred);
721 if (pred == NULL || pred == C->top()) {
722 break;
723 }
724 // Stop if going beyond definition block of val
725 if (dom_depth(pred) < val_dom_depth) {
726 break;
727 }
728 }
729 }
730 return rtn_t;
731 }
734 //------------------------------dump_spec--------------------------------------
735 // Dump special per-node info
736 #ifndef PRODUCT
737 void CountedLoopEndNode::dump_spec(outputStream *st) const {
738 if( in(TestValue)->is_Bool() ) {
739 BoolTest bt( test_trip()); // Added this for g++.
741 st->print("[");
742 bt.dump_on(st);
743 st->print("]");
744 }
745 st->print(" ");
746 IfNode::dump_spec(st);
747 }
748 #endif
750 //=============================================================================
751 //------------------------------is_member--------------------------------------
752 // Is 'l' a member of 'this'?
753 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
754 while( l->_nest > _nest ) l = l->_parent;
755 return l == this;
756 }
758 //------------------------------set_nest---------------------------------------
759 // Set loop tree nesting depth. Accumulate _has_call bits.
760 int IdealLoopTree::set_nest( uint depth ) {
761 _nest = depth;
762 int bits = _has_call;
763 if( _child ) bits |= _child->set_nest(depth+1);
764 if( bits ) _has_call = 1;
765 if( _next ) bits |= _next ->set_nest(depth );
766 return bits;
767 }
769 //------------------------------split_fall_in----------------------------------
770 // Split out multiple fall-in edges from the loop header. Move them to a
771 // private RegionNode before the loop. This becomes the loop landing pad.
772 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
773 PhaseIterGVN &igvn = phase->_igvn;
774 uint i;
776 // Make a new RegionNode to be the landing pad.
777 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
778 phase->set_loop(landing_pad,_parent);
779 // Gather all the fall-in control paths into the landing pad
780 uint icnt = fall_in_cnt;
781 uint oreq = _head->req();
782 for( i = oreq-1; i>0; i-- )
783 if( !phase->is_member( this, _head->in(i) ) )
784 landing_pad->set_req(icnt--,_head->in(i));
786 // Peel off PhiNode edges as well
787 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
788 Node *oj = _head->fast_out(j);
789 if( oj->is_Phi() ) {
790 PhiNode* old_phi = oj->as_Phi();
791 assert( old_phi->region() == _head, "" );
792 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
793 Node *p = PhiNode::make_blank(landing_pad, old_phi);
794 uint icnt = fall_in_cnt;
795 for( i = oreq-1; i>0; i-- ) {
796 if( !phase->is_member( this, _head->in(i) ) ) {
797 p->init_req(icnt--, old_phi->in(i));
798 // Go ahead and clean out old edges from old phi
799 old_phi->del_req(i);
800 }
801 }
802 // Search for CSE's here, because ZKM.jar does a lot of
803 // loop hackery and we need to be a little incremental
804 // with the CSE to avoid O(N^2) node blow-up.
805 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
806 if( p2 ) { // Found CSE
807 p->destruct(); // Recover useless new node
808 p = p2; // Use old node
809 } else {
810 igvn.register_new_node_with_optimizer(p, old_phi);
811 }
812 // Make old Phi refer to new Phi.
813 old_phi->add_req(p);
814 // Check for the special case of making the old phi useless and
815 // disappear it. In JavaGrande I have a case where this useless
816 // Phi is the loop limit and prevents recognizing a CountedLoop
817 // which in turn prevents removing an empty loop.
818 Node *id_old_phi = old_phi->Identity( &igvn );
819 if( id_old_phi != old_phi ) { // Found a simple identity?
820 // Note that I cannot call 'replace_node' here, because
821 // that will yank the edge from old_phi to the Region and
822 // I'm mid-iteration over the Region's uses.
823 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
824 Node* use = old_phi->last_out(i);
825 igvn.hash_delete(use);
826 igvn._worklist.push(use);
827 uint uses_found = 0;
828 for (uint j = 0; j < use->len(); j++) {
829 if (use->in(j) == old_phi) {
830 if (j < use->req()) use->set_req (j, id_old_phi);
831 else use->set_prec(j, id_old_phi);
832 uses_found++;
833 }
834 }
835 i -= uses_found; // we deleted 1 or more copies of this edge
836 }
837 }
838 igvn._worklist.push(old_phi);
839 }
840 }
841 // Finally clean out the fall-in edges from the RegionNode
842 for( i = oreq-1; i>0; i-- ) {
843 if( !phase->is_member( this, _head->in(i) ) ) {
844 _head->del_req(i);
845 }
846 }
847 // Transform landing pad
848 igvn.register_new_node_with_optimizer(landing_pad, _head);
849 // Insert landing pad into the header
850 _head->add_req(landing_pad);
851 }
853 //------------------------------split_outer_loop-------------------------------
854 // Split out the outermost loop from this shared header.
855 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
856 PhaseIterGVN &igvn = phase->_igvn;
858 // Find index of outermost loop; it should also be my tail.
859 uint outer_idx = 1;
860 while( _head->in(outer_idx) != _tail ) outer_idx++;
862 // Make a LoopNode for the outermost loop.
863 Node *ctl = _head->in(LoopNode::EntryControl);
864 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
865 outer = igvn.register_new_node_with_optimizer(outer, _head);
866 phase->set_created_loop_node();
867 // Outermost loop falls into '_head' loop
868 _head->set_req(LoopNode::EntryControl, outer);
869 _head->del_req(outer_idx);
870 // Split all the Phis up between '_head' loop and 'outer' loop.
871 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
872 Node *out = _head->fast_out(j);
873 if( out->is_Phi() ) {
874 PhiNode *old_phi = out->as_Phi();
875 assert( old_phi->region() == _head, "" );
876 Node *phi = PhiNode::make_blank(outer, old_phi);
877 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
878 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
879 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
880 // Make old Phi point to new Phi on the fall-in path
881 igvn.hash_delete(old_phi);
882 old_phi->set_req(LoopNode::EntryControl, phi);
883 old_phi->del_req(outer_idx);
884 igvn._worklist.push(old_phi);
885 }
886 }
888 // Use the new loop head instead of the old shared one
889 _head = outer;
890 phase->set_loop(_head, this);
891 }
893 //------------------------------fix_parent-------------------------------------
894 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
895 loop->_parent = parent;
896 if( loop->_child ) fix_parent( loop->_child, loop );
897 if( loop->_next ) fix_parent( loop->_next , parent );
898 }
900 //------------------------------estimate_path_freq-----------------------------
901 static float estimate_path_freq( Node *n ) {
902 // Try to extract some path frequency info
903 IfNode *iff;
904 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
905 uint nop = n->Opcode();
906 if( nop == Op_SafePoint ) { // Skip any safepoint
907 n = n->in(0);
908 continue;
909 }
910 if( nop == Op_CatchProj ) { // Get count from a prior call
911 // Assume call does not always throw exceptions: means the call-site
912 // count is also the frequency of the fall-through path.
913 assert( n->is_CatchProj(), "" );
914 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
915 return 0.0f; // Assume call exception path is rare
916 Node *call = n->in(0)->in(0)->in(0);
917 assert( call->is_Call(), "expect a call here" );
918 const JVMState *jvms = ((CallNode*)call)->jvms();
919 ciMethodData* methodData = jvms->method()->method_data();
920 if (!methodData->is_mature()) return 0.0f; // No call-site data
921 ciProfileData* data = methodData->bci_to_data(jvms->bci());
922 if ((data == NULL) || !data->is_CounterData()) {
923 // no call profile available, try call's control input
924 n = n->in(0);
925 continue;
926 }
927 return data->as_CounterData()->count()/FreqCountInvocations;
928 }
929 // See if there's a gating IF test
930 Node *n_c = n->in(0);
931 if( !n_c->is_If() ) break; // No estimate available
932 iff = n_c->as_If();
933 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
934 // Compute how much count comes on this path
935 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
936 // Have no count info. Skip dull uncommon-trap like branches.
937 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
938 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
939 break;
940 // Skip through never-taken branch; look for a real loop exit.
941 n = iff->in(0);
942 }
943 return 0.0f; // No estimate available
944 }
946 //------------------------------merge_many_backedges---------------------------
947 // Merge all the backedges from the shared header into a private Region.
948 // Feed that region as the one backedge to this loop.
949 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
950 uint i;
952 // Scan for the top 2 hottest backedges
953 float hotcnt = 0.0f;
954 float warmcnt = 0.0f;
955 uint hot_idx = 0;
956 // Loop starts at 2 because slot 1 is the fall-in path
957 for( i = 2; i < _head->req(); i++ ) {
958 float cnt = estimate_path_freq(_head->in(i));
959 if( cnt > hotcnt ) { // Grab hottest path
960 warmcnt = hotcnt;
961 hotcnt = cnt;
962 hot_idx = i;
963 } else if( cnt > warmcnt ) { // And 2nd hottest path
964 warmcnt = cnt;
965 }
966 }
968 // See if the hottest backedge is worthy of being an inner loop
969 // by being much hotter than the next hottest backedge.
970 if( hotcnt <= 0.0001 ||
971 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
973 // Peel out the backedges into a private merge point; peel
974 // them all except optionally hot_idx.
975 PhaseIterGVN &igvn = phase->_igvn;
977 Node *hot_tail = NULL;
978 // Make a Region for the merge point
979 Node *r = new (phase->C, 1) RegionNode(1);
980 for( i = 2; i < _head->req(); i++ ) {
981 if( i != hot_idx )
982 r->add_req( _head->in(i) );
983 else hot_tail = _head->in(i);
984 }
985 igvn.register_new_node_with_optimizer(r, _head);
986 // Plug region into end of loop _head, followed by hot_tail
987 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
988 _head->set_req(2, r);
989 if( hot_idx ) _head->add_req(hot_tail);
991 // Split all the Phis up between '_head' loop and the Region 'r'
992 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
993 Node *out = _head->fast_out(j);
994 if( out->is_Phi() ) {
995 PhiNode* n = out->as_Phi();
996 igvn.hash_delete(n); // Delete from hash before hacking edges
997 Node *hot_phi = NULL;
998 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
999 // Check all inputs for the ones to peel out
1000 uint j = 1;
1001 for( uint i = 2; i < n->req(); i++ ) {
1002 if( i != hot_idx )
1003 phi->set_req( j++, n->in(i) );
1004 else hot_phi = n->in(i);
1005 }
1006 // Register the phi but do not transform until whole place transforms
1007 igvn.register_new_node_with_optimizer(phi, n);
1008 // Add the merge phi to the old Phi
1009 while( n->req() > 3 ) n->del_req( n->req()-1 );
1010 n->set_req(2, phi);
1011 if( hot_idx ) n->add_req(hot_phi);
1012 }
1013 }
1016 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1017 // of self loop tree. Turn self into a loop headed by _head and with
1018 // tail being the new merge point.
1019 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1020 phase->set_loop(_tail,ilt); // Adjust tail
1021 _tail = r; // Self's tail is new merge point
1022 phase->set_loop(r,this);
1023 ilt->_child = _child; // New guy has my children
1024 _child = ilt; // Self has new guy as only child
1025 ilt->_parent = this; // new guy has self for parent
1026 ilt->_nest = _nest; // Same nesting depth (for now)
1028 // Starting with 'ilt', look for child loop trees using the same shared
1029 // header. Flatten these out; they will no longer be loops in the end.
1030 IdealLoopTree **pilt = &_child;
1031 while( ilt ) {
1032 if( ilt->_head == _head ) {
1033 uint i;
1034 for( i = 2; i < _head->req(); i++ )
1035 if( _head->in(i) == ilt->_tail )
1036 break; // Still a loop
1037 if( i == _head->req() ) { // No longer a loop
1038 // Flatten ilt. Hang ilt's "_next" list from the end of
1039 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1040 IdealLoopTree **cp = &ilt->_child;
1041 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1042 *cp = ilt->_next; // Hang next list at end of child list
1043 *pilt = ilt->_child; // Move child up to replace ilt
1044 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1045 ilt = ilt->_child; // Repeat using new ilt
1046 continue; // do not advance over ilt->_child
1047 }
1048 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1049 phase->set_loop(_head,ilt);
1050 }
1051 pilt = &ilt->_child; // Advance to next
1052 ilt = *pilt;
1053 }
1055 if( _child ) fix_parent( _child, this );
1056 }
1058 //------------------------------beautify_loops---------------------------------
1059 // Split shared headers and insert loop landing pads.
1060 // Insert a LoopNode to replace the RegionNode.
1061 // Return TRUE if loop tree is structurally changed.
1062 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1063 bool result = false;
1064 // Cache parts in locals for easy
1065 PhaseIterGVN &igvn = phase->_igvn;
1067 phase->C->print_method("Before beautify loops", 3);
1069 igvn.hash_delete(_head); // Yank from hash before hacking edges
1071 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1072 int fall_in_cnt = 0;
1073 for( uint i = 1; i < _head->req(); i++ )
1074 if( !phase->is_member( this, _head->in(i) ) )
1075 fall_in_cnt++;
1076 assert( fall_in_cnt, "at least 1 fall-in path" );
1077 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1078 split_fall_in( phase, fall_in_cnt );
1080 // Swap inputs to the _head and all Phis to move the fall-in edge to
1081 // the left.
1082 fall_in_cnt = 1;
1083 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1084 fall_in_cnt++;
1085 if( fall_in_cnt > 1 ) {
1086 // Since I am just swapping inputs I do not need to update def-use info
1087 Node *tmp = _head->in(1);
1088 _head->set_req( 1, _head->in(fall_in_cnt) );
1089 _head->set_req( fall_in_cnt, tmp );
1090 // Swap also all Phis
1091 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1092 Node* phi = _head->fast_out(i);
1093 if( phi->is_Phi() ) {
1094 igvn.hash_delete(phi); // Yank from hash before hacking edges
1095 tmp = phi->in(1);
1096 phi->set_req( 1, phi->in(fall_in_cnt) );
1097 phi->set_req( fall_in_cnt, tmp );
1098 }
1099 }
1100 }
1101 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1102 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1104 // If I am a shared header (multiple backedges), peel off the many
1105 // backedges into a private merge point and use the merge point as
1106 // the one true backedge.
1107 if( _head->req() > 3 ) {
1108 // Merge the many backedges into a single backedge.
1109 merge_many_backedges( phase );
1110 result = true;
1111 }
1113 // If I am a shared header (multiple backedges), peel off myself loop.
1114 // I better be the outermost loop.
1115 if( _head->req() > 3 ) {
1116 split_outer_loop( phase );
1117 result = true;
1119 } else if( !_head->is_Loop() && !_irreducible ) {
1120 // Make a new LoopNode to replace the old loop head
1121 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1122 l = igvn.register_new_node_with_optimizer(l, _head);
1123 phase->set_created_loop_node();
1124 // Go ahead and replace _head
1125 phase->_igvn.replace_node( _head, l );
1126 _head = l;
1127 phase->set_loop(_head, this);
1128 }
1130 // Now recursively beautify nested loops
1131 if( _child ) result |= _child->beautify_loops( phase );
1132 if( _next ) result |= _next ->beautify_loops( phase );
1133 return result;
1134 }
1136 //------------------------------allpaths_check_safepts----------------------------
1137 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1138 // encountered. Helper for check_safepts.
1139 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1140 assert(stack.size() == 0, "empty stack");
1141 stack.push(_tail);
1142 visited.Clear();
1143 visited.set(_tail->_idx);
1144 while (stack.size() > 0) {
1145 Node* n = stack.pop();
1146 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1147 // Terminate this path
1148 } else if (n->Opcode() == Op_SafePoint) {
1149 if (_phase->get_loop(n) != this) {
1150 if (_required_safept == NULL) _required_safept = new Node_List();
1151 _required_safept->push(n); // save the one closest to the tail
1152 }
1153 // Terminate this path
1154 } else {
1155 uint start = n->is_Region() ? 1 : 0;
1156 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1157 for (uint i = start; i < end; i++) {
1158 Node* in = n->in(i);
1159 assert(in->is_CFG(), "must be");
1160 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1161 stack.push(in);
1162 }
1163 }
1164 }
1165 }
1166 }
1168 //------------------------------check_safepts----------------------------
1169 // Given dominators, try to find loops with calls that must always be
1170 // executed (call dominates loop tail). These loops do not need non-call
1171 // safepoints (ncsfpt).
1172 //
1173 // A complication is that a safepoint in a inner loop may be needed
1174 // by an outer loop. In the following, the inner loop sees it has a
1175 // call (block 3) on every path from the head (block 2) to the
1176 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1177 // in block 2, _but_ this leaves the outer loop without a safepoint.
1178 //
1179 // entry 0
1180 // |
1181 // v
1182 // outer 1,2 +->1
1183 // | |
1184 // | v
1185 // | 2<---+ ncsfpt in 2
1186 // |_/|\ |
1187 // | v |
1188 // inner 2,3 / 3 | call in 3
1189 // / | |
1190 // v +--+
1191 // exit 4
1192 //
1193 //
1194 // This method creates a list (_required_safept) of ncsfpt nodes that must
1195 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1196 // is first looked for in the lists for the outer loops of the current loop.
1197 //
1198 // The insights into the problem:
1199 // A) counted loops are okay
1200 // B) innermost loops are okay (only an inner loop can delete
1201 // a ncsfpt needed by an outer loop)
1202 // C) a loop is immune from an inner loop deleting a safepoint
1203 // if the loop has a call on the idom-path
1204 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1205 // idom-path that is not in a nested loop
1206 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1207 // loop needs to be prevented from deletion by an inner loop
1208 //
1209 // There are two analyses:
1210 // 1) The first, and cheaper one, scans the loop body from
1211 // tail to head following the idom (immediate dominator)
1212 // chain, looking for the cases (C,D,E) above.
1213 // Since inner loops are scanned before outer loops, there is summary
1214 // information about inner loops. Inner loops can be skipped over
1215 // when the tail of an inner loop is encountered.
1216 //
1217 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1218 // the idom path (which is rare), scans all predecessor control paths
1219 // from the tail to the head, terminating a path when a call or sfpt
1220 // is encountered, to find the ncsfpt's that are closest to the tail.
1221 //
1222 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1223 // Bottom up traversal
1224 IdealLoopTree* ch = _child;
1225 while (ch != NULL) {
1226 ch->check_safepts(visited, stack);
1227 ch = ch->_next;
1228 }
1230 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1231 bool has_call = false; // call on dom-path
1232 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1233 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1234 // Scan the dom-path nodes from tail to head
1235 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1236 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1237 has_call = true;
1238 _has_sfpt = 1; // Then no need for a safept!
1239 break;
1240 } else if (n->Opcode() == Op_SafePoint) {
1241 if (_phase->get_loop(n) == this) {
1242 has_local_ncsfpt = true;
1243 break;
1244 }
1245 if (nonlocal_ncsfpt == NULL) {
1246 nonlocal_ncsfpt = n; // save the one closest to the tail
1247 }
1248 } else {
1249 IdealLoopTree* nlpt = _phase->get_loop(n);
1250 if (this != nlpt) {
1251 // If at an inner loop tail, see if the inner loop has already
1252 // recorded seeing a call on the dom-path (and stop.) If not,
1253 // jump to the head of the inner loop.
1254 assert(is_member(nlpt), "nested loop");
1255 Node* tail = nlpt->_tail;
1256 if (tail->in(0)->is_If()) tail = tail->in(0);
1257 if (n == tail) {
1258 // If inner loop has call on dom-path, so does outer loop
1259 if (nlpt->_has_sfpt) {
1260 has_call = true;
1261 _has_sfpt = 1;
1262 break;
1263 }
1264 // Skip to head of inner loop
1265 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1266 n = nlpt->_head;
1267 }
1268 }
1269 }
1270 }
1271 // Record safept's that this loop needs preserved when an
1272 // inner loop attempts to delete it's safepoints.
1273 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1274 if (nonlocal_ncsfpt != NULL) {
1275 if (_required_safept == NULL) _required_safept = new Node_List();
1276 _required_safept->push(nonlocal_ncsfpt);
1277 } else {
1278 // Failed to find a suitable safept on the dom-path. Now use
1279 // an all paths walk from tail to head, looking for safepoints to preserve.
1280 allpaths_check_safepts(visited, stack);
1281 }
1282 }
1283 }
1284 }
1286 //---------------------------is_deleteable_safept----------------------------
1287 // Is safept not required by an outer loop?
1288 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1289 assert(sfpt->Opcode() == Op_SafePoint, "");
1290 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1291 while (lp != NULL) {
1292 Node_List* sfpts = lp->_required_safept;
1293 if (sfpts != NULL) {
1294 for (uint i = 0; i < sfpts->size(); i++) {
1295 if (sfpt == sfpts->at(i))
1296 return false;
1297 }
1298 }
1299 lp = lp->_parent;
1300 }
1301 return true;
1302 }
1304 //---------------------------replace_parallel_iv-------------------------------
1305 // Replace parallel induction variable (parallel to trip counter)
1306 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1307 assert(loop->_head->is_CountedLoop(), "");
1308 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1309 Node *incr = cl->incr();
1310 if (incr == NULL)
1311 return; // Dead loop?
1312 Node *init = cl->init_trip();
1313 Node *phi = cl->phi();
1314 // protect against stride not being a constant
1315 if (!cl->stride_is_con())
1316 return;
1317 int stride_con = cl->stride_con();
1319 PhaseGVN *gvn = &_igvn;
1321 // Visit all children, looking for Phis
1322 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1323 Node *out = cl->out(i);
1324 // Look for other phis (secondary IVs). Skip dead ones
1325 if (!out->is_Phi() || out == phi || !has_node(out))
1326 continue;
1327 PhiNode* phi2 = out->as_Phi();
1328 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1329 // Look for induction variables of the form: X += constant
1330 if (phi2->region() != loop->_head ||
1331 incr2->req() != 3 ||
1332 incr2->in(1) != phi2 ||
1333 incr2 == incr ||
1334 incr2->Opcode() != Op_AddI ||
1335 !incr2->in(2)->is_Con())
1336 continue;
1338 // Check for parallel induction variable (parallel to trip counter)
1339 // via an affine function. In particular, count-down loops with
1340 // count-up array indices are common. We only RCE references off
1341 // the trip-counter, so we need to convert all these to trip-counter
1342 // expressions.
1343 Node *init2 = phi2->in( LoopNode::EntryControl );
1344 int stride_con2 = incr2->in(2)->get_int();
1346 // The general case here gets a little tricky. We want to find the
1347 // GCD of all possible parallel IV's and make a new IV using this
1348 // GCD for the loop. Then all possible IVs are simple multiples of
1349 // the GCD. In practice, this will cover very few extra loops.
1350 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1351 // where +/-1 is the common case, but other integer multiples are
1352 // also easy to handle.
1353 int ratio_con = stride_con2/stride_con;
1355 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1356 // Convert to using the trip counter. The parallel induction
1357 // variable differs from the trip counter by a loop-invariant
1358 // amount, the difference between their respective initial values.
1359 // It is scaled by the 'ratio_con'.
1360 // Perform local Ideal transformation since in most cases ratio == 1.
1361 Node* ratio = _igvn.intcon(ratio_con);
1362 set_ctrl(ratio, C->root());
1363 Node* hook = new (C, 3) Node(3);
1364 Node* ratio_init = gvn->transform(new (C, 3) MulINode(init, ratio));
1365 hook->init_req(0, ratio_init);
1366 Node* diff = gvn->transform(new (C, 3) SubINode(init2, ratio_init));
1367 hook->init_req(1, diff);
1368 Node* ratio_idx = gvn->transform(new (C, 3) MulINode(phi, ratio));
1369 hook->init_req(2, ratio_idx);
1370 Node* add = gvn->transform(new (C, 3) AddINode(ratio_idx, diff));
1371 set_subtree_ctrl(add);
1372 _igvn.replace_node( phi2, add );
1373 // Free up intermediate goo
1374 _igvn.remove_dead_node(hook);
1375 // Sometimes an induction variable is unused
1376 if (add->outcnt() == 0) {
1377 _igvn.remove_dead_node(add);
1378 }
1379 --i; // deleted this phi; rescan starting with next position
1380 continue;
1381 }
1382 }
1383 }
1385 //------------------------------counted_loop-----------------------------------
1386 // Convert to counted loops where possible
1387 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1389 // For grins, set the inner-loop flag here
1390 if (!_child) {
1391 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1392 }
1394 if (_head->is_CountedLoop() ||
1395 phase->is_counted_loop(_head, this)) {
1396 _has_sfpt = 1; // Indicate we do not need a safepoint here
1398 // Look for a safepoint to remove
1399 for (Node* n = tail(); n != _head; n = phase->idom(n))
1400 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1401 phase->is_deleteable_safept(n))
1402 phase->lazy_replace(n,n->in(TypeFunc::Control));
1404 // Look for induction variables
1405 phase->replace_parallel_iv(this);
1407 } else if (_parent != NULL && !_irreducible) {
1408 // Not a counted loop.
1409 // Look for a safepoint on the idom-path to remove, preserving the first one
1410 bool found = false;
1411 Node* n = tail();
1412 for (; n != _head && !found; n = phase->idom(n)) {
1413 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1414 found = true; // Found one
1415 }
1416 // Skip past it and delete the others
1417 for (; n != _head; n = phase->idom(n)) {
1418 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1419 phase->is_deleteable_safept(n))
1420 phase->lazy_replace(n,n->in(TypeFunc::Control));
1421 }
1422 }
1424 // Recursively
1425 if (_child) _child->counted_loop( phase );
1426 if (_next) _next ->counted_loop( phase );
1427 }
1429 #ifndef PRODUCT
1430 //------------------------------dump_head--------------------------------------
1431 // Dump 1 liner for loop header info
1432 void IdealLoopTree::dump_head( ) const {
1433 for (uint i=0; i<_nest; i++)
1434 tty->print(" ");
1435 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1436 if (_irreducible) tty->print(" IRREDUCIBLE");
1437 if (UseLoopPredicate) {
1438 Node* entry = _head->in(LoopNode::EntryControl);
1439 if (entry != NULL && entry->is_Proj() &&
1440 PhaseIdealLoop::is_uncommon_trap_if_pattern(entry->as_Proj(), Deoptimization::Reason_predicate)) {
1441 tty->print(" predicated");
1442 }
1443 }
1444 if (_head->is_CountedLoop()) {
1445 CountedLoopNode *cl = _head->as_CountedLoop();
1446 tty->print(" counted");
1447 if (cl->is_pre_loop ()) tty->print(" pre" );
1448 if (cl->is_main_loop()) tty->print(" main");
1449 if (cl->is_post_loop()) tty->print(" post");
1450 }
1451 tty->cr();
1452 }
1454 //------------------------------dump-------------------------------------------
1455 // Dump loops by loop tree
1456 void IdealLoopTree::dump( ) const {
1457 dump_head();
1458 if (_child) _child->dump();
1459 if (_next) _next ->dump();
1460 }
1462 #endif
1464 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1465 if (loop == root) {
1466 if (loop->_child != NULL) {
1467 log->begin_head("loop_tree");
1468 log->end_head();
1469 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1470 log->tail("loop_tree");
1471 assert(loop->_next == NULL, "what?");
1472 }
1473 } else {
1474 Node* head = loop->_head;
1475 log->begin_head("loop");
1476 log->print(" idx='%d' ", head->_idx);
1477 if (loop->_irreducible) log->print("irreducible='1' ");
1478 if (head->is_Loop()) {
1479 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1480 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1481 }
1482 if (head->is_CountedLoop()) {
1483 CountedLoopNode* cl = head->as_CountedLoop();
1484 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1485 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1486 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1487 }
1488 log->end_head();
1489 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1490 log->tail("loop");
1491 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1492 }
1493 }
1495 //---------------------collect_potentially_useful_predicates-----------------------
1496 // Helper function to collect potentially useful predicates to prevent them from
1497 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1498 void PhaseIdealLoop::collect_potentially_useful_predicates(
1499 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1500 if (loop->_child) { // child
1501 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1502 }
1504 // self (only loops that we can apply loop predication may use their predicates)
1505 if (loop->_head->is_Loop() &&
1506 !loop->_irreducible &&
1507 !loop->tail()->is_top()) {
1508 LoopNode* lpn = loop->_head->as_Loop();
1509 Node* entry = lpn->in(LoopNode::EntryControl);
1510 Node* predicate_proj = find_predicate(entry);
1511 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1512 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1513 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1514 }
1515 }
1517 if (loop->_next) { // sibling
1518 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1519 }
1520 }
1522 //------------------------eliminate_useless_predicates-----------------------------
1523 // Eliminate all inserted predicates if they could not be used by loop predication.
1524 void PhaseIdealLoop::eliminate_useless_predicates() {
1525 if (C->predicate_count() == 0)
1526 return; // no predicate left
1528 Unique_Node_List useful_predicates; // to store useful predicates
1529 if (C->has_loops()) {
1530 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1531 }
1533 for (int i = C->predicate_count(); i > 0; i--) {
1534 Node * n = C->predicate_opaque1_node(i-1);
1535 assert(n->Opcode() == Op_Opaque1, "must be");
1536 if (!useful_predicates.member(n)) { // not in the useful list
1537 _igvn.replace_node(n, n->in(1));
1538 }
1539 }
1540 }
1542 //=============================================================================
1543 //----------------------------build_and_optimize-------------------------------
1544 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1545 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1546 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool do_loop_pred) {
1547 ResourceMark rm;
1549 int old_progress = C->major_progress();
1551 // Reset major-progress flag for the driver's heuristics
1552 C->clear_major_progress();
1554 #ifndef PRODUCT
1555 // Capture for later assert
1556 uint unique = C->unique();
1557 _loop_invokes++;
1558 _loop_work += unique;
1559 #endif
1561 // True if the method has at least 1 irreducible loop
1562 _has_irreducible_loops = false;
1564 _created_loop_node = false;
1566 Arena *a = Thread::current()->resource_area();
1567 VectorSet visited(a);
1568 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1569 _nodes.map(C->unique(), NULL);
1570 memset(_nodes.adr(), 0, wordSize * C->unique());
1572 // Pre-build the top-level outermost loop tree entry
1573 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1574 // Do not need a safepoint at the top level
1575 _ltree_root->_has_sfpt = 1;
1577 // Empty pre-order array
1578 allocate_preorders();
1580 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1581 // IdealLoopTree entries. Data nodes are NOT walked.
1582 build_loop_tree();
1583 // Check for bailout, and return
1584 if (C->failing()) {
1585 return;
1586 }
1588 // No loops after all
1589 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
1591 // There should always be an outer loop containing the Root and Return nodes.
1592 // If not, we have a degenerate empty program. Bail out in this case.
1593 if (!has_node(C->root())) {
1594 if (!_verify_only) {
1595 C->clear_major_progress();
1596 C->record_method_not_compilable("empty program detected during loop optimization");
1597 }
1598 return;
1599 }
1601 // Nothing to do, so get out
1602 if( !C->has_loops() && !do_split_ifs && !_verify_me && !_verify_only ) {
1603 _igvn.optimize(); // Cleanup NeverBranches
1604 return;
1605 }
1607 // Set loop nesting depth
1608 _ltree_root->set_nest( 0 );
1610 // Split shared headers and insert loop landing pads.
1611 // Do not bother doing this on the Root loop of course.
1612 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
1613 if( _ltree_root->_child->beautify_loops( this ) ) {
1614 // Re-build loop tree!
1615 _ltree_root->_child = NULL;
1616 _nodes.clear();
1617 reallocate_preorders();
1618 build_loop_tree();
1619 // Check for bailout, and return
1620 if (C->failing()) {
1621 return;
1622 }
1623 // Reset loop nesting depth
1624 _ltree_root->set_nest( 0 );
1626 C->print_method("After beautify loops", 3);
1627 }
1628 }
1630 // Build Dominators for elision of NULL checks & loop finding.
1631 // Since nodes do not have a slot for immediate dominator, make
1632 // a persistent side array for that info indexed on node->_idx.
1633 _idom_size = C->unique();
1634 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1635 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1636 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1637 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1639 Dominators();
1641 if (!_verify_only) {
1642 // As a side effect, Dominators removed any unreachable CFG paths
1643 // into RegionNodes. It doesn't do this test against Root, so
1644 // we do it here.
1645 for( uint i = 1; i < C->root()->req(); i++ ) {
1646 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1647 _igvn.hash_delete(C->root());
1648 C->root()->del_req(i);
1649 _igvn._worklist.push(C->root());
1650 i--; // Rerun same iteration on compressed edges
1651 }
1652 }
1654 // Given dominators, try to find inner loops with calls that must
1655 // always be executed (call dominates loop tail). These loops do
1656 // not need a separate safepoint.
1657 Node_List cisstack(a);
1658 _ltree_root->check_safepts(visited, cisstack);
1659 }
1661 // Walk the DATA nodes and place into loops. Find earliest control
1662 // node. For CFG nodes, the _nodes array starts out and remains
1663 // holding the associated IdealLoopTree pointer. For DATA nodes, the
1664 // _nodes array holds the earliest legal controlling CFG node.
1666 // Allocate stack with enough space to avoid frequent realloc
1667 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
1668 Node_Stack nstack( a, stack_size );
1670 visited.Clear();
1671 Node_List worklist(a);
1672 // Don't need C->root() on worklist since
1673 // it will be processed among C->top() inputs
1674 worklist.push( C->top() );
1675 visited.set( C->top()->_idx ); // Set C->top() as visited now
1676 build_loop_early( visited, worklist, nstack );
1678 // Given early legal placement, try finding counted loops. This placement
1679 // is good enough to discover most loop invariants.
1680 if( !_verify_me && !_verify_only )
1681 _ltree_root->counted_loop( this );
1683 // Find latest loop placement. Find ideal loop placement.
1684 visited.Clear();
1685 init_dom_lca_tags();
1686 // Need C->root() on worklist when processing outs
1687 worklist.push( C->root() );
1688 NOT_PRODUCT( C->verify_graph_edges(); )
1689 worklist.push( C->top() );
1690 build_loop_late( visited, worklist, nstack );
1692 if (_verify_only) {
1693 // restore major progress flag
1694 for (int i = 0; i < old_progress; i++)
1695 C->set_major_progress();
1696 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
1697 assert(_igvn._worklist.size() == 0, "shouldn't push anything");
1698 return;
1699 }
1701 // some parser-inserted loop predicates could never be used by loop
1702 // predication. Eliminate them before loop optimization
1703 if (UseLoopPredicate) {
1704 eliminate_useless_predicates();
1705 }
1707 // clear out the dead code
1708 while(_deadlist.size()) {
1709 _igvn.remove_globally_dead_node(_deadlist.pop());
1710 }
1712 #ifndef PRODUCT
1713 C->verify_graph_edges();
1714 if (_verify_me) { // Nested verify pass?
1715 // Check to see if the verify mode is broken
1716 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
1717 return;
1718 }
1719 if(VerifyLoopOptimizations) verify();
1720 if(TraceLoopOpts && C->has_loops()) {
1721 _ltree_root->dump();
1722 }
1723 #endif
1725 if (ReassociateInvariants) {
1726 // Reassociate invariants and prep for split_thru_phi
1727 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1728 IdealLoopTree* lpt = iter.current();
1729 if (!lpt->is_counted() || !lpt->is_inner()) continue;
1731 lpt->reassociate_invariants(this);
1733 // Because RCE opportunities can be masked by split_thru_phi,
1734 // look for RCE candidates and inhibit split_thru_phi
1735 // on just their loop-phi's for this pass of loop opts
1736 if (SplitIfBlocks && do_split_ifs) {
1737 if (lpt->policy_range_check(this)) {
1738 lpt->_rce_candidate = 1; // = true
1739 }
1740 }
1741 }
1742 }
1744 // Check for aggressive application of split-if and other transforms
1745 // that require basic-block info (like cloning through Phi's)
1746 if( SplitIfBlocks && do_split_ifs ) {
1747 visited.Clear();
1748 split_if_with_blocks( visited, nstack );
1749 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
1750 }
1752 // Perform loop predication before iteration splitting
1753 if (do_loop_pred && C->has_loops() && !C->major_progress()) {
1754 _ltree_root->_child->loop_predication(this);
1755 }
1757 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
1758 if (do_intrinsify_fill()) {
1759 C->set_major_progress();
1760 }
1761 }
1763 // Perform iteration-splitting on inner loops. Split iterations to avoid
1764 // range checks or one-shot null checks.
1766 // If split-if's didn't hack the graph too bad (no CFG changes)
1767 // then do loop opts.
1768 if (C->has_loops() && !C->major_progress()) {
1769 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
1770 _ltree_root->_child->iteration_split( this, worklist );
1771 // No verify after peeling! GCM has hoisted code out of the loop.
1772 // After peeling, the hoisted code could sink inside the peeled area.
1773 // The peeling code does not try to recompute the best location for
1774 // all the code before the peeled area, so the verify pass will always
1775 // complain about it.
1776 }
1777 // Do verify graph edges in any case
1778 NOT_PRODUCT( C->verify_graph_edges(); );
1780 if (!do_split_ifs) {
1781 // We saw major progress in Split-If to get here. We forced a
1782 // pass with unrolling and not split-if, however more split-if's
1783 // might make progress. If the unrolling didn't make progress
1784 // then the major-progress flag got cleared and we won't try
1785 // another round of Split-If. In particular the ever-common
1786 // instance-of/check-cast pattern requires at least 2 rounds of
1787 // Split-If to clear out.
1788 C->set_major_progress();
1789 }
1791 // Repeat loop optimizations if new loops were seen
1792 if (created_loop_node()) {
1793 C->set_major_progress();
1794 }
1796 // Convert scalar to superword operations
1798 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
1799 // SuperWord transform
1800 SuperWord sw(this);
1801 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1802 IdealLoopTree* lpt = iter.current();
1803 if (lpt->is_counted()) {
1804 sw.transform_loop(lpt);
1805 }
1806 }
1807 }
1809 // Cleanup any modified bits
1810 _igvn.optimize();
1812 // disable assert until issue with split_flow_path is resolved (6742111)
1813 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
1814 // "shouldn't introduce irreducible loops");
1816 if (C->log() != NULL) {
1817 log_loop_tree(_ltree_root, _ltree_root, C->log());
1818 }
1819 }
1821 #ifndef PRODUCT
1822 //------------------------------print_statistics-------------------------------
1823 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
1824 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
1825 void PhaseIdealLoop::print_statistics() {
1826 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
1827 }
1829 //------------------------------verify-----------------------------------------
1830 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
1831 static int fail; // debug only, so its multi-thread dont care
1832 void PhaseIdealLoop::verify() const {
1833 int old_progress = C->major_progress();
1834 ResourceMark rm;
1835 PhaseIdealLoop loop_verify( _igvn, this );
1836 VectorSet visited(Thread::current()->resource_area());
1838 fail = 0;
1839 verify_compare( C->root(), &loop_verify, visited );
1840 assert( fail == 0, "verify loops failed" );
1841 // Verify loop structure is the same
1842 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
1843 // Reset major-progress. It was cleared by creating a verify version of
1844 // PhaseIdealLoop.
1845 for( int i=0; i<old_progress; i++ )
1846 C->set_major_progress();
1847 }
1849 //------------------------------verify_compare---------------------------------
1850 // Make sure me and the given PhaseIdealLoop agree on key data structures
1851 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
1852 if( !n ) return;
1853 if( visited.test_set( n->_idx ) ) return;
1854 if( !_nodes[n->_idx] ) { // Unreachable
1855 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
1856 return;
1857 }
1859 uint i;
1860 for( i = 0; i < n->req(); i++ )
1861 verify_compare( n->in(i), loop_verify, visited );
1863 // Check the '_nodes' block/loop structure
1864 i = n->_idx;
1865 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
1866 if( _nodes[i] != loop_verify->_nodes[i] &&
1867 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
1868 tty->print("Mismatched control setting for: ");
1869 n->dump();
1870 if( fail++ > 10 ) return;
1871 Node *c = get_ctrl_no_update(n);
1872 tty->print("We have it as: ");
1873 if( c->in(0) ) c->dump();
1874 else tty->print_cr("N%d",c->_idx);
1875 tty->print("Verify thinks: ");
1876 if( loop_verify->has_ctrl(n) )
1877 loop_verify->get_ctrl_no_update(n)->dump();
1878 else
1879 loop_verify->get_loop_idx(n)->dump();
1880 tty->cr();
1881 }
1882 } else { // We have a loop
1883 IdealLoopTree *us = get_loop_idx(n);
1884 if( loop_verify->has_ctrl(n) ) {
1885 tty->print("Mismatched loop setting for: ");
1886 n->dump();
1887 if( fail++ > 10 ) return;
1888 tty->print("We have it as: ");
1889 us->dump();
1890 tty->print("Verify thinks: ");
1891 loop_verify->get_ctrl_no_update(n)->dump();
1892 tty->cr();
1893 } else if (!C->major_progress()) {
1894 // Loop selection can be messed up if we did a major progress
1895 // operation, like split-if. Do not verify in that case.
1896 IdealLoopTree *them = loop_verify->get_loop_idx(n);
1897 if( us->_head != them->_head || us->_tail != them->_tail ) {
1898 tty->print("Unequals loops for: ");
1899 n->dump();
1900 if( fail++ > 10 ) return;
1901 tty->print("We have it as: ");
1902 us->dump();
1903 tty->print("Verify thinks: ");
1904 them->dump();
1905 tty->cr();
1906 }
1907 }
1908 }
1910 // Check for immediate dominators being equal
1911 if( i >= _idom_size ) {
1912 if( !n->is_CFG() ) return;
1913 tty->print("CFG Node with no idom: ");
1914 n->dump();
1915 return;
1916 }
1917 if( !n->is_CFG() ) return;
1918 if( n == C->root() ) return; // No IDOM here
1920 assert(n->_idx == i, "sanity");
1921 Node *id = idom_no_update(n);
1922 if( id != loop_verify->idom_no_update(n) ) {
1923 tty->print("Unequals idoms for: ");
1924 n->dump();
1925 if( fail++ > 10 ) return;
1926 tty->print("We have it as: ");
1927 id->dump();
1928 tty->print("Verify thinks: ");
1929 loop_verify->idom_no_update(n)->dump();
1930 tty->cr();
1931 }
1933 }
1935 //------------------------------verify_tree------------------------------------
1936 // Verify that tree structures match. Because the CFG can change, siblings
1937 // within the loop tree can be reordered. We attempt to deal with that by
1938 // reordering the verify's loop tree if possible.
1939 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
1940 assert( _parent == parent, "Badly formed loop tree" );
1942 // Siblings not in same order? Attempt to re-order.
1943 if( _head != loop->_head ) {
1944 // Find _next pointer to update
1945 IdealLoopTree **pp = &loop->_parent->_child;
1946 while( *pp != loop )
1947 pp = &((*pp)->_next);
1948 // Find proper sibling to be next
1949 IdealLoopTree **nn = &loop->_next;
1950 while( (*nn) && (*nn)->_head != _head )
1951 nn = &((*nn)->_next);
1953 // Check for no match.
1954 if( !(*nn) ) {
1955 // Annoyingly, irreducible loops can pick different headers
1956 // after a major_progress operation, so the rest of the loop
1957 // tree cannot be matched.
1958 if (_irreducible && Compile::current()->major_progress()) return;
1959 assert( 0, "failed to match loop tree" );
1960 }
1962 // Move (*nn) to (*pp)
1963 IdealLoopTree *hit = *nn;
1964 *nn = hit->_next;
1965 hit->_next = loop;
1966 *pp = loop;
1967 loop = hit;
1968 // Now try again to verify
1969 }
1971 assert( _head == loop->_head , "mismatched loop head" );
1972 Node *tail = _tail; // Inline a non-updating version of
1973 while( !tail->in(0) ) // the 'tail()' call.
1974 tail = tail->in(1);
1975 assert( tail == loop->_tail, "mismatched loop tail" );
1977 // Counted loops that are guarded should be able to find their guards
1978 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
1979 CountedLoopNode *cl = _head->as_CountedLoop();
1980 Node *init = cl->init_trip();
1981 Node *ctrl = cl->in(LoopNode::EntryControl);
1982 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1983 Node *iff = ctrl->in(0);
1984 assert( iff->Opcode() == Op_If, "" );
1985 Node *bol = iff->in(1);
1986 assert( bol->Opcode() == Op_Bool, "" );
1987 Node *cmp = bol->in(1);
1988 assert( cmp->Opcode() == Op_CmpI, "" );
1989 Node *add = cmp->in(1);
1990 Node *opaq;
1991 if( add->Opcode() == Op_Opaque1 ) {
1992 opaq = add;
1993 } else {
1994 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
1995 assert( add == init, "" );
1996 opaq = cmp->in(2);
1997 }
1998 assert( opaq->Opcode() == Op_Opaque1, "" );
2000 }
2002 if (_child != NULL) _child->verify_tree(loop->_child, this);
2003 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2004 // Innermost loops need to verify loop bodies,
2005 // but only if no 'major_progress'
2006 int fail = 0;
2007 if (!Compile::current()->major_progress() && _child == NULL) {
2008 for( uint i = 0; i < _body.size(); i++ ) {
2009 Node *n = _body.at(i);
2010 if (n->outcnt() == 0) continue; // Ignore dead
2011 uint j;
2012 for( j = 0; j < loop->_body.size(); j++ )
2013 if( loop->_body.at(j) == n )
2014 break;
2015 if( j == loop->_body.size() ) { // Not found in loop body
2016 // Last ditch effort to avoid assertion: Its possible that we
2017 // have some users (so outcnt not zero) but are still dead.
2018 // Try to find from root.
2019 if (Compile::current()->root()->find(n->_idx)) {
2020 fail++;
2021 tty->print("We have that verify does not: ");
2022 n->dump();
2023 }
2024 }
2025 }
2026 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2027 Node *n = loop->_body.at(i2);
2028 if (n->outcnt() == 0) continue; // Ignore dead
2029 uint j;
2030 for( j = 0; j < _body.size(); j++ )
2031 if( _body.at(j) == n )
2032 break;
2033 if( j == _body.size() ) { // Not found in loop body
2034 // Last ditch effort to avoid assertion: Its possible that we
2035 // have some users (so outcnt not zero) but are still dead.
2036 // Try to find from root.
2037 if (Compile::current()->root()->find(n->_idx)) {
2038 fail++;
2039 tty->print("Verify has that we do not: ");
2040 n->dump();
2041 }
2042 }
2043 }
2044 assert( !fail, "loop body mismatch" );
2045 }
2046 }
2048 #endif
2050 //------------------------------set_idom---------------------------------------
2051 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2052 uint idx = d->_idx;
2053 if (idx >= _idom_size) {
2054 uint newsize = _idom_size<<1;
2055 while( idx >= newsize ) {
2056 newsize <<= 1;
2057 }
2058 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2059 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2060 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2061 _idom_size = newsize;
2062 }
2063 _idom[idx] = n;
2064 _dom_depth[idx] = dom_depth;
2065 }
2067 //------------------------------recompute_dom_depth---------------------------------------
2068 // The dominator tree is constructed with only parent pointers.
2069 // This recomputes the depth in the tree by first tagging all
2070 // nodes as "no depth yet" marker. The next pass then runs up
2071 // the dom tree from each node marked "no depth yet", and computes
2072 // the depth on the way back down.
2073 void PhaseIdealLoop::recompute_dom_depth() {
2074 uint no_depth_marker = C->unique();
2075 uint i;
2076 // Initialize depth to "no depth yet"
2077 for (i = 0; i < _idom_size; i++) {
2078 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2079 _dom_depth[i] = no_depth_marker;
2080 }
2081 }
2082 if (_dom_stk == NULL) {
2083 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
2084 if (init_size < 10) init_size = 10;
2085 _dom_stk = new GrowableArray<uint>(init_size);
2086 }
2087 // Compute new depth for each node.
2088 for (i = 0; i < _idom_size; i++) {
2089 uint j = i;
2090 // Run up the dom tree to find a node with a depth
2091 while (_dom_depth[j] == no_depth_marker) {
2092 _dom_stk->push(j);
2093 j = _idom[j]->_idx;
2094 }
2095 // Compute the depth on the way back down this tree branch
2096 uint dd = _dom_depth[j] + 1;
2097 while (_dom_stk->length() > 0) {
2098 uint j = _dom_stk->pop();
2099 _dom_depth[j] = dd;
2100 dd++;
2101 }
2102 }
2103 }
2105 //------------------------------sort-------------------------------------------
2106 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2107 // loop tree, not the root.
2108 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2109 if( !innermost ) return loop; // New innermost loop
2111 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2112 assert( loop_preorder, "not yet post-walked loop" );
2113 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2114 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2116 // Insert at start of list
2117 while( l ) { // Insertion sort based on pre-order
2118 if( l == loop ) return innermost; // Already on list!
2119 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2120 assert( l_preorder, "not yet post-walked l" );
2121 // Check header pre-order number to figure proper nesting
2122 if( loop_preorder > l_preorder )
2123 break; // End of insertion
2124 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2125 // Since I split shared headers, you'd think this could not happen.
2126 // BUT: I must first do the preorder numbering before I can discover I
2127 // have shared headers, so the split headers all get the same preorder
2128 // number as the RegionNode they split from.
2129 if( loop_preorder == l_preorder &&
2130 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2131 break; // Also check for shared headers (same pre#)
2132 pp = &l->_parent; // Chain up list
2133 l = *pp;
2134 }
2135 // Link into list
2136 // Point predecessor to me
2137 *pp = loop;
2138 // Point me to successor
2139 IdealLoopTree *p = loop->_parent;
2140 loop->_parent = l; // Point me to successor
2141 if( p ) sort( p, innermost ); // Insert my parents into list as well
2142 return innermost;
2143 }
2145 //------------------------------build_loop_tree--------------------------------
2146 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2147 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2148 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2149 // tightest enclosing IdealLoopTree for post-walked.
2150 //
2151 // During my forward walk I do a short 1-layer lookahead to see if I can find
2152 // a loop backedge with that doesn't have any work on the backedge. This
2153 // helps me construct nested loops with shared headers better.
2154 //
2155 // Once I've done the forward recursion, I do the post-work. For each child
2156 // I check to see if there is a backedge. Backedges define a loop! I
2157 // insert an IdealLoopTree at the target of the backedge.
2158 //
2159 // During the post-work I also check to see if I have several children
2160 // belonging to different loops. If so, then this Node is a decision point
2161 // where control flow can choose to change loop nests. It is at this
2162 // decision point where I can figure out how loops are nested. At this
2163 // time I can properly order the different loop nests from my children.
2164 // Note that there may not be any backedges at the decision point!
2165 //
2166 // Since the decision point can be far removed from the backedges, I can't
2167 // order my loops at the time I discover them. Thus at the decision point
2168 // I need to inspect loop header pre-order numbers to properly nest my
2169 // loops. This means I need to sort my childrens' loops by pre-order.
2170 // The sort is of size number-of-control-children, which generally limits
2171 // it to size 2 (i.e., I just choose between my 2 target loops).
2172 void PhaseIdealLoop::build_loop_tree() {
2173 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2174 GrowableArray <Node *> bltstack(C->unique() >> 1);
2175 Node *n = C->root();
2176 bltstack.push(n);
2177 int pre_order = 1;
2178 int stack_size;
2180 while ( ( stack_size = bltstack.length() ) != 0 ) {
2181 n = bltstack.top(); // Leave node on stack
2182 if ( !is_visited(n) ) {
2183 // ---- Pre-pass Work ----
2184 // Pre-walked but not post-walked nodes need a pre_order number.
2186 set_preorder_visited( n, pre_order ); // set as visited
2188 // ---- Scan over children ----
2189 // Scan first over control projections that lead to loop headers.
2190 // This helps us find inner-to-outer loops with shared headers better.
2192 // Scan children's children for loop headers.
2193 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2194 Node* m = n->raw_out(i); // Child
2195 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2196 // Scan over children's children to find loop
2197 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2198 Node* l = m->fast_out(j);
2199 if( is_visited(l) && // Been visited?
2200 !is_postvisited(l) && // But not post-visited
2201 get_preorder(l) < pre_order ) { // And smaller pre-order
2202 // Found! Scan the DFS down this path before doing other paths
2203 bltstack.push(m);
2204 break;
2205 }
2206 }
2207 }
2208 }
2209 pre_order++;
2210 }
2211 else if ( !is_postvisited(n) ) {
2212 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2213 // such as com.sun.rsasign.am::a.
2214 // For non-recursive version, first, process current children.
2215 // On next iteration, check if additional children were added.
2216 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2217 Node* u = n->raw_out(k);
2218 if ( u->is_CFG() && !is_visited(u) ) {
2219 bltstack.push(u);
2220 }
2221 }
2222 if ( bltstack.length() == stack_size ) {
2223 // There were no additional children, post visit node now
2224 (void)bltstack.pop(); // Remove node from stack
2225 pre_order = build_loop_tree_impl( n, pre_order );
2226 // Check for bailout
2227 if (C->failing()) {
2228 return;
2229 }
2230 // Check to grow _preorders[] array for the case when
2231 // build_loop_tree_impl() adds new nodes.
2232 check_grow_preorders();
2233 }
2234 }
2235 else {
2236 (void)bltstack.pop(); // Remove post-visited node from stack
2237 }
2238 }
2239 }
2241 //------------------------------build_loop_tree_impl---------------------------
2242 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2243 // ---- Post-pass Work ----
2244 // Pre-walked but not post-walked nodes need a pre_order number.
2246 // Tightest enclosing loop for this Node
2247 IdealLoopTree *innermost = NULL;
2249 // For all children, see if any edge is a backedge. If so, make a loop
2250 // for it. Then find the tightest enclosing loop for the self Node.
2251 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2252 Node* m = n->fast_out(i); // Child
2253 if( n == m ) continue; // Ignore control self-cycles
2254 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2256 IdealLoopTree *l; // Child's loop
2257 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2258 // Found a backedge
2259 assert( get_preorder(m) < pre_order, "should be backedge" );
2260 // Check for the RootNode, which is already a LoopNode and is allowed
2261 // to have multiple "backedges".
2262 if( m == C->root()) { // Found the root?
2263 l = _ltree_root; // Root is the outermost LoopNode
2264 } else { // Else found a nested loop
2265 // Insert a LoopNode to mark this loop.
2266 l = new IdealLoopTree(this, m, n);
2267 } // End of Else found a nested loop
2268 if( !has_loop(m) ) // If 'm' does not already have a loop set
2269 set_loop(m, l); // Set loop header to loop now
2271 } else { // Else not a nested loop
2272 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2273 l = get_loop(m); // Get previously determined loop
2274 // If successor is header of a loop (nest), move up-loop till it
2275 // is a member of some outer enclosing loop. Since there are no
2276 // shared headers (I've split them already) I only need to go up
2277 // at most 1 level.
2278 while( l && l->_head == m ) // Successor heads loop?
2279 l = l->_parent; // Move up 1 for me
2280 // If this loop is not properly parented, then this loop
2281 // has no exit path out, i.e. its an infinite loop.
2282 if( !l ) {
2283 // Make loop "reachable" from root so the CFG is reachable. Basically
2284 // insert a bogus loop exit that is never taken. 'm', the loop head,
2285 // points to 'n', one (of possibly many) fall-in paths. There may be
2286 // many backedges as well.
2288 // Here I set the loop to be the root loop. I could have, after
2289 // inserting a bogus loop exit, restarted the recursion and found my
2290 // new loop exit. This would make the infinite loop a first-class
2291 // loop and it would then get properly optimized. What's the use of
2292 // optimizing an infinite loop?
2293 l = _ltree_root; // Oops, found infinite loop
2295 if (!_verify_only) {
2296 // Insert the NeverBranch between 'm' and it's control user.
2297 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2298 _igvn.register_new_node_with_optimizer(iff);
2299 set_loop(iff, l);
2300 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2301 _igvn.register_new_node_with_optimizer(if_t);
2302 set_loop(if_t, l);
2304 Node* cfg = NULL; // Find the One True Control User of m
2305 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2306 Node* x = m->fast_out(j);
2307 if (x->is_CFG() && x != m && x != iff)
2308 { cfg = x; break; }
2309 }
2310 assert(cfg != NULL, "must find the control user of m");
2311 uint k = 0; // Probably cfg->in(0)
2312 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2313 cfg->set_req( k, if_t ); // Now point to NeverBranch
2315 // Now create the never-taken loop exit
2316 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2317 _igvn.register_new_node_with_optimizer(if_f);
2318 set_loop(if_f, l);
2319 // Find frame ptr for Halt. Relies on the optimizer
2320 // V-N'ing. Easier and quicker than searching through
2321 // the program structure.
2322 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2323 _igvn.register_new_node_with_optimizer(frame);
2324 // Halt & Catch Fire
2325 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2326 _igvn.register_new_node_with_optimizer(halt);
2327 set_loop(halt, l);
2328 C->root()->add_req(halt);
2329 }
2330 set_loop(C->root(), _ltree_root);
2331 }
2332 }
2333 // Weeny check for irreducible. This child was already visited (this
2334 // IS the post-work phase). Is this child's loop header post-visited
2335 // as well? If so, then I found another entry into the loop.
2336 if (!_verify_only) {
2337 while( is_postvisited(l->_head) ) {
2338 // found irreducible
2339 l->_irreducible = 1; // = true
2340 l = l->_parent;
2341 _has_irreducible_loops = true;
2342 // Check for bad CFG here to prevent crash, and bailout of compile
2343 if (l == NULL) {
2344 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2345 return pre_order;
2346 }
2347 }
2348 }
2350 // This Node might be a decision point for loops. It is only if
2351 // it's children belong to several different loops. The sort call
2352 // does a trivial amount of work if there is only 1 child or all
2353 // children belong to the same loop. If however, the children
2354 // belong to different loops, the sort call will properly set the
2355 // _parent pointers to show how the loops nest.
2356 //
2357 // In any case, it returns the tightest enclosing loop.
2358 innermost = sort( l, innermost );
2359 }
2361 // Def-use info will have some dead stuff; dead stuff will have no
2362 // loop decided on.
2364 // Am I a loop header? If so fix up my parent's child and next ptrs.
2365 if( innermost && innermost->_head == n ) {
2366 assert( get_loop(n) == innermost, "" );
2367 IdealLoopTree *p = innermost->_parent;
2368 IdealLoopTree *l = innermost;
2369 while( p && l->_head == n ) {
2370 l->_next = p->_child; // Put self on parents 'next child'
2371 p->_child = l; // Make self as first child of parent
2372 l = p; // Now walk up the parent chain
2373 p = l->_parent;
2374 }
2375 } else {
2376 // Note that it is possible for a LoopNode to reach here, if the
2377 // backedge has been made unreachable (hence the LoopNode no longer
2378 // denotes a Loop, and will eventually be removed).
2380 // Record tightest enclosing loop for self. Mark as post-visited.
2381 set_loop(n, innermost);
2382 // Also record has_call flag early on
2383 if( innermost ) {
2384 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2385 // Do not count uncommon calls
2386 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2387 Node *iff = n->in(0)->in(0);
2388 if( !iff->is_If() ||
2389 (n->in(0)->Opcode() == Op_IfFalse &&
2390 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2391 (iff->as_If()->_prob >= 0.01) )
2392 innermost->_has_call = 1;
2393 }
2394 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2395 // Disable loop optimizations if the loop has a scalar replaceable
2396 // allocation. This disabling may cause a potential performance lost
2397 // if the allocation is not eliminated for some reason.
2398 innermost->_allow_optimizations = false;
2399 innermost->_has_call = 1; // = true
2400 }
2401 }
2402 }
2404 // Flag as post-visited now
2405 set_postvisited(n);
2406 return pre_order;
2407 }
2410 //------------------------------build_loop_early-------------------------------
2411 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2412 // First pass computes the earliest controlling node possible. This is the
2413 // controlling input with the deepest dominating depth.
2414 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2415 while (worklist.size() != 0) {
2416 // Use local variables nstack_top_n & nstack_top_i to cache values
2417 // on nstack's top.
2418 Node *nstack_top_n = worklist.pop();
2419 uint nstack_top_i = 0;
2420 //while_nstack_nonempty:
2421 while (true) {
2422 // Get parent node and next input's index from stack's top.
2423 Node *n = nstack_top_n;
2424 uint i = nstack_top_i;
2425 uint cnt = n->req(); // Count of inputs
2426 if (i == 0) { // Pre-process the node.
2427 if( has_node(n) && // Have either loop or control already?
2428 !has_ctrl(n) ) { // Have loop picked out already?
2429 // During "merge_many_backedges" we fold up several nested loops
2430 // into a single loop. This makes the members of the original
2431 // loop bodies pointing to dead loops; they need to move up
2432 // to the new UNION'd larger loop. I set the _head field of these
2433 // dead loops to NULL and the _parent field points to the owning
2434 // loop. Shades of UNION-FIND algorithm.
2435 IdealLoopTree *ilt;
2436 while( !(ilt = get_loop(n))->_head ) {
2437 // Normally I would use a set_loop here. But in this one special
2438 // case, it is legal (and expected) to change what loop a Node
2439 // belongs to.
2440 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2441 }
2442 // Remove safepoints ONLY if I've already seen I don't need one.
2443 // (the old code here would yank a 2nd safepoint after seeing a
2444 // first one, even though the 1st did not dominate in the loop body
2445 // and thus could be avoided indefinitely)
2446 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2447 is_deleteable_safept(n)) {
2448 Node *in = n->in(TypeFunc::Control);
2449 lazy_replace(n,in); // Pull safepoint now
2450 // Carry on with the recursion "as if" we are walking
2451 // only the control input
2452 if( !visited.test_set( in->_idx ) ) {
2453 worklist.push(in); // Visit this guy later, using worklist
2454 }
2455 // Get next node from nstack:
2456 // - skip n's inputs processing by setting i > cnt;
2457 // - we also will not call set_early_ctrl(n) since
2458 // has_node(n) == true (see the condition above).
2459 i = cnt + 1;
2460 }
2461 }
2462 } // if (i == 0)
2464 // Visit all inputs
2465 bool done = true; // Assume all n's inputs will be processed
2466 while (i < cnt) {
2467 Node *in = n->in(i);
2468 ++i;
2469 if (in == NULL) continue;
2470 if (in->pinned() && !in->is_CFG())
2471 set_ctrl(in, in->in(0));
2472 int is_visited = visited.test_set( in->_idx );
2473 if (!has_node(in)) { // No controlling input yet?
2474 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2475 assert( !is_visited, "visit only once" );
2476 nstack.push(n, i); // Save parent node and next input's index.
2477 nstack_top_n = in; // Process current input now.
2478 nstack_top_i = 0;
2479 done = false; // Not all n's inputs processed.
2480 break; // continue while_nstack_nonempty;
2481 } else if (!is_visited) {
2482 // This guy has a location picked out for him, but has not yet
2483 // been visited. Happens to all CFG nodes, for instance.
2484 // Visit him using the worklist instead of recursion, to break
2485 // cycles. Since he has a location already we do not need to
2486 // find his location before proceeding with the current Node.
2487 worklist.push(in); // Visit this guy later, using worklist
2488 }
2489 }
2490 if (done) {
2491 // All of n's inputs have been processed, complete post-processing.
2493 // Compute earliest point this Node can go.
2494 // CFG, Phi, pinned nodes already know their controlling input.
2495 if (!has_node(n)) {
2496 // Record earliest legal location
2497 set_early_ctrl( n );
2498 }
2499 if (nstack.is_empty()) {
2500 // Finished all nodes on stack.
2501 // Process next node on the worklist.
2502 break;
2503 }
2504 // Get saved parent node and next input's index.
2505 nstack_top_n = nstack.node();
2506 nstack_top_i = nstack.index();
2507 nstack.pop();
2508 }
2509 } // while (true)
2510 }
2511 }
2513 //------------------------------dom_lca_internal--------------------------------
2514 // Pair-wise LCA
2515 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2516 if( !n1 ) return n2; // Handle NULL original LCA
2517 assert( n1->is_CFG(), "" );
2518 assert( n2->is_CFG(), "" );
2519 // find LCA of all uses
2520 uint d1 = dom_depth(n1);
2521 uint d2 = dom_depth(n2);
2522 while (n1 != n2) {
2523 if (d1 > d2) {
2524 n1 = idom(n1);
2525 d1 = dom_depth(n1);
2526 } else if (d1 < d2) {
2527 n2 = idom(n2);
2528 d2 = dom_depth(n2);
2529 } else {
2530 // Here d1 == d2. Due to edits of the dominator-tree, sections
2531 // of the tree might have the same depth. These sections have
2532 // to be searched more carefully.
2534 // Scan up all the n1's with equal depth, looking for n2.
2535 Node *t1 = idom(n1);
2536 while (dom_depth(t1) == d1) {
2537 if (t1 == n2) return n2;
2538 t1 = idom(t1);
2539 }
2540 // Scan up all the n2's with equal depth, looking for n1.
2541 Node *t2 = idom(n2);
2542 while (dom_depth(t2) == d2) {
2543 if (t2 == n1) return n1;
2544 t2 = idom(t2);
2545 }
2546 // Move up to a new dominator-depth value as well as up the dom-tree.
2547 n1 = t1;
2548 n2 = t2;
2549 d1 = dom_depth(n1);
2550 d2 = dom_depth(n2);
2551 }
2552 }
2553 return n1;
2554 }
2556 //------------------------------compute_idom-----------------------------------
2557 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2558 // IDOMs are correct.
2559 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2560 assert( region->is_Region(), "" );
2561 Node *LCA = NULL;
2562 for( uint i = 1; i < region->req(); i++ ) {
2563 if( region->in(i) != C->top() )
2564 LCA = dom_lca( LCA, region->in(i) );
2565 }
2566 return LCA;
2567 }
2569 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
2570 bool had_error = false;
2571 #ifdef ASSERT
2572 if (early != C->root()) {
2573 // Make sure that there's a dominance path from use to LCA
2574 Node* d = use;
2575 while (d != LCA) {
2576 d = idom(d);
2577 if (d == C->root()) {
2578 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
2579 n->dump();
2580 use->dump();
2581 had_error = true;
2582 break;
2583 }
2584 }
2585 }
2586 #endif
2587 return had_error;
2588 }
2591 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
2592 // Compute LCA over list of uses
2593 bool had_error = false;
2594 Node *LCA = NULL;
2595 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2596 Node* c = n->fast_out(i);
2597 if (_nodes[c->_idx] == NULL)
2598 continue; // Skip the occasional dead node
2599 if( c->is_Phi() ) { // For Phis, we must land above on the path
2600 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2601 if( c->in(j) == n ) { // Found matching input?
2602 Node *use = c->in(0)->in(j);
2603 if (_verify_only && use->is_top()) continue;
2604 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2605 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2606 }
2607 }
2608 } else {
2609 // For CFG data-users, use is in the block just prior
2610 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2611 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2612 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2613 }
2614 }
2615 assert(!had_error, "bad dominance");
2616 return LCA;
2617 }
2619 //------------------------------get_late_ctrl----------------------------------
2620 // Compute latest legal control.
2621 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2622 assert(early != NULL, "early control should not be NULL");
2624 Node* LCA = compute_lca_of_uses(n, early);
2625 #ifdef ASSERT
2626 if (LCA == C->root() && LCA != early) {
2627 // def doesn't dominate uses so print some useful debugging output
2628 compute_lca_of_uses(n, early, true);
2629 }
2630 #endif
2632 // if this is a load, check for anti-dependent stores
2633 // We use a conservative algorithm to identify potential interfering
2634 // instructions and for rescheduling the load. The users of the memory
2635 // input of this load are examined. Any use which is not a load and is
2636 // dominated by early is considered a potentially interfering store.
2637 // This can produce false positives.
2638 if (n->is_Load() && LCA != early) {
2639 Node_List worklist;
2641 Node *mem = n->in(MemNode::Memory);
2642 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2643 Node* s = mem->fast_out(i);
2644 worklist.push(s);
2645 }
2646 while(worklist.size() != 0 && LCA != early) {
2647 Node* s = worklist.pop();
2648 if (s->is_Load()) {
2649 continue;
2650 } else if (s->is_MergeMem()) {
2651 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
2652 Node* s1 = s->fast_out(i);
2653 worklist.push(s1);
2654 }
2655 } else {
2656 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
2657 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
2658 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
2659 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
2660 }
2661 }
2662 }
2663 }
2665 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
2666 return LCA;
2667 }
2669 // true if CFG node d dominates CFG node n
2670 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
2671 if (d == n)
2672 return true;
2673 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
2674 uint dd = dom_depth(d);
2675 while (dom_depth(n) >= dd) {
2676 if (n == d)
2677 return true;
2678 n = idom(n);
2679 }
2680 return false;
2681 }
2683 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
2684 // Pair-wise LCA with tags.
2685 // Tag each index with the node 'tag' currently being processed
2686 // before advancing up the dominator chain using idom().
2687 // Later calls that find a match to 'tag' know that this path has already
2688 // been considered in the current LCA (which is input 'n1' by convention).
2689 // Since get_late_ctrl() is only called once for each node, the tag array
2690 // does not need to be cleared between calls to get_late_ctrl().
2691 // Algorithm trades a larger constant factor for better asymptotic behavior
2692 //
2693 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
2694 uint d1 = dom_depth(n1);
2695 uint d2 = dom_depth(n2);
2697 do {
2698 if (d1 > d2) {
2699 // current lca is deeper than n2
2700 _dom_lca_tags.map(n1->_idx, tag);
2701 n1 = idom(n1);
2702 d1 = dom_depth(n1);
2703 } else if (d1 < d2) {
2704 // n2 is deeper than current lca
2705 Node *memo = _dom_lca_tags[n2->_idx];
2706 if( memo == tag ) {
2707 return n1; // Return the current LCA
2708 }
2709 _dom_lca_tags.map(n2->_idx, tag);
2710 n2 = idom(n2);
2711 d2 = dom_depth(n2);
2712 } else {
2713 // Here d1 == d2. Due to edits of the dominator-tree, sections
2714 // of the tree might have the same depth. These sections have
2715 // to be searched more carefully.
2717 // Scan up all the n1's with equal depth, looking for n2.
2718 _dom_lca_tags.map(n1->_idx, tag);
2719 Node *t1 = idom(n1);
2720 while (dom_depth(t1) == d1) {
2721 if (t1 == n2) return n2;
2722 _dom_lca_tags.map(t1->_idx, tag);
2723 t1 = idom(t1);
2724 }
2725 // Scan up all the n2's with equal depth, looking for n1.
2726 _dom_lca_tags.map(n2->_idx, tag);
2727 Node *t2 = idom(n2);
2728 while (dom_depth(t2) == d2) {
2729 if (t2 == n1) return n1;
2730 _dom_lca_tags.map(t2->_idx, tag);
2731 t2 = idom(t2);
2732 }
2733 // Move up to a new dominator-depth value as well as up the dom-tree.
2734 n1 = t1;
2735 n2 = t2;
2736 d1 = dom_depth(n1);
2737 d2 = dom_depth(n2);
2738 }
2739 } while (n1 != n2);
2740 return n1;
2741 }
2743 //------------------------------init_dom_lca_tags------------------------------
2744 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2745 // Intended use does not involve any growth for the array, so it could
2746 // be of fixed size.
2747 void PhaseIdealLoop::init_dom_lca_tags() {
2748 uint limit = C->unique() + 1;
2749 _dom_lca_tags.map( limit, NULL );
2750 #ifdef ASSERT
2751 for( uint i = 0; i < limit; ++i ) {
2752 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2753 }
2754 #endif // ASSERT
2755 }
2757 //------------------------------clear_dom_lca_tags------------------------------
2758 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
2759 // Intended use does not involve any growth for the array, so it could
2760 // be of fixed size.
2761 void PhaseIdealLoop::clear_dom_lca_tags() {
2762 uint limit = C->unique() + 1;
2763 _dom_lca_tags.map( limit, NULL );
2764 _dom_lca_tags.clear();
2765 #ifdef ASSERT
2766 for( uint i = 0; i < limit; ++i ) {
2767 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
2768 }
2769 #endif // ASSERT
2770 }
2772 //------------------------------build_loop_late--------------------------------
2773 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2774 // Second pass finds latest legal placement, and ideal loop placement.
2775 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2776 while (worklist.size() != 0) {
2777 Node *n = worklist.pop();
2778 // Only visit once
2779 if (visited.test_set(n->_idx)) continue;
2780 uint cnt = n->outcnt();
2781 uint i = 0;
2782 while (true) {
2783 assert( _nodes[n->_idx], "no dead nodes" );
2784 // Visit all children
2785 if (i < cnt) {
2786 Node* use = n->raw_out(i);
2787 ++i;
2788 // Check for dead uses. Aggressively prune such junk. It might be
2789 // dead in the global sense, but still have local uses so I cannot
2790 // easily call 'remove_dead_node'.
2791 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
2792 // Due to cycles, we might not hit the same fixed point in the verify
2793 // pass as we do in the regular pass. Instead, visit such phis as
2794 // simple uses of the loop head.
2795 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
2796 if( !visited.test(use->_idx) )
2797 worklist.push(use);
2798 } else if( !visited.test_set(use->_idx) ) {
2799 nstack.push(n, i); // Save parent and next use's index.
2800 n = use; // Process all children of current use.
2801 cnt = use->outcnt();
2802 i = 0;
2803 }
2804 } else {
2805 // Do not visit around the backedge of loops via data edges.
2806 // push dead code onto a worklist
2807 _deadlist.push(use);
2808 }
2809 } else {
2810 // All of n's children have been processed, complete post-processing.
2811 build_loop_late_post(n);
2812 if (nstack.is_empty()) {
2813 // Finished all nodes on stack.
2814 // Process next node on the worklist.
2815 break;
2816 }
2817 // Get saved parent node and next use's index. Visit the rest of uses.
2818 n = nstack.node();
2819 cnt = n->outcnt();
2820 i = nstack.index();
2821 nstack.pop();
2822 }
2823 }
2824 }
2825 }
2827 //------------------------------build_loop_late_post---------------------------
2828 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2829 // Second pass finds latest legal placement, and ideal loop placement.
2830 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
2832 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
2833 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
2834 }
2836 // CFG and pinned nodes already handled
2837 if( n->in(0) ) {
2838 if( n->in(0)->is_top() ) return; // Dead?
2840 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
2841 // _must_ be pinned (they have to observe their control edge of course).
2842 // Unlike Stores (which modify an unallocable resource, the memory
2843 // state), Mods/Loads can float around. So free them up.
2844 bool pinned = true;
2845 switch( n->Opcode() ) {
2846 case Op_DivI:
2847 case Op_DivF:
2848 case Op_DivD:
2849 case Op_ModI:
2850 case Op_ModF:
2851 case Op_ModD:
2852 case Op_LoadB: // Same with Loads; they can sink
2853 case Op_LoadUS: // during loop optimizations.
2854 case Op_LoadD:
2855 case Op_LoadF:
2856 case Op_LoadI:
2857 case Op_LoadKlass:
2858 case Op_LoadNKlass:
2859 case Op_LoadL:
2860 case Op_LoadS:
2861 case Op_LoadP:
2862 case Op_LoadN:
2863 case Op_LoadRange:
2864 case Op_LoadD_unaligned:
2865 case Op_LoadL_unaligned:
2866 case Op_StrComp: // Does a bunch of load-like effects
2867 case Op_StrEquals:
2868 case Op_StrIndexOf:
2869 case Op_AryEq:
2870 pinned = false;
2871 }
2872 if( pinned ) {
2873 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
2874 if( !chosen_loop->_child ) // Inner loop?
2875 chosen_loop->_body.push(n); // Collect inner loops
2876 return;
2877 }
2878 } else { // No slot zero
2879 if( n->is_CFG() ) { // CFG with no slot 0 is dead
2880 _nodes.map(n->_idx,0); // No block setting, it's globally dead
2881 return;
2882 }
2883 assert(!n->is_CFG() || n->outcnt() == 0, "");
2884 }
2886 // Do I have a "safe range" I can select over?
2887 Node *early = get_ctrl(n);// Early location already computed
2889 // Compute latest point this Node can go
2890 Node *LCA = get_late_ctrl( n, early );
2891 // LCA is NULL due to uses being dead
2892 if( LCA == NULL ) {
2893 #ifdef ASSERT
2894 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
2895 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
2896 }
2897 #endif
2898 _nodes.map(n->_idx, 0); // This node is useless
2899 _deadlist.push(n);
2900 return;
2901 }
2902 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
2904 Node *legal = LCA; // Walk 'legal' up the IDOM chain
2905 Node *least = legal; // Best legal position so far
2906 while( early != legal ) { // While not at earliest legal
2907 #ifdef ASSERT
2908 if (legal->is_Start() && !early->is_Root()) {
2909 // Bad graph. Print idom path and fail.
2910 tty->print_cr( "Bad graph detected in build_loop_late");
2911 tty->print("n: ");n->dump(); tty->cr();
2912 tty->print("early: ");early->dump(); tty->cr();
2913 int ct = 0;
2914 Node *dbg_legal = LCA;
2915 while(!dbg_legal->is_Start() && ct < 100) {
2916 tty->print("idom[%d] ",ct); dbg_legal->dump(); tty->cr();
2917 ct++;
2918 dbg_legal = idom(dbg_legal);
2919 }
2920 assert(false, "Bad graph detected in build_loop_late");
2921 }
2922 #endif
2923 // Find least loop nesting depth
2924 legal = idom(legal); // Bump up the IDOM tree
2925 // Check for lower nesting depth
2926 if( get_loop(legal)->_nest < get_loop(least)->_nest )
2927 least = legal;
2928 }
2929 assert(early == legal || legal != C->root(), "bad dominance of inputs");
2931 // Try not to place code on a loop entry projection
2932 // which can inhibit range check elimination.
2933 if (least != early) {
2934 Node* ctrl_out = least->unique_ctrl_out();
2935 if (ctrl_out && ctrl_out->is_CountedLoop() &&
2936 least == ctrl_out->in(LoopNode::EntryControl)) {
2937 Node* least_dom = idom(least);
2938 if (get_loop(least_dom)->is_member(get_loop(least))) {
2939 least = least_dom;
2940 }
2941 }
2942 }
2944 #ifdef ASSERT
2945 // If verifying, verify that 'verify_me' has a legal location
2946 // and choose it as our location.
2947 if( _verify_me ) {
2948 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
2949 Node *legal = LCA;
2950 while( early != legal ) { // While not at earliest legal
2951 if( legal == v_ctrl ) break; // Check for prior good location
2952 legal = idom(legal) ;// Bump up the IDOM tree
2953 }
2954 // Check for prior good location
2955 if( legal == v_ctrl ) least = legal; // Keep prior if found
2956 }
2957 #endif
2959 // Assign discovered "here or above" point
2960 least = find_non_split_ctrl(least);
2961 set_ctrl(n, least);
2963 // Collect inner loop bodies
2964 IdealLoopTree *chosen_loop = get_loop(least);
2965 if( !chosen_loop->_child ) // Inner loop?
2966 chosen_loop->_body.push(n);// Collect inner loops
2967 }
2969 #ifndef PRODUCT
2970 //------------------------------dump-------------------------------------------
2971 void PhaseIdealLoop::dump( ) const {
2972 ResourceMark rm;
2973 Arena* arena = Thread::current()->resource_area();
2974 Node_Stack stack(arena, C->unique() >> 2);
2975 Node_List rpo_list;
2976 VectorSet visited(arena);
2977 visited.set(C->top()->_idx);
2978 rpo( C->root(), stack, visited, rpo_list );
2979 // Dump root loop indexed by last element in PO order
2980 dump( _ltree_root, rpo_list.size(), rpo_list );
2981 }
2983 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
2984 loop->dump_head();
2986 // Now scan for CFG nodes in the same loop
2987 for( uint j=idx; j > 0; j-- ) {
2988 Node *n = rpo_list[j-1];
2989 if( !_nodes[n->_idx] ) // Skip dead nodes
2990 continue;
2991 if( get_loop(n) != loop ) { // Wrong loop nest
2992 if( get_loop(n)->_head == n && // Found nested loop?
2993 get_loop(n)->_parent == loop )
2994 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
2995 continue;
2996 }
2998 // Dump controlling node
2999 for( uint x = 0; x < loop->_nest; x++ )
3000 tty->print(" ");
3001 tty->print("C");
3002 if( n == C->root() ) {
3003 n->dump();
3004 } else {
3005 Node* cached_idom = idom_no_update(n);
3006 Node *computed_idom = n->in(0);
3007 if( n->is_Region() ) {
3008 computed_idom = compute_idom(n);
3009 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3010 // any MultiBranch ctrl node), so apply a similar transform to
3011 // the cached idom returned from idom_no_update.
3012 cached_idom = find_non_split_ctrl(cached_idom);
3013 }
3014 tty->print(" ID:%d",computed_idom->_idx);
3015 n->dump();
3016 if( cached_idom != computed_idom ) {
3017 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3018 computed_idom->_idx, cached_idom->_idx);
3019 }
3020 }
3021 // Dump nodes it controls
3022 for( uint k = 0; k < _nodes.Size(); k++ ) {
3023 // (k < C->unique() && get_ctrl(find(k)) == n)
3024 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3025 Node *m = C->root()->find(k);
3026 if( m && m->outcnt() > 0 ) {
3027 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3028 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3029 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3030 }
3031 for( uint j = 0; j < loop->_nest; j++ )
3032 tty->print(" ");
3033 tty->print(" ");
3034 m->dump();
3035 }
3036 }
3037 }
3038 }
3039 }
3041 // Collect a R-P-O for the whole CFG.
3042 // Result list is in post-order (scan backwards for RPO)
3043 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3044 stk.push(start, 0);
3045 visited.set(start->_idx);
3047 while (stk.is_nonempty()) {
3048 Node* m = stk.node();
3049 uint idx = stk.index();
3050 if (idx < m->outcnt()) {
3051 stk.set_index(idx + 1);
3052 Node* n = m->raw_out(idx);
3053 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3054 stk.push(n, 0);
3055 }
3056 } else {
3057 rpo_list.push(m);
3058 stk.pop();
3059 }
3060 }
3061 }
3062 #endif
3065 //=============================================================================
3066 //------------------------------LoopTreeIterator-----------------------------------
3068 // Advance to next loop tree using a preorder, left-to-right traversal.
3069 void LoopTreeIterator::next() {
3070 assert(!done(), "must not be done.");
3071 if (_curnt->_child != NULL) {
3072 _curnt = _curnt->_child;
3073 } else if (_curnt->_next != NULL) {
3074 _curnt = _curnt->_next;
3075 } else {
3076 while (_curnt != _root && _curnt->_next == NULL) {
3077 _curnt = _curnt->_parent;
3078 }
3079 if (_curnt == _root) {
3080 _curnt = NULL;
3081 assert(done(), "must be done.");
3082 } else {
3083 assert(_curnt->_next != NULL, "must be more to do");
3084 _curnt = _curnt->_next;
3085 }
3086 }
3087 }